xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision f5057be340e44f3edc0fe90fa875eb89a4c49b4f)
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  * Spawn an Ethernet device from Verbs information.
516  *
517  * @param dpdk_dev
518  *   Backing DPDK device.
519  * @param spawn
520  *   Verbs device parameters (name, port, switch_info) to spawn.
521  * @param config
522  *   Device configuration parameters.
523  *
524  * @return
525  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
526  *   is set. The following errors are defined:
527  *
528  *   EBUSY: device is not supposed to be spawned.
529  *   EEXIST: device is already spawned
530  */
531 static struct rte_eth_dev *
532 mlx5_dev_spawn(struct rte_device *dpdk_dev,
533 	       struct mlx5_dev_spawn_data *spawn,
534 	       struct mlx5_dev_config *config)
535 {
536 	const struct mlx5_switch_info *switch_info = &spawn->info;
537 	struct mlx5_dev_ctx_shared *sh = NULL;
538 	struct ibv_port_attr port_attr;
539 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
540 	struct rte_eth_dev *eth_dev = NULL;
541 	struct mlx5_priv *priv = NULL;
542 	int err = 0;
543 	unsigned int hw_padding = 0;
544 	unsigned int mps;
545 	unsigned int cqe_comp;
546 	unsigned int cqe_pad = 0;
547 	unsigned int tunnel_en = 0;
548 	unsigned int mpls_en = 0;
549 	unsigned int swp = 0;
550 	unsigned int mprq = 0;
551 	unsigned int mprq_min_stride_size_n = 0;
552 	unsigned int mprq_max_stride_size_n = 0;
553 	unsigned int mprq_min_stride_num_n = 0;
554 	unsigned int mprq_max_stride_num_n = 0;
555 	struct rte_ether_addr mac;
556 	char name[RTE_ETH_NAME_MAX_LEN];
557 	int own_domain_id = 0;
558 	uint16_t port_id;
559 	unsigned int i;
560 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
561 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
562 #endif
563 
564 	/* Determine if this port representor is supposed to be spawned. */
565 	if (switch_info->representor && dpdk_dev->devargs) {
566 		struct rte_eth_devargs eth_da;
567 
568 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
569 		if (err) {
570 			rte_errno = -err;
571 			DRV_LOG(ERR, "failed to process device arguments: %s",
572 				strerror(rte_errno));
573 			return NULL;
574 		}
575 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
576 			if (eth_da.representor_ports[i] ==
577 			    (uint16_t)switch_info->port_name)
578 				break;
579 		if (i == eth_da.nb_representor_ports) {
580 			rte_errno = EBUSY;
581 			return NULL;
582 		}
583 	}
584 	/* Build device name. */
585 	if (spawn->pf_bond <  0) {
586 		/* Single device. */
587 		if (!switch_info->representor)
588 			strlcpy(name, dpdk_dev->name, sizeof(name));
589 		else
590 			snprintf(name, sizeof(name), "%s_representor_%u",
591 				 dpdk_dev->name, switch_info->port_name);
592 	} else {
593 		/* Bonding device. */
594 		if (!switch_info->representor)
595 			snprintf(name, sizeof(name), "%s_%s",
596 				 dpdk_dev->name,
597 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
598 		else
599 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
600 				 dpdk_dev->name,
601 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
602 				 switch_info->port_name);
603 	}
604 	/* check if the device is already spawned */
605 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
606 		rte_errno = EEXIST;
607 		return NULL;
608 	}
609 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
610 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
611 		struct mlx5_mp_id mp_id;
612 
613 		eth_dev = rte_eth_dev_attach_secondary(name);
614 		if (eth_dev == NULL) {
615 			DRV_LOG(ERR, "can not attach rte ethdev");
616 			rte_errno = ENOMEM;
617 			return NULL;
618 		}
619 		eth_dev->device = dpdk_dev;
620 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
621 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
622 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
623 		err = mlx5_proc_priv_init(eth_dev);
624 		if (err)
625 			return NULL;
626 		mp_id.port_id = eth_dev->data->port_id;
627 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
628 		/* Receive command fd from primary process */
629 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
630 		if (err < 0)
631 			goto err_secondary;
632 		/* Remap UAR for Tx queues. */
633 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
634 		if (err)
635 			goto err_secondary;
636 		/*
637 		 * Ethdev pointer is still required as input since
638 		 * the primary device is not accessible from the
639 		 * secondary process.
640 		 */
641 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
642 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
643 		return eth_dev;
644 err_secondary:
645 		mlx5_dev_close(eth_dev);
646 		return NULL;
647 	}
648 	/*
649 	 * Some parameters ("tx_db_nc" in particularly) are needed in
650 	 * advance to create dv/verbs device context. We proceed the
651 	 * devargs here to get ones, and later proceed devargs again
652 	 * to override some hardware settings.
653 	 */
654 	err = mlx5_args(config, dpdk_dev->devargs);
655 	if (err) {
656 		err = rte_errno;
657 		DRV_LOG(ERR, "failed to process device arguments: %s",
658 			strerror(rte_errno));
659 		goto error;
660 	}
661 	mlx5_malloc_mem_select(config->sys_mem_en);
662 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
663 	if (!sh)
664 		return NULL;
665 	config->devx = sh->devx;
666 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
667 	config->dest_tir = 1;
668 #endif
669 #ifdef HAVE_IBV_MLX5_MOD_SWP
670 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
671 #endif
672 	/*
673 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
674 	 * as all ConnectX-5 devices.
675 	 */
676 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
677 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
678 #endif
679 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
680 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
681 #endif
682 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
683 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
684 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
685 			DRV_LOG(DEBUG, "enhanced MPW is supported");
686 			mps = MLX5_MPW_ENHANCED;
687 		} else {
688 			DRV_LOG(DEBUG, "MPW is supported");
689 			mps = MLX5_MPW;
690 		}
691 	} else {
692 		DRV_LOG(DEBUG, "MPW isn't supported");
693 		mps = MLX5_MPW_DISABLED;
694 	}
695 #ifdef HAVE_IBV_MLX5_MOD_SWP
696 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
697 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
698 	DRV_LOG(DEBUG, "SWP support: %u", swp);
699 #endif
700 	config->swp = !!swp;
701 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
702 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
703 		struct mlx5dv_striding_rq_caps mprq_caps =
704 			dv_attr.striding_rq_caps;
705 
706 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
707 			mprq_caps.min_single_stride_log_num_of_bytes);
708 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
709 			mprq_caps.max_single_stride_log_num_of_bytes);
710 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
711 			mprq_caps.min_single_wqe_log_num_of_strides);
712 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
713 			mprq_caps.max_single_wqe_log_num_of_strides);
714 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
715 			mprq_caps.supported_qpts);
716 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
717 		mprq = 1;
718 		mprq_min_stride_size_n =
719 			mprq_caps.min_single_stride_log_num_of_bytes;
720 		mprq_max_stride_size_n =
721 			mprq_caps.max_single_stride_log_num_of_bytes;
722 		mprq_min_stride_num_n =
723 			mprq_caps.min_single_wqe_log_num_of_strides;
724 		mprq_max_stride_num_n =
725 			mprq_caps.max_single_wqe_log_num_of_strides;
726 	}
727 #endif
728 	if (RTE_CACHE_LINE_SIZE == 128 &&
729 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
730 		cqe_comp = 0;
731 	else
732 		cqe_comp = 1;
733 	config->cqe_comp = cqe_comp;
734 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
735 	/* Whether device supports 128B Rx CQE padding. */
736 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
737 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
738 #endif
739 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
740 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
741 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
742 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
743 			     (dv_attr.tunnel_offloads_caps &
744 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
745 			     (dv_attr.tunnel_offloads_caps &
746 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
747 	}
748 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
749 		tunnel_en ? "" : "not ");
750 #else
751 	DRV_LOG(WARNING,
752 		"tunnel offloading disabled due to old OFED/rdma-core version");
753 #endif
754 	config->tunnel_en = tunnel_en;
755 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
756 	mpls_en = ((dv_attr.tunnel_offloads_caps &
757 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
758 		   (dv_attr.tunnel_offloads_caps &
759 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
760 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
761 		mpls_en ? "" : "not ");
762 #else
763 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
764 		" old OFED/rdma-core version or firmware configuration");
765 #endif
766 	config->mpls_en = mpls_en;
767 	/* Check port status. */
768 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
769 	if (err) {
770 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
771 		goto error;
772 	}
773 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
774 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
775 		err = EINVAL;
776 		goto error;
777 	}
778 	if (port_attr.state != IBV_PORT_ACTIVE)
779 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
780 			mlx5_glue->port_state_str(port_attr.state),
781 			port_attr.state);
782 	/* Allocate private eth device data. */
783 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
784 			   sizeof(*priv),
785 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
786 	if (priv == NULL) {
787 		DRV_LOG(ERR, "priv allocation failure");
788 		err = ENOMEM;
789 		goto error;
790 	}
791 	priv->sh = sh;
792 	priv->dev_port = spawn->phys_port;
793 	priv->pci_dev = spawn->pci_dev;
794 	priv->mtu = RTE_ETHER_MTU;
795 	priv->mp_id.port_id = port_id;
796 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
797 	/* Some internal functions rely on Netlink sockets, open them now. */
798 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
799 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
800 	priv->representor = !!switch_info->representor;
801 	priv->master = !!switch_info->master;
802 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
803 	priv->vport_meta_tag = 0;
804 	priv->vport_meta_mask = 0;
805 	priv->pf_bond = spawn->pf_bond;
806 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
807 	/*
808 	 * The DevX port query API is implemented. E-Switch may use
809 	 * either vport or reg_c[0] metadata register to match on
810 	 * vport index. The engaged part of metadata register is
811 	 * defined by mask.
812 	 */
813 	if (switch_info->representor || switch_info->master) {
814 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
815 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
816 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
817 						 &devx_port);
818 		if (err) {
819 			DRV_LOG(WARNING,
820 				"can't query devx port %d on device %s",
821 				spawn->phys_port,
822 				mlx5_os_get_dev_device_name(spawn->phys_dev));
823 			devx_port.comp_mask = 0;
824 		}
825 	}
826 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
827 		priv->vport_meta_tag = devx_port.reg_c_0.value;
828 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
829 		if (!priv->vport_meta_mask) {
830 			DRV_LOG(ERR, "vport zero mask for port %d"
831 				     " on bonding device %s",
832 				     spawn->phys_port,
833 				     mlx5_os_get_dev_device_name
834 							(spawn->phys_dev));
835 			err = ENOTSUP;
836 			goto error;
837 		}
838 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
839 			DRV_LOG(ERR, "invalid vport tag for port %d"
840 				     " on bonding device %s",
841 				     spawn->phys_port,
842 				     mlx5_os_get_dev_device_name
843 							(spawn->phys_dev));
844 			err = ENOTSUP;
845 			goto error;
846 		}
847 	}
848 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
849 		priv->vport_id = devx_port.vport_num;
850 	} else if (spawn->pf_bond >= 0) {
851 		DRV_LOG(ERR, "can't deduce vport index for port %d"
852 			     " on bonding device %s",
853 			     spawn->phys_port,
854 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
855 		err = ENOTSUP;
856 		goto error;
857 	} else {
858 		/* Suppose vport index in compatible way. */
859 		priv->vport_id = switch_info->representor ?
860 				 switch_info->port_name + 1 : -1;
861 	}
862 #else
863 	/*
864 	 * Kernel/rdma_core support single E-Switch per PF configurations
865 	 * only and vport_id field contains the vport index for
866 	 * associated VF, which is deduced from representor port name.
867 	 * For example, let's have the IB device port 10, it has
868 	 * attached network device eth0, which has port name attribute
869 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
870 	 * as 3 (2+1). This assigning schema should be changed if the
871 	 * multiple E-Switch instances per PF configurations or/and PCI
872 	 * subfunctions are added.
873 	 */
874 	priv->vport_id = switch_info->representor ?
875 			 switch_info->port_name + 1 : -1;
876 #endif
877 	/* representor_id field keeps the unmodified VF index. */
878 	priv->representor_id = switch_info->representor ?
879 			       switch_info->port_name : -1;
880 	/*
881 	 * Look for sibling devices in order to reuse their switch domain
882 	 * if any, otherwise allocate one.
883 	 */
884 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
885 		const struct mlx5_priv *opriv =
886 			rte_eth_devices[port_id].data->dev_private;
887 
888 		if (!opriv ||
889 		    opriv->sh != priv->sh ||
890 			opriv->domain_id ==
891 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
892 			continue;
893 		priv->domain_id = opriv->domain_id;
894 		break;
895 	}
896 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
897 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
898 		if (err) {
899 			err = rte_errno;
900 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
901 				strerror(rte_errno));
902 			goto error;
903 		}
904 		own_domain_id = 1;
905 	}
906 	/* Override some values set by hardware configuration. */
907 	mlx5_args(config, dpdk_dev->devargs);
908 	err = mlx5_dev_check_sibling_config(priv, config);
909 	if (err)
910 		goto error;
911 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
912 			    IBV_DEVICE_RAW_IP_CSUM);
913 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
914 		(config->hw_csum ? "" : "not "));
915 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
916 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
917 	DRV_LOG(DEBUG, "counters are not supported");
918 #endif
919 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
920 	if (config->dv_flow_en) {
921 		DRV_LOG(WARNING, "DV flow is not supported");
922 		config->dv_flow_en = 0;
923 	}
924 #endif
925 	config->ind_table_max_size =
926 		sh->device_attr.max_rwq_indirection_table_size;
927 	/*
928 	 * Remove this check once DPDK supports larger/variable
929 	 * indirection tables.
930 	 */
931 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
932 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
933 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
934 		config->ind_table_max_size);
935 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
936 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
937 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
938 		(config->hw_vlan_strip ? "" : "not "));
939 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
940 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
941 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
942 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
943 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
944 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
945 			IBV_DEVICE_PCI_WRITE_END_PADDING);
946 #endif
947 	if (config->hw_padding && !hw_padding) {
948 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
949 		config->hw_padding = 0;
950 	} else if (config->hw_padding) {
951 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
952 	}
953 	config->tso = (sh->device_attr.max_tso > 0 &&
954 		      (sh->device_attr.tso_supported_qpts &
955 		       (1 << IBV_QPT_RAW_PACKET)));
956 	if (config->tso)
957 		config->tso_max_payload_sz = sh->device_attr.max_tso;
958 	/*
959 	 * MPW is disabled by default, while the Enhanced MPW is enabled
960 	 * by default.
961 	 */
962 	if (config->mps == MLX5_ARG_UNSET)
963 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
964 							  MLX5_MPW_DISABLED;
965 	else
966 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
967 	DRV_LOG(INFO, "%sMPS is %s",
968 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
969 		config->mps == MLX5_MPW ? "legacy " : "",
970 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
971 	if (config->cqe_comp && !cqe_comp) {
972 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
973 		config->cqe_comp = 0;
974 	}
975 	if (config->cqe_pad && !cqe_pad) {
976 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
977 		config->cqe_pad = 0;
978 	} else if (config->cqe_pad) {
979 		DRV_LOG(INFO, "Rx CQE padding is enabled");
980 	}
981 	if (config->devx) {
982 		priv->counter_fallback = 0;
983 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
984 		if (err) {
985 			err = -err;
986 			goto error;
987 		}
988 		if (!config->hca_attr.flow_counters_dump)
989 			priv->counter_fallback = 1;
990 #ifndef HAVE_IBV_DEVX_ASYNC
991 		priv->counter_fallback = 1;
992 #endif
993 		if (priv->counter_fallback)
994 			DRV_LOG(INFO, "Use fall-back DV counter management");
995 		/* Check for LRO support. */
996 		if (config->dest_tir && config->hca_attr.lro_cap &&
997 		    config->dv_flow_en) {
998 			/* TBD check tunnel lro caps. */
999 			config->lro.supported = config->hca_attr.lro_cap;
1000 			DRV_LOG(DEBUG, "Device supports LRO");
1001 			/*
1002 			 * If LRO timeout is not configured by application,
1003 			 * use the minimal supported value.
1004 			 */
1005 			if (!config->lro.timeout)
1006 				config->lro.timeout =
1007 				config->hca_attr.lro_timer_supported_periods[0];
1008 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1009 				config->lro.timeout);
1010 		}
1011 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1012 		if (config->hca_attr.qos.sup &&
1013 		    config->hca_attr.qos.srtcm_sup &&
1014 		    config->dv_flow_en) {
1015 			uint8_t reg_c_mask =
1016 				config->hca_attr.qos.flow_meter_reg_c_ids;
1017 			/*
1018 			 * Meter needs two REG_C's for color match and pre-sfx
1019 			 * flow match. Here get the REG_C for color match.
1020 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1021 			 */
1022 			reg_c_mask &= 0xfc;
1023 			if (__builtin_popcount(reg_c_mask) < 1) {
1024 				priv->mtr_en = 0;
1025 				DRV_LOG(WARNING, "No available register for"
1026 					" meter.");
1027 			} else {
1028 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1029 						      REG_C_0;
1030 				priv->mtr_en = 1;
1031 				priv->mtr_reg_share =
1032 				      config->hca_attr.qos.flow_meter_reg_share;
1033 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1034 					priv->mtr_color_reg);
1035 			}
1036 		}
1037 #endif
1038 	}
1039 	if (config->tx_pp) {
1040 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1041 			config->hca_attr.dev_freq_khz);
1042 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1043 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1044 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1045 			config->hca_attr.cross_channel ? "" : "not ");
1046 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1047 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1048 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1049 			config->hca_attr.non_wire_sq ? "" : "not ");
1050 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1051 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1052 			config->hca_attr.log_max_static_sq_wq);
1053 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1054 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1055 		if (!config->devx) {
1056 			DRV_LOG(ERR, "DevX is required for packet pacing");
1057 			err = ENODEV;
1058 			goto error;
1059 		}
1060 		if (!config->hca_attr.qos.packet_pacing) {
1061 			DRV_LOG(ERR, "Packet pacing is not supported");
1062 			err = ENODEV;
1063 			goto error;
1064 		}
1065 		if (!config->hca_attr.cross_channel) {
1066 			DRV_LOG(ERR, "Cross channel operations are"
1067 				     " required for packet pacing");
1068 			err = ENODEV;
1069 			goto error;
1070 		}
1071 		if (!config->hca_attr.wqe_index_ignore) {
1072 			DRV_LOG(ERR, "WQE index ignore feature is"
1073 				     " required for packet pacing");
1074 			err = ENODEV;
1075 			goto error;
1076 		}
1077 		if (!config->hca_attr.non_wire_sq) {
1078 			DRV_LOG(ERR, "Non-wire SQ feature is"
1079 				     " required for packet pacing");
1080 			err = ENODEV;
1081 			goto error;
1082 		}
1083 		if (!config->hca_attr.log_max_static_sq_wq) {
1084 			DRV_LOG(ERR, "Static WQE SQ feature is"
1085 				     " required for packet pacing");
1086 			err = ENODEV;
1087 			goto error;
1088 		}
1089 		if (!config->hca_attr.qos.wqe_rate_pp) {
1090 			DRV_LOG(ERR, "WQE rate mode is required"
1091 				     " for packet pacing");
1092 			err = ENODEV;
1093 			goto error;
1094 		}
1095 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1096 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1097 			     " can't create queues for packet pacing");
1098 		err = ENODEV;
1099 		goto error;
1100 #endif
1101 	}
1102 	if (config->devx) {
1103 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1104 
1105 		err = config->hca_attr.access_register_user ?
1106 			mlx5_devx_cmd_register_read
1107 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1108 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1109 		if (!err) {
1110 			uint32_t ts_mode;
1111 
1112 			/* MTUTC register is read successfully. */
1113 			ts_mode = MLX5_GET(register_mtutc, reg,
1114 					   time_stamp_mode);
1115 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1116 				config->rt_timestamp = 1;
1117 		} else {
1118 			/* Kernel does not support register reading. */
1119 			if (config->hca_attr.dev_freq_khz ==
1120 						 (NS_PER_S / MS_PER_S))
1121 				config->rt_timestamp = 1;
1122 		}
1123 	}
1124 	/*
1125 	 * If HW has bug working with tunnel packet decapsulation and
1126 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1127 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1128 	 */
1129 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1130 		config->hw_fcs_strip = 0;
1131 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1132 		(config->hw_fcs_strip ? "" : "not "));
1133 	if (config->mprq.enabled && mprq) {
1134 		if (config->mprq.stride_num_n &&
1135 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1136 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1137 			config->mprq.stride_num_n =
1138 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1139 						mprq_min_stride_num_n),
1140 					mprq_max_stride_num_n);
1141 			DRV_LOG(WARNING,
1142 				"the number of strides"
1143 				" for Multi-Packet RQ is out of range,"
1144 				" setting default value (%u)",
1145 				1 << config->mprq.stride_num_n);
1146 		}
1147 		if (config->mprq.stride_size_n &&
1148 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1149 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1150 			config->mprq.stride_size_n =
1151 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1152 						mprq_min_stride_size_n),
1153 					mprq_max_stride_size_n);
1154 			DRV_LOG(WARNING,
1155 				"the size of a stride"
1156 				" for Multi-Packet RQ is out of range,"
1157 				" setting default value (%u)",
1158 				1 << config->mprq.stride_size_n);
1159 		}
1160 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1161 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1162 	} else if (config->mprq.enabled && !mprq) {
1163 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1164 		config->mprq.enabled = 0;
1165 	}
1166 	if (config->max_dump_files_num == 0)
1167 		config->max_dump_files_num = 128;
1168 	eth_dev = rte_eth_dev_allocate(name);
1169 	if (eth_dev == NULL) {
1170 		DRV_LOG(ERR, "can not allocate rte ethdev");
1171 		err = ENOMEM;
1172 		goto error;
1173 	}
1174 	if (priv->representor) {
1175 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1176 		eth_dev->data->representor_id = priv->representor_id;
1177 	}
1178 	/*
1179 	 * Store associated network device interface index. This index
1180 	 * is permanent throughout the lifetime of device. So, we may store
1181 	 * the ifindex here and use the cached value further.
1182 	 */
1183 	MLX5_ASSERT(spawn->ifindex);
1184 	priv->if_index = spawn->ifindex;
1185 	if (priv->pf_bond >= 0 && priv->master) {
1186 		/* Get bond interface info */
1187 		err = mlx5_sysfs_bond_info(priv->if_index,
1188 				     &priv->bond_ifindex,
1189 				     priv->bond_name);
1190 		if (err)
1191 			DRV_LOG(ERR, "unable to get bond info: %s",
1192 				strerror(rte_errno));
1193 		else
1194 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1195 				priv->if_index, priv->bond_ifindex,
1196 				priv->bond_name);
1197 	}
1198 	eth_dev->data->dev_private = priv;
1199 	priv->dev_data = eth_dev->data;
1200 	eth_dev->data->mac_addrs = priv->mac;
1201 	eth_dev->device = dpdk_dev;
1202 	/* Configure the first MAC address by default. */
1203 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1204 		DRV_LOG(ERR,
1205 			"port %u cannot get MAC address, is mlx5_en"
1206 			" loaded? (errno: %s)",
1207 			eth_dev->data->port_id, strerror(rte_errno));
1208 		err = ENODEV;
1209 		goto error;
1210 	}
1211 	DRV_LOG(INFO,
1212 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1213 		eth_dev->data->port_id,
1214 		mac.addr_bytes[0], mac.addr_bytes[1],
1215 		mac.addr_bytes[2], mac.addr_bytes[3],
1216 		mac.addr_bytes[4], mac.addr_bytes[5]);
1217 #ifdef RTE_LIBRTE_MLX5_DEBUG
1218 	{
1219 		char ifname[IF_NAMESIZE];
1220 
1221 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1222 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1223 				eth_dev->data->port_id, ifname);
1224 		else
1225 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1226 				eth_dev->data->port_id);
1227 	}
1228 #endif
1229 	/* Get actual MTU if possible. */
1230 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1231 	if (err) {
1232 		err = rte_errno;
1233 		goto error;
1234 	}
1235 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1236 		priv->mtu);
1237 	/* Initialize burst functions to prevent crashes before link-up. */
1238 	eth_dev->rx_pkt_burst = removed_rx_burst;
1239 	eth_dev->tx_pkt_burst = removed_tx_burst;
1240 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1241 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1242 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1243 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1244 	/* Register MAC address. */
1245 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1246 	if (config->vf && config->vf_nl_en)
1247 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1248 				      mlx5_ifindex(eth_dev),
1249 				      eth_dev->data->mac_addrs,
1250 				      MLX5_MAX_MAC_ADDRESSES);
1251 	priv->flows = 0;
1252 	priv->ctrl_flows = 0;
1253 	TAILQ_INIT(&priv->flow_meters);
1254 	TAILQ_INIT(&priv->flow_meter_profiles);
1255 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1256 	mlx5_glue->dv_set_context_attr(sh->ctx,
1257 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1258 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1259 				.alloc = &mlx5_alloc_verbs_buf,
1260 				.free = &mlx5_free_verbs_buf,
1261 				.data = priv,
1262 			}));
1263 	/* Bring Ethernet device up. */
1264 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1265 		eth_dev->data->port_id);
1266 	mlx5_set_link_up(eth_dev);
1267 	/*
1268 	 * Even though the interrupt handler is not installed yet,
1269 	 * interrupts will still trigger on the async_fd from
1270 	 * Verbs context returned by ibv_open_device().
1271 	 */
1272 	mlx5_link_update(eth_dev, 0);
1273 #ifdef HAVE_MLX5DV_DR_ESWITCH
1274 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1275 	      (switch_info->representor || switch_info->master)))
1276 		config->dv_esw_en = 0;
1277 #else
1278 	config->dv_esw_en = 0;
1279 #endif
1280 	/* Detect minimal data bytes to inline. */
1281 	mlx5_set_min_inline(spawn, config);
1282 	/* Store device configuration on private structure. */
1283 	priv->config = *config;
1284 	/* Create context for virtual machine VLAN workaround. */
1285 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1286 	if (config->dv_flow_en) {
1287 		err = mlx5_alloc_shared_dr(priv);
1288 		if (err)
1289 			goto error;
1290 		/*
1291 		 * RSS id is shared with meter flow id. Meter flow id can only
1292 		 * use the 24 MSB of the register.
1293 		 */
1294 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1295 				     MLX5_MTR_COLOR_BITS);
1296 		if (!priv->qrss_id_pool) {
1297 			DRV_LOG(ERR, "can't create flow id pool");
1298 			err = ENOMEM;
1299 			goto error;
1300 		}
1301 	}
1302 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1303 		priv->obj_ops = devx_obj_ops;
1304 		priv->obj_ops.drop_action_create =
1305 						ibv_obj_ops.drop_action_create;
1306 		priv->obj_ops.drop_action_destroy =
1307 						ibv_obj_ops.drop_action_destroy;
1308 	} else {
1309 		priv->obj_ops = ibv_obj_ops;
1310 	}
1311 	/* Supported Verbs flow priority number detection. */
1312 	err = mlx5_flow_discover_priorities(eth_dev);
1313 	if (err < 0) {
1314 		err = -err;
1315 		goto error;
1316 	}
1317 	priv->config.flow_prio = err;
1318 	if (!priv->config.dv_esw_en &&
1319 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1320 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1321 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1322 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1323 	}
1324 	mlx5_set_metadata_mask(eth_dev);
1325 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1326 	    !priv->sh->dv_regc0_mask) {
1327 		DRV_LOG(ERR, "metadata mode %u is not supported "
1328 			     "(no metadata reg_c[0] is available)",
1329 			     priv->config.dv_xmeta_en);
1330 			err = ENOTSUP;
1331 			goto error;
1332 	}
1333 	/*
1334 	 * Allocate the buffer for flow creating, just once.
1335 	 * The allocation must be done before any flow creating.
1336 	 */
1337 	mlx5_flow_alloc_intermediate(eth_dev);
1338 	/* Query availability of metadata reg_c's. */
1339 	err = mlx5_flow_discover_mreg_c(eth_dev);
1340 	if (err < 0) {
1341 		err = -err;
1342 		goto error;
1343 	}
1344 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1345 		DRV_LOG(DEBUG,
1346 			"port %u extensive metadata register is not supported",
1347 			eth_dev->data->port_id);
1348 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1349 			DRV_LOG(ERR, "metadata mode %u is not supported "
1350 				     "(no metadata registers available)",
1351 				     priv->config.dv_xmeta_en);
1352 			err = ENOTSUP;
1353 			goto error;
1354 		}
1355 	}
1356 	if (priv->config.dv_flow_en &&
1357 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1358 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1359 	    priv->sh->dv_regc0_mask) {
1360 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1361 						      MLX5_FLOW_MREG_HTABLE_SZ);
1362 		if (!priv->mreg_cp_tbl) {
1363 			err = ENOMEM;
1364 			goto error;
1365 		}
1366 	}
1367 	return eth_dev;
1368 error:
1369 	if (priv) {
1370 		if (priv->mreg_cp_tbl)
1371 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1372 		if (priv->sh)
1373 			mlx5_os_free_shared_dr(priv);
1374 		if (priv->nl_socket_route >= 0)
1375 			close(priv->nl_socket_route);
1376 		if (priv->nl_socket_rdma >= 0)
1377 			close(priv->nl_socket_rdma);
1378 		if (priv->vmwa_context)
1379 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1380 		if (priv->qrss_id_pool)
1381 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1382 		if (own_domain_id)
1383 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1384 		mlx5_free(priv);
1385 		if (eth_dev != NULL)
1386 			eth_dev->data->dev_private = NULL;
1387 	}
1388 	if (eth_dev != NULL) {
1389 		/* mac_addrs must not be freed alone because part of
1390 		 * dev_private
1391 		 **/
1392 		eth_dev->data->mac_addrs = NULL;
1393 		rte_eth_dev_release_port(eth_dev);
1394 	}
1395 	if (sh)
1396 		mlx5_free_shared_dev_ctx(sh);
1397 	MLX5_ASSERT(err > 0);
1398 	rte_errno = err;
1399 	return NULL;
1400 }
1401 
1402 /**
1403  * Comparison callback to sort device data.
1404  *
1405  * This is meant to be used with qsort().
1406  *
1407  * @param a[in]
1408  *   Pointer to pointer to first data object.
1409  * @param b[in]
1410  *   Pointer to pointer to second data object.
1411  *
1412  * @return
1413  *   0 if both objects are equal, less than 0 if the first argument is less
1414  *   than the second, greater than 0 otherwise.
1415  */
1416 static int
1417 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1418 {
1419 	const struct mlx5_switch_info *si_a =
1420 		&((const struct mlx5_dev_spawn_data *)a)->info;
1421 	const struct mlx5_switch_info *si_b =
1422 		&((const struct mlx5_dev_spawn_data *)b)->info;
1423 	int ret;
1424 
1425 	/* Master device first. */
1426 	ret = si_b->master - si_a->master;
1427 	if (ret)
1428 		return ret;
1429 	/* Then representor devices. */
1430 	ret = si_b->representor - si_a->representor;
1431 	if (ret)
1432 		return ret;
1433 	/* Unidentified devices come last in no specific order. */
1434 	if (!si_a->representor)
1435 		return 0;
1436 	/* Order representors by name. */
1437 	return si_a->port_name - si_b->port_name;
1438 }
1439 
1440 /**
1441  * Match PCI information for possible slaves of bonding device.
1442  *
1443  * @param[in] ibv_dev
1444  *   Pointer to Infiniband device structure.
1445  * @param[in] pci_dev
1446  *   Pointer to PCI device structure to match PCI address.
1447  * @param[in] nl_rdma
1448  *   Netlink RDMA group socket handle.
1449  *
1450  * @return
1451  *   negative value if no bonding device found, otherwise
1452  *   positive index of slave PF in bonding.
1453  */
1454 static int
1455 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1456 			   const struct rte_pci_device *pci_dev,
1457 			   int nl_rdma)
1458 {
1459 	char ifname[IF_NAMESIZE + 1];
1460 	unsigned int ifindex;
1461 	unsigned int np, i;
1462 	FILE *file = NULL;
1463 	int pf = -1;
1464 
1465 	/*
1466 	 * Try to get master device name. If something goes
1467 	 * wrong suppose the lack of kernel support and no
1468 	 * bonding devices.
1469 	 */
1470 	if (nl_rdma < 0)
1471 		return -1;
1472 	if (!strstr(ibv_dev->name, "bond"))
1473 		return -1;
1474 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1475 	if (!np)
1476 		return -1;
1477 	/*
1478 	 * The Master device might not be on the predefined
1479 	 * port (not on port index 1, it is not garanted),
1480 	 * we have to scan all Infiniband device port and
1481 	 * find master.
1482 	 */
1483 	for (i = 1; i <= np; ++i) {
1484 		/* Check whether Infiniband port is populated. */
1485 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1486 		if (!ifindex)
1487 			continue;
1488 		if (!if_indextoname(ifindex, ifname))
1489 			continue;
1490 		/* Try to read bonding slave names from sysfs. */
1491 		MKSTR(slaves,
1492 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1493 		file = fopen(slaves, "r");
1494 		if (file)
1495 			break;
1496 	}
1497 	if (!file)
1498 		return -1;
1499 	/* Use safe format to check maximal buffer length. */
1500 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1501 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1502 		char tmp_str[IF_NAMESIZE + 32];
1503 		struct rte_pci_addr pci_addr;
1504 		struct mlx5_switch_info	info;
1505 
1506 		/* Process slave interface names in the loop. */
1507 		snprintf(tmp_str, sizeof(tmp_str),
1508 			 "/sys/class/net/%s", ifname);
1509 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1510 			DRV_LOG(WARNING, "can not get PCI address"
1511 					 " for netdev \"%s\"", ifname);
1512 			continue;
1513 		}
1514 		if (pci_dev->addr.domain != pci_addr.domain ||
1515 		    pci_dev->addr.bus != pci_addr.bus ||
1516 		    pci_dev->addr.devid != pci_addr.devid ||
1517 		    pci_dev->addr.function != pci_addr.function)
1518 			continue;
1519 		/* Slave interface PCI address match found. */
1520 		fclose(file);
1521 		snprintf(tmp_str, sizeof(tmp_str),
1522 			 "/sys/class/net/%s/phys_port_name", ifname);
1523 		file = fopen(tmp_str, "rb");
1524 		if (!file)
1525 			break;
1526 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1527 		if (fscanf(file, "%32s", tmp_str) == 1)
1528 			mlx5_translate_port_name(tmp_str, &info);
1529 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1530 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1531 			pf = info.port_name;
1532 		break;
1533 	}
1534 	if (file)
1535 		fclose(file);
1536 	return pf;
1537 }
1538 
1539 /**
1540  * DPDK callback to register a PCI device.
1541  *
1542  * This function spawns Ethernet devices out of a given PCI device.
1543  *
1544  * @param[in] pci_drv
1545  *   PCI driver structure (mlx5_driver).
1546  * @param[in] pci_dev
1547  *   PCI device information.
1548  *
1549  * @return
1550  *   0 on success, a negative errno value otherwise and rte_errno is set.
1551  */
1552 int
1553 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1554 		  struct rte_pci_device *pci_dev)
1555 {
1556 	struct ibv_device **ibv_list;
1557 	/*
1558 	 * Number of found IB Devices matching with requested PCI BDF.
1559 	 * nd != 1 means there are multiple IB devices over the same
1560 	 * PCI device and we have representors and master.
1561 	 */
1562 	unsigned int nd = 0;
1563 	/*
1564 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1565 	 * we have the single multiport IB device, and there may be
1566 	 * representors attached to some of found ports.
1567 	 */
1568 	unsigned int np = 0;
1569 	/*
1570 	 * Number of DPDK ethernet devices to Spawn - either over
1571 	 * multiple IB devices or multiple ports of single IB device.
1572 	 * Actually this is the number of iterations to spawn.
1573 	 */
1574 	unsigned int ns = 0;
1575 	/*
1576 	 * Bonding device
1577 	 *   < 0 - no bonding device (single one)
1578 	 *  >= 0 - bonding device (value is slave PF index)
1579 	 */
1580 	int bd = -1;
1581 	struct mlx5_dev_spawn_data *list = NULL;
1582 	struct mlx5_dev_config dev_config;
1583 	unsigned int dev_config_vf;
1584 	int ret;
1585 
1586 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1587 		mlx5_pmd_socket_init();
1588 	ret = mlx5_init_once();
1589 	if (ret) {
1590 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1591 			strerror(rte_errno));
1592 		return -rte_errno;
1593 	}
1594 	errno = 0;
1595 	ibv_list = mlx5_glue->get_device_list(&ret);
1596 	if (!ibv_list) {
1597 		rte_errno = errno ? errno : ENOSYS;
1598 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1599 		return -rte_errno;
1600 	}
1601 	/*
1602 	 * First scan the list of all Infiniband devices to find
1603 	 * matching ones, gathering into the list.
1604 	 */
1605 	struct ibv_device *ibv_match[ret + 1];
1606 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1607 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1608 	unsigned int i;
1609 
1610 	while (ret-- > 0) {
1611 		struct rte_pci_addr pci_addr;
1612 
1613 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1614 		bd = mlx5_device_bond_pci_match
1615 				(ibv_list[ret], pci_dev, nl_rdma);
1616 		if (bd >= 0) {
1617 			/*
1618 			 * Bonding device detected. Only one match is allowed,
1619 			 * the bonding is supported over multi-port IB device,
1620 			 * there should be no matches on representor PCI
1621 			 * functions or non VF LAG bonding devices with
1622 			 * specified address.
1623 			 */
1624 			if (nd) {
1625 				DRV_LOG(ERR,
1626 					"multiple PCI match on bonding device"
1627 					"\"%s\" found", ibv_list[ret]->name);
1628 				rte_errno = ENOENT;
1629 				ret = -rte_errno;
1630 				goto exit;
1631 			}
1632 			DRV_LOG(INFO, "PCI information matches for"
1633 				      " slave %d bonding device \"%s\"",
1634 				      bd, ibv_list[ret]->name);
1635 			ibv_match[nd++] = ibv_list[ret];
1636 			break;
1637 		}
1638 		if (mlx5_dev_to_pci_addr
1639 			(ibv_list[ret]->ibdev_path, &pci_addr))
1640 			continue;
1641 		if (pci_dev->addr.domain != pci_addr.domain ||
1642 		    pci_dev->addr.bus != pci_addr.bus ||
1643 		    pci_dev->addr.devid != pci_addr.devid ||
1644 		    pci_dev->addr.function != pci_addr.function)
1645 			continue;
1646 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1647 			ibv_list[ret]->name);
1648 		ibv_match[nd++] = ibv_list[ret];
1649 	}
1650 	ibv_match[nd] = NULL;
1651 	if (!nd) {
1652 		/* No device matches, just complain and bail out. */
1653 		DRV_LOG(WARNING,
1654 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1655 			" are kernel drivers loaded?",
1656 			pci_dev->addr.domain, pci_dev->addr.bus,
1657 			pci_dev->addr.devid, pci_dev->addr.function);
1658 		rte_errno = ENOENT;
1659 		ret = -rte_errno;
1660 		goto exit;
1661 	}
1662 	if (nd == 1) {
1663 		/*
1664 		 * Found single matching device may have multiple ports.
1665 		 * Each port may be representor, we have to check the port
1666 		 * number and check the representors existence.
1667 		 */
1668 		if (nl_rdma >= 0)
1669 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1670 		if (!np)
1671 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1672 					 " ports number", ibv_match[0]->name);
1673 		if (bd >= 0 && !np) {
1674 			DRV_LOG(ERR, "can not get ports"
1675 				     " for bonding device");
1676 			rte_errno = ENOENT;
1677 			ret = -rte_errno;
1678 			goto exit;
1679 		}
1680 	}
1681 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1682 	if (bd >= 0) {
1683 		/*
1684 		 * This may happen if there is VF LAG kernel support and
1685 		 * application is compiled with older rdma_core library.
1686 		 */
1687 		DRV_LOG(ERR,
1688 			"No kernel/verbs support for VF LAG bonding found.");
1689 		rte_errno = ENOTSUP;
1690 		ret = -rte_errno;
1691 		goto exit;
1692 	}
1693 #endif
1694 	/*
1695 	 * Now we can determine the maximal
1696 	 * amount of devices to be spawned.
1697 	 */
1698 	list = mlx5_malloc(MLX5_MEM_ZERO,
1699 			   sizeof(struct mlx5_dev_spawn_data) *
1700 			   (np ? np : nd),
1701 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1702 	if (!list) {
1703 		DRV_LOG(ERR, "spawn data array allocation failure");
1704 		rte_errno = ENOMEM;
1705 		ret = -rte_errno;
1706 		goto exit;
1707 	}
1708 	if (bd >= 0 || np > 1) {
1709 		/*
1710 		 * Single IB device with multiple ports found,
1711 		 * it may be E-Switch master device and representors.
1712 		 * We have to perform identification through the ports.
1713 		 */
1714 		MLX5_ASSERT(nl_rdma >= 0);
1715 		MLX5_ASSERT(ns == 0);
1716 		MLX5_ASSERT(nd == 1);
1717 		MLX5_ASSERT(np);
1718 		for (i = 1; i <= np; ++i) {
1719 			list[ns].max_port = np;
1720 			list[ns].phys_port = i;
1721 			list[ns].phys_dev = ibv_match[0];
1722 			list[ns].eth_dev = NULL;
1723 			list[ns].pci_dev = pci_dev;
1724 			list[ns].pf_bond = bd;
1725 			list[ns].ifindex = mlx5_nl_ifindex
1726 				(nl_rdma,
1727 				mlx5_os_get_dev_device_name
1728 						(list[ns].phys_dev), i);
1729 			if (!list[ns].ifindex) {
1730 				/*
1731 				 * No network interface index found for the
1732 				 * specified port, it means there is no
1733 				 * representor on this port. It's OK,
1734 				 * there can be disabled ports, for example
1735 				 * if sriov_numvfs < sriov_totalvfs.
1736 				 */
1737 				continue;
1738 			}
1739 			ret = -1;
1740 			if (nl_route >= 0)
1741 				ret = mlx5_nl_switch_info
1742 					       (nl_route,
1743 						list[ns].ifindex,
1744 						&list[ns].info);
1745 			if (ret || (!list[ns].info.representor &&
1746 				    !list[ns].info.master)) {
1747 				/*
1748 				 * We failed to recognize representors with
1749 				 * Netlink, let's try to perform the task
1750 				 * with sysfs.
1751 				 */
1752 				ret =  mlx5_sysfs_switch_info
1753 						(list[ns].ifindex,
1754 						 &list[ns].info);
1755 			}
1756 			if (!ret && bd >= 0) {
1757 				switch (list[ns].info.name_type) {
1758 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1759 					if (list[ns].info.port_name == bd)
1760 						ns++;
1761 					break;
1762 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1763 					/* Fallthrough */
1764 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1765 					if (list[ns].info.pf_num == bd)
1766 						ns++;
1767 					break;
1768 				default:
1769 					break;
1770 				}
1771 				continue;
1772 			}
1773 			if (!ret && (list[ns].info.representor ^
1774 				     list[ns].info.master))
1775 				ns++;
1776 		}
1777 		if (!ns) {
1778 			DRV_LOG(ERR,
1779 				"unable to recognize master/representors"
1780 				" on the IB device with multiple ports");
1781 			rte_errno = ENOENT;
1782 			ret = -rte_errno;
1783 			goto exit;
1784 		}
1785 	} else {
1786 		/*
1787 		 * The existence of several matching entries (nd > 1) means
1788 		 * port representors have been instantiated. No existing Verbs
1789 		 * call nor sysfs entries can tell them apart, this can only
1790 		 * be done through Netlink calls assuming kernel drivers are
1791 		 * recent enough to support them.
1792 		 *
1793 		 * In the event of identification failure through Netlink,
1794 		 * try again through sysfs, then:
1795 		 *
1796 		 * 1. A single IB device matches (nd == 1) with single
1797 		 *    port (np=0/1) and is not a representor, assume
1798 		 *    no switch support.
1799 		 *
1800 		 * 2. Otherwise no safe assumptions can be made;
1801 		 *    complain louder and bail out.
1802 		 */
1803 		for (i = 0; i != nd; ++i) {
1804 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1805 			list[ns].max_port = 1;
1806 			list[ns].phys_port = 1;
1807 			list[ns].phys_dev = ibv_match[i];
1808 			list[ns].eth_dev = NULL;
1809 			list[ns].pci_dev = pci_dev;
1810 			list[ns].pf_bond = -1;
1811 			list[ns].ifindex = 0;
1812 			if (nl_rdma >= 0)
1813 				list[ns].ifindex = mlx5_nl_ifindex
1814 				(nl_rdma,
1815 				mlx5_os_get_dev_device_name
1816 						(list[ns].phys_dev), 1);
1817 			if (!list[ns].ifindex) {
1818 				char ifname[IF_NAMESIZE];
1819 
1820 				/*
1821 				 * Netlink failed, it may happen with old
1822 				 * ib_core kernel driver (before 4.16).
1823 				 * We can assume there is old driver because
1824 				 * here we are processing single ports IB
1825 				 * devices. Let's try sysfs to retrieve
1826 				 * the ifindex. The method works for
1827 				 * master device only.
1828 				 */
1829 				if (nd > 1) {
1830 					/*
1831 					 * Multiple devices found, assume
1832 					 * representors, can not distinguish
1833 					 * master/representor and retrieve
1834 					 * ifindex via sysfs.
1835 					 */
1836 					continue;
1837 				}
1838 				ret = mlx5_get_ifname_sysfs
1839 					(ibv_match[i]->ibdev_path, ifname);
1840 				if (!ret)
1841 					list[ns].ifindex =
1842 						if_nametoindex(ifname);
1843 				if (!list[ns].ifindex) {
1844 					/*
1845 					 * No network interface index found
1846 					 * for the specified device, it means
1847 					 * there it is neither representor
1848 					 * nor master.
1849 					 */
1850 					continue;
1851 				}
1852 			}
1853 			ret = -1;
1854 			if (nl_route >= 0)
1855 				ret = mlx5_nl_switch_info
1856 					       (nl_route,
1857 						list[ns].ifindex,
1858 						&list[ns].info);
1859 			if (ret || (!list[ns].info.representor &&
1860 				    !list[ns].info.master)) {
1861 				/*
1862 				 * We failed to recognize representors with
1863 				 * Netlink, let's try to perform the task
1864 				 * with sysfs.
1865 				 */
1866 				ret =  mlx5_sysfs_switch_info
1867 						(list[ns].ifindex,
1868 						 &list[ns].info);
1869 			}
1870 			if (!ret && (list[ns].info.representor ^
1871 				     list[ns].info.master)) {
1872 				ns++;
1873 			} else if ((nd == 1) &&
1874 				   !list[ns].info.representor &&
1875 				   !list[ns].info.master) {
1876 				/*
1877 				 * Single IB device with
1878 				 * one physical port and
1879 				 * attached network device.
1880 				 * May be SRIOV is not enabled
1881 				 * or there is no representors.
1882 				 */
1883 				DRV_LOG(INFO, "no E-Switch support detected");
1884 				ns++;
1885 				break;
1886 			}
1887 		}
1888 		if (!ns) {
1889 			DRV_LOG(ERR,
1890 				"unable to recognize master/representors"
1891 				" on the multiple IB devices");
1892 			rte_errno = ENOENT;
1893 			ret = -rte_errno;
1894 			goto exit;
1895 		}
1896 	}
1897 	MLX5_ASSERT(ns);
1898 	/*
1899 	 * Sort list to probe devices in natural order for users convenience
1900 	 * (i.e. master first, then representors from lowest to highest ID).
1901 	 */
1902 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1903 	/* Device specific configuration. */
1904 	switch (pci_dev->id.device_id) {
1905 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1906 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1907 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1908 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1909 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1910 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1911 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1912 		dev_config_vf = 1;
1913 		break;
1914 	default:
1915 		dev_config_vf = 0;
1916 		break;
1917 	}
1918 	for (i = 0; i != ns; ++i) {
1919 		uint32_t restore;
1920 
1921 		/* Default configuration. */
1922 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1923 		dev_config.vf = dev_config_vf;
1924 		dev_config.mps = MLX5_ARG_UNSET;
1925 		dev_config.dbnc = MLX5_ARG_UNSET;
1926 		dev_config.rx_vec_en = 1;
1927 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
1928 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
1929 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
1930 		dev_config.txqs_inline = MLX5_ARG_UNSET;
1931 		dev_config.vf_nl_en = 1;
1932 		dev_config.mr_ext_memseg_en = 1;
1933 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1934 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1935 		dev_config.dv_esw_en = 1;
1936 		dev_config.dv_flow_en = 1;
1937 		dev_config.decap_en = 1;
1938 		dev_config.log_hp_size = MLX5_ARG_UNSET;
1939 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1940 						 &list[i],
1941 						 &dev_config);
1942 		if (!list[i].eth_dev) {
1943 			if (rte_errno != EBUSY && rte_errno != EEXIST)
1944 				break;
1945 			/* Device is disabled or already spawned. Ignore it. */
1946 			continue;
1947 		}
1948 		restore = list[i].eth_dev->data->dev_flags;
1949 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1950 		/* Restore non-PCI flags cleared by the above call. */
1951 		list[i].eth_dev->data->dev_flags |= restore;
1952 		rte_eth_dev_probing_finish(list[i].eth_dev);
1953 	}
1954 	if (i != ns) {
1955 		DRV_LOG(ERR,
1956 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1957 			" encountering an error: %s",
1958 			pci_dev->addr.domain, pci_dev->addr.bus,
1959 			pci_dev->addr.devid, pci_dev->addr.function,
1960 			strerror(rte_errno));
1961 		ret = -rte_errno;
1962 		/* Roll back. */
1963 		while (i--) {
1964 			if (!list[i].eth_dev)
1965 				continue;
1966 			mlx5_dev_close(list[i].eth_dev);
1967 			/* mac_addrs must not be freed because in dev_private */
1968 			list[i].eth_dev->data->mac_addrs = NULL;
1969 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1970 		}
1971 		/* Restore original error. */
1972 		rte_errno = -ret;
1973 	} else {
1974 		ret = 0;
1975 	}
1976 exit:
1977 	/*
1978 	 * Do the routine cleanup:
1979 	 * - close opened Netlink sockets
1980 	 * - free allocated spawn data array
1981 	 * - free the Infiniband device list
1982 	 */
1983 	if (nl_rdma >= 0)
1984 		close(nl_rdma);
1985 	if (nl_route >= 0)
1986 		close(nl_route);
1987 	if (list)
1988 		mlx5_free(list);
1989 	MLX5_ASSERT(ibv_list);
1990 	mlx5_glue->free_device_list(ibv_list);
1991 	return ret;
1992 }
1993 
1994 static int
1995 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1996 {
1997 	char *env;
1998 	int value;
1999 
2000 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2001 	/* Get environment variable to store. */
2002 	env = getenv(MLX5_SHUT_UP_BF);
2003 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2004 	if (config->dbnc == MLX5_ARG_UNSET)
2005 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2006 	else
2007 		setenv(MLX5_SHUT_UP_BF,
2008 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2009 	return value;
2010 }
2011 
2012 static void
2013 mlx5_restore_doorbell_mapping_env(int value)
2014 {
2015 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2016 	/* Restore the original environment variable state. */
2017 	if (value == MLX5_ARG_UNSET)
2018 		unsetenv(MLX5_SHUT_UP_BF);
2019 	else
2020 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2021 }
2022 
2023 /**
2024  * Extract pdn of PD object using DV API.
2025  *
2026  * @param[in] pd
2027  *   Pointer to the verbs PD object.
2028  * @param[out] pdn
2029  *   Pointer to the PD object number variable.
2030  *
2031  * @return
2032  *   0 on success, error value otherwise.
2033  */
2034 int
2035 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2036 {
2037 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2038 	struct mlx5dv_obj obj;
2039 	struct mlx5dv_pd pd_info;
2040 	int ret = 0;
2041 
2042 	obj.pd.in = pd;
2043 	obj.pd.out = &pd_info;
2044 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2045 	if (ret) {
2046 		DRV_LOG(DEBUG, "Fail to get PD object info");
2047 		return ret;
2048 	}
2049 	*pdn = pd_info.pdn;
2050 	return 0;
2051 #else
2052 	(void)pd;
2053 	(void)pdn;
2054 	return -ENOTSUP;
2055 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2056 }
2057 
2058 /**
2059  * Function API to open IB device.
2060  *
2061  * This function calls the Linux glue APIs to open a device.
2062  *
2063  * @param[in] spawn
2064  *   Pointer to the IB device attributes (name, port, etc).
2065  * @param[out] config
2066  *   Pointer to device configuration structure.
2067  * @param[out] sh
2068  *   Pointer to shared context structure.
2069  *
2070  * @return
2071  *   0 on success, a positive error value otherwise.
2072  */
2073 int
2074 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2075 		     const struct mlx5_dev_config *config,
2076 		     struct mlx5_dev_ctx_shared *sh)
2077 {
2078 	int dbmap_env;
2079 	int err = 0;
2080 
2081 	sh->numa_node = spawn->pci_dev->device.numa_node;
2082 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2083 	/*
2084 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2085 	 * before the device creation. The rdma_core library
2086 	 * checks the variable at device creation and
2087 	 * stores the result internally.
2088 	 */
2089 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2090 	/* Try to open IB device with DV first, then usual Verbs. */
2091 	errno = 0;
2092 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2093 	if (sh->ctx) {
2094 		sh->devx = 1;
2095 		DRV_LOG(DEBUG, "DevX is supported");
2096 		/* The device is created, no need for environment. */
2097 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2098 	} else {
2099 		/* The environment variable is still configured. */
2100 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2101 		err = errno ? errno : ENODEV;
2102 		/*
2103 		 * The environment variable is not needed anymore,
2104 		 * all device creation attempts are completed.
2105 		 */
2106 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2107 		if (!sh->ctx)
2108 			return err;
2109 		DRV_LOG(DEBUG, "DevX is NOT supported");
2110 		err = 0;
2111 	}
2112 	return err;
2113 }
2114 
2115 /**
2116  * Install shared asynchronous device events handler.
2117  * This function is implemented to support event sharing
2118  * between multiple ports of single IB device.
2119  *
2120  * @param sh
2121  *   Pointer to mlx5_dev_ctx_shared object.
2122  */
2123 void
2124 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2125 {
2126 	int ret;
2127 	int flags;
2128 
2129 	sh->intr_handle.fd = -1;
2130 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2131 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2132 		    F_SETFL, flags | O_NONBLOCK);
2133 	if (ret) {
2134 		DRV_LOG(INFO, "failed to change file descriptor async event"
2135 			" queue");
2136 	} else {
2137 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2138 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2139 		if (rte_intr_callback_register(&sh->intr_handle,
2140 					mlx5_dev_interrupt_handler, sh)) {
2141 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2142 			sh->intr_handle.fd = -1;
2143 		}
2144 	}
2145 	if (sh->devx) {
2146 #ifdef HAVE_IBV_DEVX_ASYNC
2147 		sh->intr_handle_devx.fd = -1;
2148 		sh->devx_comp =
2149 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2150 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2151 		if (!devx_comp) {
2152 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2153 			return;
2154 		}
2155 		flags = fcntl(devx_comp->fd, F_GETFL);
2156 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2157 		if (ret) {
2158 			DRV_LOG(INFO, "failed to change file descriptor"
2159 				" devx comp");
2160 			return;
2161 		}
2162 		sh->intr_handle_devx.fd = devx_comp->fd;
2163 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2164 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2165 					mlx5_dev_interrupt_handler_devx, sh)) {
2166 			DRV_LOG(INFO, "Fail to install the devx shared"
2167 				" interrupt.");
2168 			sh->intr_handle_devx.fd = -1;
2169 		}
2170 #endif /* HAVE_IBV_DEVX_ASYNC */
2171 	}
2172 }
2173 
2174 /**
2175  * Uninstall shared asynchronous device events handler.
2176  * This function is implemented to support event sharing
2177  * between multiple ports of single IB device.
2178  *
2179  * @param dev
2180  *   Pointer to mlx5_dev_ctx_shared object.
2181  */
2182 void
2183 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2184 {
2185 	if (sh->intr_handle.fd >= 0)
2186 		mlx5_intr_callback_unregister(&sh->intr_handle,
2187 					      mlx5_dev_interrupt_handler, sh);
2188 #ifdef HAVE_IBV_DEVX_ASYNC
2189 	if (sh->intr_handle_devx.fd >= 0)
2190 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2191 				  mlx5_dev_interrupt_handler_devx, sh);
2192 	if (sh->devx_comp)
2193 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2194 #endif
2195 }
2196 
2197 /**
2198  * Read statistics by a named counter.
2199  *
2200  * @param[in] priv
2201  *   Pointer to the private device data structure.
2202  * @param[in] ctr_name
2203  *   Pointer to the name of the statistic counter to read
2204  * @param[out] stat
2205  *   Pointer to read statistic value.
2206  * @return
2207  *   0 on success and stat is valud, 1 if failed to read the value
2208  *   rte_errno is set.
2209  *
2210  */
2211 int
2212 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2213 		      uint64_t *stat)
2214 {
2215 	int fd;
2216 
2217 	if (priv->sh) {
2218 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2219 		      priv->sh->ibdev_path,
2220 		      priv->dev_port,
2221 		      ctr_name);
2222 		fd = open(path, O_RDONLY);
2223 		/*
2224 		 * in switchdev the file location is not per port
2225 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2226 		 */
2227 		if (fd == -1) {
2228 			MKSTR(path1, "%s/hw_counters/%s",
2229 			      priv->sh->ibdev_path,
2230 			      ctr_name);
2231 			fd = open(path1, O_RDONLY);
2232 		}
2233 		if (fd != -1) {
2234 			char buf[21] = {'\0'};
2235 			ssize_t n = read(fd, buf, sizeof(buf));
2236 
2237 			close(fd);
2238 			if (n != -1) {
2239 				*stat = strtoull(buf, NULL, 10);
2240 				return 0;
2241 			}
2242 		}
2243 	}
2244 	*stat = 0;
2245 	return 1;
2246 }
2247 
2248 /**
2249  * Set the reg_mr and dereg_mr call backs
2250  *
2251  * @param reg_mr_cb[out]
2252  *   Pointer to reg_mr func
2253  * @param dereg_mr_cb[out]
2254  *   Pointer to dereg_mr func
2255  *
2256  */
2257 void
2258 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2259 		      mlx5_dereg_mr_t *dereg_mr_cb)
2260 {
2261 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2262 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2263 }
2264 
2265 /**
2266  * Remove a MAC address from device
2267  *
2268  * @param dev
2269  *   Pointer to Ethernet device structure.
2270  * @param index
2271  *   MAC address index.
2272  */
2273 void
2274 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2275 {
2276 	struct mlx5_priv *priv = dev->data->dev_private;
2277 	const int vf = priv->config.vf;
2278 
2279 	if (vf)
2280 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2281 					mlx5_ifindex(dev), priv->mac_own,
2282 					&dev->data->mac_addrs[index], index);
2283 }
2284 
2285 /**
2286  * Adds a MAC address to the device
2287  *
2288  * @param dev
2289  *   Pointer to Ethernet device structure.
2290  * @param mac_addr
2291  *   MAC address to register.
2292  * @param index
2293  *   MAC address index.
2294  *
2295  * @return
2296  *   0 on success, a negative errno value otherwise
2297  */
2298 int
2299 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2300 		     uint32_t index)
2301 {
2302 	struct mlx5_priv *priv = dev->data->dev_private;
2303 	const int vf = priv->config.vf;
2304 	int ret = 0;
2305 
2306 	if (vf)
2307 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2308 					   mlx5_ifindex(dev), priv->mac_own,
2309 					   mac, index);
2310 	return ret;
2311 }
2312 
2313 /**
2314  * Modify a VF MAC address
2315  *
2316  * @param priv
2317  *   Pointer to device private data.
2318  * @param mac_addr
2319  *   MAC address to modify into.
2320  * @param iface_idx
2321  *   Net device interface index
2322  * @param vf_index
2323  *   VF index
2324  *
2325  * @return
2326  *   0 on success, a negative errno value otherwise
2327  */
2328 int
2329 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2330 			   unsigned int iface_idx,
2331 			   struct rte_ether_addr *mac_addr,
2332 			   int vf_index)
2333 {
2334 	return mlx5_nl_vf_mac_addr_modify
2335 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2336 }
2337 
2338 /**
2339  * Set device promiscuous mode
2340  *
2341  * @param dev
2342  *   Pointer to Ethernet device structure.
2343  * @param enable
2344  *   0 - promiscuous is disabled, otherwise - enabled
2345  *
2346  * @return
2347  *   0 on success, a negative error value otherwise
2348  */
2349 int
2350 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2351 {
2352 	struct mlx5_priv *priv = dev->data->dev_private;
2353 
2354 	return mlx5_nl_promisc(priv->nl_socket_route,
2355 			       mlx5_ifindex(dev), !!enable);
2356 }
2357 
2358 /**
2359  * Set device promiscuous mode
2360  *
2361  * @param dev
2362  *   Pointer to Ethernet device structure.
2363  * @param enable
2364  *   0 - all multicase is disabled, otherwise - enabled
2365  *
2366  * @return
2367  *   0 on success, a negative error value otherwise
2368  */
2369 int
2370 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2371 {
2372 	struct mlx5_priv *priv = dev->data->dev_private;
2373 
2374 	return mlx5_nl_allmulti(priv->nl_socket_route,
2375 				mlx5_ifindex(dev), !!enable);
2376 }
2377 
2378 /**
2379  * Flush device MAC addresses
2380  *
2381  * @param dev
2382  *   Pointer to Ethernet device structure.
2383  *
2384  */
2385 void
2386 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2387 {
2388 	struct mlx5_priv *priv = dev->data->dev_private;
2389 
2390 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2391 			       dev->data->mac_addrs,
2392 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2393 }
2394 
2395 const struct eth_dev_ops mlx5_os_dev_ops = {
2396 	.dev_configure = mlx5_dev_configure,
2397 	.dev_start = mlx5_dev_start,
2398 	.dev_stop = mlx5_dev_stop,
2399 	.dev_set_link_down = mlx5_set_link_down,
2400 	.dev_set_link_up = mlx5_set_link_up,
2401 	.dev_close = mlx5_dev_close,
2402 	.promiscuous_enable = mlx5_promiscuous_enable,
2403 	.promiscuous_disable = mlx5_promiscuous_disable,
2404 	.allmulticast_enable = mlx5_allmulticast_enable,
2405 	.allmulticast_disable = mlx5_allmulticast_disable,
2406 	.link_update = mlx5_link_update,
2407 	.stats_get = mlx5_stats_get,
2408 	.stats_reset = mlx5_stats_reset,
2409 	.xstats_get = mlx5_xstats_get,
2410 	.xstats_reset = mlx5_xstats_reset,
2411 	.xstats_get_names = mlx5_xstats_get_names,
2412 	.fw_version_get = mlx5_fw_version_get,
2413 	.dev_infos_get = mlx5_dev_infos_get,
2414 	.read_clock = mlx5_txpp_read_clock,
2415 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2416 	.vlan_filter_set = mlx5_vlan_filter_set,
2417 	.rx_queue_setup = mlx5_rx_queue_setup,
2418 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2419 	.tx_queue_setup = mlx5_tx_queue_setup,
2420 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2421 	.rx_queue_release = mlx5_rx_queue_release,
2422 	.tx_queue_release = mlx5_tx_queue_release,
2423 	.rx_queue_start = mlx5_rx_queue_start,
2424 	.rx_queue_stop = mlx5_rx_queue_stop,
2425 	.tx_queue_start = mlx5_tx_queue_start,
2426 	.tx_queue_stop = mlx5_tx_queue_stop,
2427 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2428 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2429 	.mac_addr_remove = mlx5_mac_addr_remove,
2430 	.mac_addr_add = mlx5_mac_addr_add,
2431 	.mac_addr_set = mlx5_mac_addr_set,
2432 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2433 	.mtu_set = mlx5_dev_set_mtu,
2434 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2435 	.vlan_offload_set = mlx5_vlan_offload_set,
2436 	.reta_update = mlx5_dev_rss_reta_update,
2437 	.reta_query = mlx5_dev_rss_reta_query,
2438 	.rss_hash_update = mlx5_rss_hash_update,
2439 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2440 	.filter_ctrl = mlx5_dev_filter_ctrl,
2441 	.rxq_info_get = mlx5_rxq_info_get,
2442 	.txq_info_get = mlx5_txq_info_get,
2443 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2444 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2445 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2446 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2447 	.is_removed = mlx5_is_removed,
2448 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2449 	.get_module_info = mlx5_get_module_info,
2450 	.get_module_eeprom = mlx5_get_module_eeprom,
2451 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2452 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2453 };
2454 
2455 /* Available operations from secondary process. */
2456 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2457 	.stats_get = mlx5_stats_get,
2458 	.stats_reset = mlx5_stats_reset,
2459 	.xstats_get = mlx5_xstats_get,
2460 	.xstats_reset = mlx5_xstats_reset,
2461 	.xstats_get_names = mlx5_xstats_get_names,
2462 	.fw_version_get = mlx5_fw_version_get,
2463 	.dev_infos_get = mlx5_dev_infos_get,
2464 	.read_clock = mlx5_txpp_read_clock,
2465 	.rx_queue_start = mlx5_rx_queue_start,
2466 	.rx_queue_stop = mlx5_rx_queue_stop,
2467 	.tx_queue_start = mlx5_tx_queue_start,
2468 	.tx_queue_stop = mlx5_tx_queue_stop,
2469 	.rxq_info_get = mlx5_rxq_info_get,
2470 	.txq_info_get = mlx5_txq_info_get,
2471 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2472 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2473 	.get_module_info = mlx5_get_module_info,
2474 	.get_module_eeprom = mlx5_get_module_eeprom,
2475 };
2476 
2477 /* Available operations in flow isolated mode. */
2478 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2479 	.dev_configure = mlx5_dev_configure,
2480 	.dev_start = mlx5_dev_start,
2481 	.dev_stop = mlx5_dev_stop,
2482 	.dev_set_link_down = mlx5_set_link_down,
2483 	.dev_set_link_up = mlx5_set_link_up,
2484 	.dev_close = mlx5_dev_close,
2485 	.promiscuous_enable = mlx5_promiscuous_enable,
2486 	.promiscuous_disable = mlx5_promiscuous_disable,
2487 	.allmulticast_enable = mlx5_allmulticast_enable,
2488 	.allmulticast_disable = mlx5_allmulticast_disable,
2489 	.link_update = mlx5_link_update,
2490 	.stats_get = mlx5_stats_get,
2491 	.stats_reset = mlx5_stats_reset,
2492 	.xstats_get = mlx5_xstats_get,
2493 	.xstats_reset = mlx5_xstats_reset,
2494 	.xstats_get_names = mlx5_xstats_get_names,
2495 	.fw_version_get = mlx5_fw_version_get,
2496 	.dev_infos_get = mlx5_dev_infos_get,
2497 	.read_clock = mlx5_txpp_read_clock,
2498 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2499 	.vlan_filter_set = mlx5_vlan_filter_set,
2500 	.rx_queue_setup = mlx5_rx_queue_setup,
2501 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2502 	.tx_queue_setup = mlx5_tx_queue_setup,
2503 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2504 	.rx_queue_release = mlx5_rx_queue_release,
2505 	.tx_queue_release = mlx5_tx_queue_release,
2506 	.rx_queue_start = mlx5_rx_queue_start,
2507 	.rx_queue_stop = mlx5_rx_queue_stop,
2508 	.tx_queue_start = mlx5_tx_queue_start,
2509 	.tx_queue_stop = mlx5_tx_queue_stop,
2510 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2511 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2512 	.mac_addr_remove = mlx5_mac_addr_remove,
2513 	.mac_addr_add = mlx5_mac_addr_add,
2514 	.mac_addr_set = mlx5_mac_addr_set,
2515 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2516 	.mtu_set = mlx5_dev_set_mtu,
2517 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2518 	.vlan_offload_set = mlx5_vlan_offload_set,
2519 	.filter_ctrl = mlx5_dev_filter_ctrl,
2520 	.rxq_info_get = mlx5_rxq_info_get,
2521 	.txq_info_get = mlx5_txq_info_get,
2522 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2523 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2524 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2525 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2526 	.is_removed = mlx5_is_removed,
2527 	.get_module_info = mlx5_get_module_info,
2528 	.get_module_eeprom = mlx5_get_module_eeprom,
2529 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2530 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2531 };
2532