xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 1b9e9826ad80e5ae56b484c6f6c4bb2df6e8fa70)
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 <ethdev_driver.h>
20 #include <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_cqe = attr_ex.orig_attr.max_cqe;
120 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
121 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
122 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
123 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
124 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
125 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
126 	device_attr->max_rwq_indirection_table_size =
127 		attr_ex.rss_caps.max_rwq_indirection_table_size;
128 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
129 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
130 
131 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
132 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
133 	if (err)
134 		return err;
135 
136 	device_attr->flags = dv_attr.flags;
137 	device_attr->comp_mask = dv_attr.comp_mask;
138 #ifdef HAVE_IBV_MLX5_MOD_SWP
139 	device_attr->sw_parsing_offloads =
140 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
141 #endif
142 	device_attr->min_single_stride_log_num_of_bytes =
143 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
144 	device_attr->max_single_stride_log_num_of_bytes =
145 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
146 	device_attr->min_single_wqe_log_num_of_strides =
147 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
148 	device_attr->max_single_wqe_log_num_of_strides =
149 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
150 	device_attr->stride_supported_qpts =
151 		dv_attr.striding_rq_caps.supported_qpts;
152 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
153 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
154 #endif
155 
156 	return err;
157 }
158 
159 /**
160  * Verbs callback to allocate a memory. This function should allocate the space
161  * according to the size provided residing inside a huge page.
162  * Please note that all allocation must respect the alignment from libmlx5
163  * (i.e. currently rte_mem_page_size()).
164  *
165  * @param[in] size
166  *   The size in bytes of the memory to allocate.
167  * @param[in] data
168  *   A pointer to the callback data.
169  *
170  * @return
171  *   Allocated buffer, NULL otherwise and rte_errno is set.
172  */
173 static void *
174 mlx5_alloc_verbs_buf(size_t size, void *data)
175 {
176 	struct mlx5_dev_ctx_shared *sh = data;
177 	void *ret;
178 	size_t alignment = rte_mem_page_size();
179 	if (alignment == (size_t)-1) {
180 		DRV_LOG(ERR, "Failed to get mem page size");
181 		rte_errno = ENOMEM;
182 		return NULL;
183 	}
184 
185 	MLX5_ASSERT(data != NULL);
186 	ret = mlx5_malloc(0, size, alignment, sh->numa_node);
187 	if (!ret && size)
188 		rte_errno = ENOMEM;
189 	return ret;
190 }
191 
192 /**
193  * Verbs callback to free a memory.
194  *
195  * @param[in] ptr
196  *   A pointer to the memory to free.
197  * @param[in] data
198  *   A pointer to the callback data.
199  */
200 static void
201 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
202 {
203 	MLX5_ASSERT(data != NULL);
204 	mlx5_free(ptr);
205 }
206 
207 /**
208  * Initialize DR related data within private structure.
209  * Routine checks the reference counter and does actual
210  * resources creation/initialization only if counter is zero.
211  *
212  * @param[in] priv
213  *   Pointer to the private device data structure.
214  *
215  * @return
216  *   Zero on success, positive error code otherwise.
217  */
218 static int
219 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
220 {
221 	struct mlx5_dev_ctx_shared *sh = priv->sh;
222 	char s[MLX5_HLIST_NAMESIZE] __rte_unused;
223 	int err;
224 
225 	MLX5_ASSERT(sh && sh->refcnt);
226 	if (sh->refcnt > 1)
227 		return 0;
228 	err = mlx5_alloc_table_hash_list(priv);
229 	if (err)
230 		goto error;
231 	/* The resources below are only valid with DV support. */
232 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
233 	/* Init port id action cache list. */
234 	snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
235 	mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
236 			     flow_dv_port_id_create_cb,
237 			     flow_dv_port_id_match_cb,
238 			     flow_dv_port_id_remove_cb);
239 	/* Init push vlan action cache list. */
240 	snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
241 	mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
242 			     flow_dv_push_vlan_create_cb,
243 			     flow_dv_push_vlan_match_cb,
244 			     flow_dv_push_vlan_remove_cb);
245 	/* Init sample action cache list. */
246 	snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
247 	mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
248 			     flow_dv_sample_create_cb,
249 			     flow_dv_sample_match_cb,
250 			     flow_dv_sample_remove_cb);
251 	/* Init dest array action cache list. */
252 	snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
253 	mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
254 			     flow_dv_dest_array_create_cb,
255 			     flow_dv_dest_array_match_cb,
256 			     flow_dv_dest_array_remove_cb);
257 	/* Create tags hash list table. */
258 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
259 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
260 					  MLX5_HLIST_WRITE_MOST,
261 					  flow_dv_tag_create_cb,
262 					  flow_dv_tag_match_cb,
263 					  flow_dv_tag_remove_cb);
264 	if (!sh->tag_table) {
265 		DRV_LOG(ERR, "tags with hash creation failed.");
266 		err = ENOMEM;
267 		goto error;
268 	}
269 	sh->tag_table->ctx = sh;
270 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
271 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
272 					    0, MLX5_HLIST_WRITE_MOST |
273 					    MLX5_HLIST_DIRECT_KEY,
274 					    flow_dv_modify_create_cb,
275 					    flow_dv_modify_match_cb,
276 					    flow_dv_modify_remove_cb);
277 	if (!sh->modify_cmds) {
278 		DRV_LOG(ERR, "hdr modify hash creation failed");
279 		err = ENOMEM;
280 		goto error;
281 	}
282 	sh->modify_cmds->ctx = sh;
283 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
284 	sh->encaps_decaps = mlx5_hlist_create(s,
285 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
286 					      0, MLX5_HLIST_DIRECT_KEY |
287 					      MLX5_HLIST_WRITE_MOST,
288 					      flow_dv_encap_decap_create_cb,
289 					      flow_dv_encap_decap_match_cb,
290 					      flow_dv_encap_decap_remove_cb);
291 	if (!sh->encaps_decaps) {
292 		DRV_LOG(ERR, "encap decap hash creation failed");
293 		err = ENOMEM;
294 		goto error;
295 	}
296 	sh->encaps_decaps->ctx = sh;
297 #endif
298 #ifdef HAVE_MLX5DV_DR
299 	void *domain;
300 
301 	/* Reference counter is zero, we should initialize structures. */
302 	domain = mlx5_glue->dr_create_domain(sh->ctx,
303 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
304 	if (!domain) {
305 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
306 		err = errno;
307 		goto error;
308 	}
309 	sh->rx_domain = domain;
310 	domain = mlx5_glue->dr_create_domain(sh->ctx,
311 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
312 	if (!domain) {
313 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
314 		err = errno;
315 		goto error;
316 	}
317 	sh->tx_domain = domain;
318 #ifdef HAVE_MLX5DV_DR_ESWITCH
319 	if (priv->config.dv_esw_en) {
320 		domain  = mlx5_glue->dr_create_domain
321 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
322 		if (!domain) {
323 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
324 			err = errno;
325 			goto error;
326 		}
327 		sh->fdb_domain = domain;
328 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
329 	}
330 #endif
331 	if (!sh->tunnel_hub)
332 		err = mlx5_alloc_tunnel_hub(sh);
333 	if (err) {
334 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
335 		goto error;
336 	}
337 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
338 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
339 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
340 		if (sh->fdb_domain)
341 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
342 	}
343 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
344 #endif /* HAVE_MLX5DV_DR */
345 	sh->default_miss_action =
346 			mlx5_glue->dr_create_flow_action_default_miss();
347 	if (!sh->default_miss_action)
348 		DRV_LOG(WARNING, "Default miss action is not supported.");
349 	return 0;
350 error:
351 	/* Rollback the created objects. */
352 	if (sh->rx_domain) {
353 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
354 		sh->rx_domain = NULL;
355 	}
356 	if (sh->tx_domain) {
357 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
358 		sh->tx_domain = NULL;
359 	}
360 	if (sh->fdb_domain) {
361 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
362 		sh->fdb_domain = NULL;
363 	}
364 	if (sh->esw_drop_action) {
365 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
366 		sh->esw_drop_action = NULL;
367 	}
368 	if (sh->pop_vlan_action) {
369 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
370 		sh->pop_vlan_action = NULL;
371 	}
372 	if (sh->encaps_decaps) {
373 		mlx5_hlist_destroy(sh->encaps_decaps);
374 		sh->encaps_decaps = NULL;
375 	}
376 	if (sh->modify_cmds) {
377 		mlx5_hlist_destroy(sh->modify_cmds);
378 		sh->modify_cmds = NULL;
379 	}
380 	if (sh->tag_table) {
381 		/* tags should be destroyed with flow before. */
382 		mlx5_hlist_destroy(sh->tag_table);
383 		sh->tag_table = NULL;
384 	}
385 	if (sh->tunnel_hub) {
386 		mlx5_release_tunnel_hub(sh, priv->dev_port);
387 		sh->tunnel_hub = NULL;
388 	}
389 	mlx5_free_table_hash_list(priv);
390 	return err;
391 }
392 
393 /**
394  * Destroy DR related data within private structure.
395  *
396  * @param[in] priv
397  *   Pointer to the private device data structure.
398  */
399 void
400 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
401 {
402 	struct mlx5_dev_ctx_shared *sh = priv->sh;
403 
404 	MLX5_ASSERT(sh && sh->refcnt);
405 	if (sh->refcnt > 1)
406 		return;
407 #ifdef HAVE_MLX5DV_DR
408 	if (sh->rx_domain) {
409 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
410 		sh->rx_domain = NULL;
411 	}
412 	if (sh->tx_domain) {
413 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
414 		sh->tx_domain = NULL;
415 	}
416 #ifdef HAVE_MLX5DV_DR_ESWITCH
417 	if (sh->fdb_domain) {
418 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
419 		sh->fdb_domain = NULL;
420 	}
421 	if (sh->esw_drop_action) {
422 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
423 		sh->esw_drop_action = NULL;
424 	}
425 #endif
426 	if (sh->pop_vlan_action) {
427 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
428 		sh->pop_vlan_action = NULL;
429 	}
430 #endif /* HAVE_MLX5DV_DR */
431 	if (sh->default_miss_action)
432 		mlx5_glue->destroy_flow_action
433 				(sh->default_miss_action);
434 	if (sh->encaps_decaps) {
435 		mlx5_hlist_destroy(sh->encaps_decaps);
436 		sh->encaps_decaps = NULL;
437 	}
438 	if (sh->modify_cmds) {
439 		mlx5_hlist_destroy(sh->modify_cmds);
440 		sh->modify_cmds = NULL;
441 	}
442 	if (sh->tag_table) {
443 		/* tags should be destroyed with flow before. */
444 		mlx5_hlist_destroy(sh->tag_table);
445 		sh->tag_table = NULL;
446 	}
447 	if (sh->tunnel_hub) {
448 		mlx5_release_tunnel_hub(sh, priv->dev_port);
449 		sh->tunnel_hub = NULL;
450 	}
451 	mlx5_cache_list_destroy(&sh->port_id_action_list);
452 	mlx5_cache_list_destroy(&sh->push_vlan_action_list);
453 	mlx5_free_table_hash_list(priv);
454 }
455 
456 /**
457  * Initialize shared data between primary and secondary process.
458  *
459  * A memzone is reserved by primary process and secondary processes attach to
460  * the memzone.
461  *
462  * @return
463  *   0 on success, a negative errno value otherwise and rte_errno is set.
464  */
465 static int
466 mlx5_init_shared_data(void)
467 {
468 	const struct rte_memzone *mz;
469 	int ret = 0;
470 
471 	rte_spinlock_lock(&mlx5_shared_data_lock);
472 	if (mlx5_shared_data == NULL) {
473 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
474 			/* Allocate shared memory. */
475 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
476 						 sizeof(*mlx5_shared_data),
477 						 SOCKET_ID_ANY, 0);
478 			if (mz == NULL) {
479 				DRV_LOG(ERR,
480 					"Cannot allocate mlx5 shared data");
481 				ret = -rte_errno;
482 				goto error;
483 			}
484 			mlx5_shared_data = mz->addr;
485 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
486 			rte_spinlock_init(&mlx5_shared_data->lock);
487 		} else {
488 			/* Lookup allocated shared memory. */
489 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
490 			if (mz == NULL) {
491 				DRV_LOG(ERR,
492 					"Cannot attach mlx5 shared data");
493 				ret = -rte_errno;
494 				goto error;
495 			}
496 			mlx5_shared_data = mz->addr;
497 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
498 		}
499 	}
500 error:
501 	rte_spinlock_unlock(&mlx5_shared_data_lock);
502 	return ret;
503 }
504 
505 /**
506  * PMD global initialization.
507  *
508  * Independent from individual device, this function initializes global
509  * per-PMD data structures distinguishing primary and secondary processes.
510  * Hence, each initialization is called once per a process.
511  *
512  * @return
513  *   0 on success, a negative errno value otherwise and rte_errno is set.
514  */
515 static int
516 mlx5_init_once(void)
517 {
518 	struct mlx5_shared_data *sd;
519 	struct mlx5_local_data *ld = &mlx5_local_data;
520 	int ret = 0;
521 
522 	if (mlx5_init_shared_data())
523 		return -rte_errno;
524 	sd = mlx5_shared_data;
525 	MLX5_ASSERT(sd);
526 	rte_spinlock_lock(&sd->lock);
527 	switch (rte_eal_process_type()) {
528 	case RTE_PROC_PRIMARY:
529 		if (sd->init_done)
530 			break;
531 		LIST_INIT(&sd->mem_event_cb_list);
532 		rte_rwlock_init(&sd->mem_event_rwlock);
533 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
534 						mlx5_mr_mem_event_cb, NULL);
535 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
536 					   mlx5_mp_os_primary_handle);
537 		if (ret)
538 			goto out;
539 		sd->init_done = true;
540 		break;
541 	case RTE_PROC_SECONDARY:
542 		if (ld->init_done)
543 			break;
544 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
545 					     mlx5_mp_os_secondary_handle);
546 		if (ret)
547 			goto out;
548 		++sd->secondary_cnt;
549 		ld->init_done = true;
550 		break;
551 	default:
552 		break;
553 	}
554 out:
555 	rte_spinlock_unlock(&sd->lock);
556 	return ret;
557 }
558 
559 /**
560  * Create the Tx queue DevX/Verbs object.
561  *
562  * @param dev
563  *   Pointer to Ethernet device.
564  * @param idx
565  *   Queue index in DPDK Tx queue array.
566  *
567  * @return
568  *   0 on success, a negative errno value otherwise and rte_errno is set.
569  */
570 static int
571 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
572 {
573 	struct mlx5_priv *priv = dev->data->dev_private;
574 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
575 	struct mlx5_txq_ctrl *txq_ctrl =
576 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
577 
578 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
579 		return mlx5_txq_devx_obj_new(dev, idx);
580 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
581 	if (!priv->config.dv_esw_en)
582 		return mlx5_txq_devx_obj_new(dev, idx);
583 #endif
584 	return mlx5_txq_ibv_obj_new(dev, idx);
585 }
586 
587 /**
588  * Release an Tx DevX/verbs queue object.
589  *
590  * @param txq_obj
591  *   DevX/Verbs Tx queue object.
592  */
593 static void
594 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
595 {
596 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
597 		mlx5_txq_devx_obj_release(txq_obj);
598 		return;
599 	}
600 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
601 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
602 		mlx5_txq_devx_obj_release(txq_obj);
603 		return;
604 	}
605 #endif
606 	mlx5_txq_ibv_obj_release(txq_obj);
607 }
608 
609 /**
610  * DV flow counter mode detect and config.
611  *
612  * @param dev
613  *   Pointer to rte_eth_dev structure.
614  *
615  */
616 static void
617 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
618 {
619 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
620 	struct mlx5_priv *priv = dev->data->dev_private;
621 	struct mlx5_dev_ctx_shared *sh = priv->sh;
622 	bool fallback;
623 
624 #ifndef HAVE_IBV_DEVX_ASYNC
625 	fallback = true;
626 #else
627 	fallback = false;
628 	if (!priv->config.devx || !priv->config.dv_flow_en ||
629 	    !priv->config.hca_attr.flow_counters_dump ||
630 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
631 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
632 		fallback = true;
633 #endif
634 	if (fallback)
635 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
636 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
637 			priv->config.hca_attr.flow_counters_dump,
638 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
639 	/* Initialize fallback mode only on the port initializes sh. */
640 	if (sh->refcnt == 1)
641 		sh->cmng.counter_fallback = fallback;
642 	else if (fallback != sh->cmng.counter_fallback)
643 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
644 			"with others:%d.", PORT_ID(priv), fallback);
645 #endif
646 }
647 
648 static void
649 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
650 {
651 	struct mlx5_priv *priv = dev->data->dev_private;
652 	void *ctx = priv->sh->ctx;
653 
654 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
655 	if (!priv->q_counters) {
656 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
657 		struct ibv_wq *wq;
658 
659 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
660 			"by DevX - fall-back to use the kernel driver global "
661 			"queue counter.", dev->data->port_id);
662 		/* Create WQ by kernel and query its queue counter ID. */
663 		if (cq) {
664 			wq = mlx5_glue->create_wq(ctx,
665 						  &(struct ibv_wq_init_attr){
666 						    .wq_type = IBV_WQT_RQ,
667 						    .max_wr = 1,
668 						    .max_sge = 1,
669 						    .pd = priv->sh->pd,
670 						    .cq = cq,
671 						});
672 			if (wq) {
673 				/* Counter is assigned only on RDY state. */
674 				int ret = mlx5_glue->modify_wq(wq,
675 						 &(struct ibv_wq_attr){
676 						 .attr_mask = IBV_WQ_ATTR_STATE,
677 						 .wq_state = IBV_WQS_RDY,
678 						});
679 
680 				if (ret == 0)
681 					mlx5_devx_cmd_wq_query(wq,
682 							 &priv->counter_set_id);
683 				claim_zero(mlx5_glue->destroy_wq(wq));
684 			}
685 			claim_zero(mlx5_glue->destroy_cq(cq));
686 		}
687 	} else {
688 		priv->counter_set_id = priv->q_counters->id;
689 	}
690 	if (priv->counter_set_id == 0)
691 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
692 			"available.", dev->data->port_id);
693 }
694 
695 /**
696  * Spawn an Ethernet device from Verbs information.
697  *
698  * @param dpdk_dev
699  *   Backing DPDK device.
700  * @param spawn
701  *   Verbs device parameters (name, port, switch_info) to spawn.
702  * @param config
703  *   Device configuration parameters.
704  *
705  * @return
706  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
707  *   is set. The following errors are defined:
708  *
709  *   EBUSY: device is not supposed to be spawned.
710  *   EEXIST: device is already spawned
711  */
712 static struct rte_eth_dev *
713 mlx5_dev_spawn(struct rte_device *dpdk_dev,
714 	       struct mlx5_dev_spawn_data *spawn,
715 	       struct mlx5_dev_config *config)
716 {
717 	const struct mlx5_switch_info *switch_info = &spawn->info;
718 	struct mlx5_dev_ctx_shared *sh = NULL;
719 	struct ibv_port_attr port_attr;
720 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
721 	struct rte_eth_dev *eth_dev = NULL;
722 	struct mlx5_priv *priv = NULL;
723 	int err = 0;
724 	unsigned int hw_padding = 0;
725 	unsigned int mps;
726 	unsigned int tunnel_en = 0;
727 	unsigned int mpls_en = 0;
728 	unsigned int swp = 0;
729 	unsigned int mprq = 0;
730 	unsigned int mprq_min_stride_size_n = 0;
731 	unsigned int mprq_max_stride_size_n = 0;
732 	unsigned int mprq_min_stride_num_n = 0;
733 	unsigned int mprq_max_stride_num_n = 0;
734 	struct rte_ether_addr mac;
735 	char name[RTE_ETH_NAME_MAX_LEN];
736 	int own_domain_id = 0;
737 	uint16_t port_id;
738 	unsigned int i;
739 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
740 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
741 #endif
742 
743 	/* Determine if this port representor is supposed to be spawned. */
744 	if (switch_info->representor && dpdk_dev->devargs) {
745 		struct rte_eth_devargs eth_da;
746 
747 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
748 		if (err) {
749 			rte_errno = -err;
750 			DRV_LOG(ERR, "failed to process device arguments: %s",
751 				strerror(rte_errno));
752 			return NULL;
753 		}
754 		if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
755 			/* Representor not specified. */
756 			rte_errno = EBUSY;
757 			return NULL;
758 		}
759 		if (eth_da.type != RTE_ETH_REPRESENTOR_VF) {
760 			rte_errno = ENOTSUP;
761 			DRV_LOG(ERR, "unsupported representor type: %s",
762 				dpdk_dev->devargs->args);
763 			return NULL;
764 		}
765 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
766 			if (eth_da.representor_ports[i] ==
767 			    (uint16_t)switch_info->port_name)
768 				break;
769 		if (i == eth_da.nb_representor_ports) {
770 			rte_errno = EBUSY;
771 			return NULL;
772 		}
773 	}
774 	/* Build device name. */
775 	if (spawn->pf_bond <  0) {
776 		/* Single device. */
777 		if (!switch_info->representor)
778 			strlcpy(name, dpdk_dev->name, sizeof(name));
779 		else
780 			snprintf(name, sizeof(name), "%s_representor_%u",
781 				 dpdk_dev->name, switch_info->port_name);
782 	} else {
783 		/* Bonding device. */
784 		if (!switch_info->representor)
785 			snprintf(name, sizeof(name), "%s_%s",
786 				 dpdk_dev->name,
787 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
788 		else
789 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
790 				 dpdk_dev->name,
791 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
792 				 switch_info->port_name);
793 	}
794 	/* check if the device is already spawned */
795 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
796 		rte_errno = EEXIST;
797 		return NULL;
798 	}
799 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
800 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
801 		struct mlx5_mp_id mp_id;
802 
803 		eth_dev = rte_eth_dev_attach_secondary(name);
804 		if (eth_dev == NULL) {
805 			DRV_LOG(ERR, "can not attach rte ethdev");
806 			rte_errno = ENOMEM;
807 			return NULL;
808 		}
809 		priv = eth_dev->data->dev_private;
810 		if (priv->sh->bond_dev != UINT16_MAX)
811 			/* For bonding port, use primary PCI device. */
812 			eth_dev->device =
813 				rte_eth_devices[priv->sh->bond_dev].device;
814 		else
815 			eth_dev->device = dpdk_dev;
816 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
817 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
818 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
819 		err = mlx5_proc_priv_init(eth_dev);
820 		if (err)
821 			return NULL;
822 		mp_id.port_id = eth_dev->data->port_id;
823 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
824 		/* Receive command fd from primary process */
825 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
826 		if (err < 0)
827 			goto err_secondary;
828 		/* Remap UAR for Tx queues. */
829 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
830 		if (err)
831 			goto err_secondary;
832 		/*
833 		 * Ethdev pointer is still required as input since
834 		 * the primary device is not accessible from the
835 		 * secondary process.
836 		 */
837 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
838 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
839 		return eth_dev;
840 err_secondary:
841 		mlx5_dev_close(eth_dev);
842 		return NULL;
843 	}
844 	/*
845 	 * Some parameters ("tx_db_nc" in particularly) are needed in
846 	 * advance to create dv/verbs device context. We proceed the
847 	 * devargs here to get ones, and later proceed devargs again
848 	 * to override some hardware settings.
849 	 */
850 	err = mlx5_args(config, dpdk_dev->devargs);
851 	if (err) {
852 		err = rte_errno;
853 		DRV_LOG(ERR, "failed to process device arguments: %s",
854 			strerror(rte_errno));
855 		goto error;
856 	}
857 	if (config->dv_miss_info) {
858 		if (switch_info->master || switch_info->representor)
859 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
860 	}
861 	mlx5_malloc_mem_select(config->sys_mem_en);
862 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
863 	if (!sh)
864 		return NULL;
865 	config->devx = sh->devx;
866 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
867 	config->dest_tir = 1;
868 #endif
869 #ifdef HAVE_IBV_MLX5_MOD_SWP
870 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
871 #endif
872 	/*
873 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
874 	 * as all ConnectX-5 devices.
875 	 */
876 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
877 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
878 #endif
879 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
880 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
881 #endif
882 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
883 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
884 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
885 			DRV_LOG(DEBUG, "enhanced MPW is supported");
886 			mps = MLX5_MPW_ENHANCED;
887 		} else {
888 			DRV_LOG(DEBUG, "MPW is supported");
889 			mps = MLX5_MPW;
890 		}
891 	} else {
892 		DRV_LOG(DEBUG, "MPW isn't supported");
893 		mps = MLX5_MPW_DISABLED;
894 	}
895 #ifdef HAVE_IBV_MLX5_MOD_SWP
896 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
897 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
898 	DRV_LOG(DEBUG, "SWP support: %u", swp);
899 #endif
900 	config->swp = !!swp;
901 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
902 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
903 		struct mlx5dv_striding_rq_caps mprq_caps =
904 			dv_attr.striding_rq_caps;
905 
906 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
907 			mprq_caps.min_single_stride_log_num_of_bytes);
908 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
909 			mprq_caps.max_single_stride_log_num_of_bytes);
910 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
911 			mprq_caps.min_single_wqe_log_num_of_strides);
912 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
913 			mprq_caps.max_single_wqe_log_num_of_strides);
914 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
915 			mprq_caps.supported_qpts);
916 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
917 		mprq = 1;
918 		mprq_min_stride_size_n =
919 			mprq_caps.min_single_stride_log_num_of_bytes;
920 		mprq_max_stride_size_n =
921 			mprq_caps.max_single_stride_log_num_of_bytes;
922 		mprq_min_stride_num_n =
923 			mprq_caps.min_single_wqe_log_num_of_strides;
924 		mprq_max_stride_num_n =
925 			mprq_caps.max_single_wqe_log_num_of_strides;
926 	}
927 #endif
928 	/* Rx CQE compression is enabled by default. */
929 	config->cqe_comp = 1;
930 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
931 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
932 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
933 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
934 			     (dv_attr.tunnel_offloads_caps &
935 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
936 			     (dv_attr.tunnel_offloads_caps &
937 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
938 	}
939 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
940 		tunnel_en ? "" : "not ");
941 #else
942 	DRV_LOG(WARNING,
943 		"tunnel offloading disabled due to old OFED/rdma-core version");
944 #endif
945 	config->tunnel_en = tunnel_en;
946 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
947 	mpls_en = ((dv_attr.tunnel_offloads_caps &
948 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
949 		   (dv_attr.tunnel_offloads_caps &
950 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
951 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
952 		mpls_en ? "" : "not ");
953 #else
954 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
955 		" old OFED/rdma-core version or firmware configuration");
956 #endif
957 	config->mpls_en = mpls_en;
958 	/* Check port status. */
959 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
960 	if (err) {
961 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
962 		goto error;
963 	}
964 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
965 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
966 		err = EINVAL;
967 		goto error;
968 	}
969 	if (port_attr.state != IBV_PORT_ACTIVE)
970 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
971 			mlx5_glue->port_state_str(port_attr.state),
972 			port_attr.state);
973 	/* Allocate private eth device data. */
974 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
975 			   sizeof(*priv),
976 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
977 	if (priv == NULL) {
978 		DRV_LOG(ERR, "priv allocation failure");
979 		err = ENOMEM;
980 		goto error;
981 	}
982 	priv->sh = sh;
983 	priv->dev_port = spawn->phys_port;
984 	priv->pci_dev = spawn->pci_dev;
985 	priv->mtu = RTE_ETHER_MTU;
986 	/* Some internal functions rely on Netlink sockets, open them now. */
987 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
988 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
989 	priv->representor = !!switch_info->representor;
990 	priv->master = !!switch_info->master;
991 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
992 	priv->vport_meta_tag = 0;
993 	priv->vport_meta_mask = 0;
994 	priv->pf_bond = spawn->pf_bond;
995 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
996 	/*
997 	 * The DevX port query API is implemented. E-Switch may use
998 	 * either vport or reg_c[0] metadata register to match on
999 	 * vport index. The engaged part of metadata register is
1000 	 * defined by mask.
1001 	 */
1002 	if (switch_info->representor || switch_info->master) {
1003 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1004 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1005 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1006 						 &devx_port);
1007 		if (err) {
1008 			DRV_LOG(WARNING,
1009 				"can't query devx port %d on device %s",
1010 				spawn->phys_port,
1011 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1012 			devx_port.comp_mask = 0;
1013 		}
1014 	}
1015 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1016 		priv->vport_meta_tag = devx_port.reg_c_0.value;
1017 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
1018 		if (!priv->vport_meta_mask) {
1019 			DRV_LOG(ERR, "vport zero mask for port %d"
1020 				     " on bonding device %s",
1021 				     spawn->phys_port,
1022 				     mlx5_os_get_dev_device_name
1023 							(spawn->phys_dev));
1024 			err = ENOTSUP;
1025 			goto error;
1026 		}
1027 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1028 			DRV_LOG(ERR, "invalid vport tag for port %d"
1029 				     " on bonding device %s",
1030 				     spawn->phys_port,
1031 				     mlx5_os_get_dev_device_name
1032 							(spawn->phys_dev));
1033 			err = ENOTSUP;
1034 			goto error;
1035 		}
1036 	}
1037 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1038 		priv->vport_id = devx_port.vport_num;
1039 	} else if (spawn->pf_bond >= 0) {
1040 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1041 			     " on bonding device %s",
1042 			     spawn->phys_port,
1043 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1044 		err = ENOTSUP;
1045 		goto error;
1046 	} else {
1047 		/* Suppose vport index in compatible way. */
1048 		priv->vport_id = switch_info->representor ?
1049 				 switch_info->port_name + 1 : -1;
1050 	}
1051 #else
1052 	/*
1053 	 * Kernel/rdma_core support single E-Switch per PF configurations
1054 	 * only and vport_id field contains the vport index for
1055 	 * associated VF, which is deduced from representor port name.
1056 	 * For example, let's have the IB device port 10, it has
1057 	 * attached network device eth0, which has port name attribute
1058 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1059 	 * as 3 (2+1). This assigning schema should be changed if the
1060 	 * multiple E-Switch instances per PF configurations or/and PCI
1061 	 * subfunctions are added.
1062 	 */
1063 	priv->vport_id = switch_info->representor ?
1064 			 switch_info->port_name + 1 : -1;
1065 #endif
1066 	/* representor_id field keeps the unmodified VF index. */
1067 	priv->representor_id = switch_info->representor ?
1068 			       switch_info->port_name : -1;
1069 	/*
1070 	 * Look for sibling devices in order to reuse their switch domain
1071 	 * if any, otherwise allocate one.
1072 	 */
1073 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1074 		const struct mlx5_priv *opriv =
1075 			rte_eth_devices[port_id].data->dev_private;
1076 
1077 		if (!opriv ||
1078 		    opriv->sh != priv->sh ||
1079 			opriv->domain_id ==
1080 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1081 			continue;
1082 		priv->domain_id = opriv->domain_id;
1083 		break;
1084 	}
1085 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1086 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1087 		if (err) {
1088 			err = rte_errno;
1089 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1090 				strerror(rte_errno));
1091 			goto error;
1092 		}
1093 		own_domain_id = 1;
1094 	}
1095 	/* Override some values set by hardware configuration. */
1096 	mlx5_args(config, dpdk_dev->devargs);
1097 	err = mlx5_dev_check_sibling_config(priv, config);
1098 	if (err)
1099 		goto error;
1100 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1101 			    IBV_DEVICE_RAW_IP_CSUM);
1102 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1103 		(config->hw_csum ? "" : "not "));
1104 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1105 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1106 	DRV_LOG(DEBUG, "counters are not supported");
1107 #endif
1108 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1109 	if (config->dv_flow_en) {
1110 		DRV_LOG(WARNING, "DV flow is not supported");
1111 		config->dv_flow_en = 0;
1112 	}
1113 #endif
1114 	config->ind_table_max_size =
1115 		sh->device_attr.max_rwq_indirection_table_size;
1116 	/*
1117 	 * Remove this check once DPDK supports larger/variable
1118 	 * indirection tables.
1119 	 */
1120 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1121 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1122 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1123 		config->ind_table_max_size);
1124 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1125 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1126 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1127 		(config->hw_vlan_strip ? "" : "not "));
1128 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1129 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1130 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1131 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1132 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1133 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1134 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1135 #endif
1136 	if (config->hw_padding && !hw_padding) {
1137 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1138 		config->hw_padding = 0;
1139 	} else if (config->hw_padding) {
1140 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1141 	}
1142 	config->tso = (sh->device_attr.max_tso > 0 &&
1143 		      (sh->device_attr.tso_supported_qpts &
1144 		       (1 << IBV_QPT_RAW_PACKET)));
1145 	if (config->tso)
1146 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1147 	/*
1148 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1149 	 * by default.
1150 	 */
1151 	if (config->mps == MLX5_ARG_UNSET)
1152 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1153 							  MLX5_MPW_DISABLED;
1154 	else
1155 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1156 	DRV_LOG(INFO, "%sMPS is %s",
1157 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1158 		config->mps == MLX5_MPW ? "legacy " : "",
1159 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1160 	if (config->devx) {
1161 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1162 		if (err) {
1163 			err = -err;
1164 			goto error;
1165 		}
1166 		/* Check relax ordering support. */
1167 		if (!haswell_broadwell_cpu) {
1168 			sh->cmng.relaxed_ordering_write =
1169 				config->hca_attr.relaxed_ordering_write;
1170 			sh->cmng.relaxed_ordering_read =
1171 				config->hca_attr.relaxed_ordering_read;
1172 		} else {
1173 			sh->cmng.relaxed_ordering_read = 0;
1174 			sh->cmng.relaxed_ordering_write = 0;
1175 		}
1176 		/* Check for LRO support. */
1177 		if (config->dest_tir && config->hca_attr.lro_cap &&
1178 		    config->dv_flow_en) {
1179 			/* TBD check tunnel lro caps. */
1180 			config->lro.supported = config->hca_attr.lro_cap;
1181 			DRV_LOG(DEBUG, "Device supports LRO");
1182 			/*
1183 			 * If LRO timeout is not configured by application,
1184 			 * use the minimal supported value.
1185 			 */
1186 			if (!config->lro.timeout)
1187 				config->lro.timeout =
1188 				config->hca_attr.lro_timer_supported_periods[0];
1189 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1190 				config->lro.timeout);
1191 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1192 				"required for coalescing is %d bytes",
1193 				config->hca_attr.lro_min_mss_size);
1194 		}
1195 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1196 		if (config->hca_attr.qos.sup &&
1197 		    config->hca_attr.qos.flow_meter_old &&
1198 		    config->dv_flow_en) {
1199 			uint8_t reg_c_mask =
1200 				config->hca_attr.qos.flow_meter_reg_c_ids;
1201 			/*
1202 			 * Meter needs two REG_C's for color match and pre-sfx
1203 			 * flow match. Here get the REG_C for color match.
1204 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1205 			 */
1206 			reg_c_mask &= 0xfc;
1207 			if (__builtin_popcount(reg_c_mask) < 1) {
1208 				priv->mtr_en = 0;
1209 				DRV_LOG(WARNING, "No available register for"
1210 					" meter.");
1211 			} else {
1212 				/*
1213 				 * The meter color register is used by the
1214 				 * flow-hit feature as well.
1215 				 * The flow-hit feature must use REG_C_3
1216 				 * Prefer REG_C_3 if it is available.
1217 				 */
1218 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1219 					priv->mtr_color_reg = REG_C_3;
1220 				else
1221 					priv->mtr_color_reg = ffs(reg_c_mask)
1222 							      - 1 + REG_C_0;
1223 				priv->mtr_en = 1;
1224 				priv->mtr_reg_share =
1225 				      config->hca_attr.qos.flow_meter;
1226 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1227 					priv->mtr_color_reg);
1228 			}
1229 		}
1230 #endif
1231 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1232 		if (config->hca_attr.flow_hit_aso &&
1233 		    priv->mtr_color_reg == REG_C_3) {
1234 			sh->flow_hit_aso_en = 1;
1235 			err = mlx5_flow_aso_age_mng_init(sh);
1236 			if (err) {
1237 				err = -err;
1238 				goto error;
1239 			}
1240 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1241 		}
1242 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1243 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1244 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1245 		    config->dv_flow_en) {
1246 			priv->sampler_en = 1;
1247 			DRV_LOG(DEBUG, "Sampler enabled!");
1248 		} else {
1249 			priv->sampler_en = 0;
1250 			if (!config->hca_attr.log_max_ft_sampler_num)
1251 				DRV_LOG(WARNING,
1252 					"No available register for sampler.");
1253 			else
1254 				DRV_LOG(DEBUG, "DV flow is not supported!");
1255 		}
1256 #endif
1257 	}
1258 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1259 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1260 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1261 		config->cqe_comp = 0;
1262 	}
1263 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1264 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1265 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1266 				 " format isn't supported.");
1267 		config->cqe_comp = 0;
1268 	}
1269 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1270 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1271 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1272 				 " format isn't supported.");
1273 		config->cqe_comp = 0;
1274 	}
1275 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1276 			config->cqe_comp ? "" : "not ");
1277 	if (config->tx_pp) {
1278 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1279 			config->hca_attr.dev_freq_khz);
1280 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1281 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1282 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1283 			config->hca_attr.cross_channel ? "" : "not ");
1284 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1285 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1286 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1287 			config->hca_attr.non_wire_sq ? "" : "not ");
1288 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1289 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1290 			config->hca_attr.log_max_static_sq_wq);
1291 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1292 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1293 		if (!config->devx) {
1294 			DRV_LOG(ERR, "DevX is required for packet pacing");
1295 			err = ENODEV;
1296 			goto error;
1297 		}
1298 		if (!config->hca_attr.qos.packet_pacing) {
1299 			DRV_LOG(ERR, "Packet pacing is not supported");
1300 			err = ENODEV;
1301 			goto error;
1302 		}
1303 		if (!config->hca_attr.cross_channel) {
1304 			DRV_LOG(ERR, "Cross channel operations are"
1305 				     " required for packet pacing");
1306 			err = ENODEV;
1307 			goto error;
1308 		}
1309 		if (!config->hca_attr.wqe_index_ignore) {
1310 			DRV_LOG(ERR, "WQE index ignore feature is"
1311 				     " required for packet pacing");
1312 			err = ENODEV;
1313 			goto error;
1314 		}
1315 		if (!config->hca_attr.non_wire_sq) {
1316 			DRV_LOG(ERR, "Non-wire SQ feature is"
1317 				     " required for packet pacing");
1318 			err = ENODEV;
1319 			goto error;
1320 		}
1321 		if (!config->hca_attr.log_max_static_sq_wq) {
1322 			DRV_LOG(ERR, "Static WQE SQ feature is"
1323 				     " required for packet pacing");
1324 			err = ENODEV;
1325 			goto error;
1326 		}
1327 		if (!config->hca_attr.qos.wqe_rate_pp) {
1328 			DRV_LOG(ERR, "WQE rate mode is required"
1329 				     " for packet pacing");
1330 			err = ENODEV;
1331 			goto error;
1332 		}
1333 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1334 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1335 			     " can't create queues for packet pacing");
1336 		err = ENODEV;
1337 		goto error;
1338 #endif
1339 	}
1340 	if (config->devx) {
1341 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1342 
1343 		err = config->hca_attr.access_register_user ?
1344 			mlx5_devx_cmd_register_read
1345 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1346 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1347 		if (!err) {
1348 			uint32_t ts_mode;
1349 
1350 			/* MTUTC register is read successfully. */
1351 			ts_mode = MLX5_GET(register_mtutc, reg,
1352 					   time_stamp_mode);
1353 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1354 				config->rt_timestamp = 1;
1355 		} else {
1356 			/* Kernel does not support register reading. */
1357 			if (config->hca_attr.dev_freq_khz ==
1358 						 (NS_PER_S / MS_PER_S))
1359 				config->rt_timestamp = 1;
1360 		}
1361 	}
1362 	/*
1363 	 * If HW has bug working with tunnel packet decapsulation and
1364 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1365 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1366 	 */
1367 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1368 		config->hw_fcs_strip = 0;
1369 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1370 		(config->hw_fcs_strip ? "" : "not "));
1371 	if (config->mprq.enabled && mprq) {
1372 		if (config->mprq.stride_num_n &&
1373 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1374 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1375 			config->mprq.stride_num_n =
1376 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1377 						mprq_min_stride_num_n),
1378 					mprq_max_stride_num_n);
1379 			DRV_LOG(WARNING,
1380 				"the number of strides"
1381 				" for Multi-Packet RQ is out of range,"
1382 				" setting default value (%u)",
1383 				1 << config->mprq.stride_num_n);
1384 		}
1385 		if (config->mprq.stride_size_n &&
1386 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1387 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1388 			config->mprq.stride_size_n =
1389 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1390 						mprq_min_stride_size_n),
1391 					mprq_max_stride_size_n);
1392 			DRV_LOG(WARNING,
1393 				"the size of a stride"
1394 				" for Multi-Packet RQ is out of range,"
1395 				" setting default value (%u)",
1396 				1 << config->mprq.stride_size_n);
1397 		}
1398 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1399 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1400 	} else if (config->mprq.enabled && !mprq) {
1401 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1402 		config->mprq.enabled = 0;
1403 	}
1404 	if (config->max_dump_files_num == 0)
1405 		config->max_dump_files_num = 128;
1406 	eth_dev = rte_eth_dev_allocate(name);
1407 	if (eth_dev == NULL) {
1408 		DRV_LOG(ERR, "can not allocate rte ethdev");
1409 		err = ENOMEM;
1410 		goto error;
1411 	}
1412 	if (priv->representor) {
1413 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1414 		eth_dev->data->representor_id = priv->representor_id;
1415 	}
1416 	priv->mp_id.port_id = eth_dev->data->port_id;
1417 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1418 	/*
1419 	 * Store associated network device interface index. This index
1420 	 * is permanent throughout the lifetime of device. So, we may store
1421 	 * the ifindex here and use the cached value further.
1422 	 */
1423 	MLX5_ASSERT(spawn->ifindex);
1424 	priv->if_index = spawn->ifindex;
1425 	if (priv->pf_bond >= 0 && priv->master) {
1426 		/* Get bond interface info */
1427 		err = mlx5_sysfs_bond_info(priv->if_index,
1428 				     &priv->bond_ifindex,
1429 				     priv->bond_name);
1430 		if (err)
1431 			DRV_LOG(ERR, "unable to get bond info: %s",
1432 				strerror(rte_errno));
1433 		else
1434 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1435 				priv->if_index, priv->bond_ifindex,
1436 				priv->bond_name);
1437 	}
1438 	eth_dev->data->dev_private = priv;
1439 	priv->dev_data = eth_dev->data;
1440 	eth_dev->data->mac_addrs = priv->mac;
1441 	if (spawn->pf_bond < 0) {
1442 		eth_dev->device = dpdk_dev;
1443 	} else {
1444 		/* Use primary bond PCI as device. */
1445 		if (sh->bond_dev == UINT16_MAX) {
1446 			sh->bond_dev = eth_dev->data->port_id;
1447 			eth_dev->device = dpdk_dev;
1448 		} else {
1449 			eth_dev->device = rte_eth_devices[sh->bond_dev].device;
1450 		}
1451 	}
1452 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1453 	/* Configure the first MAC address by default. */
1454 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1455 		DRV_LOG(ERR,
1456 			"port %u cannot get MAC address, is mlx5_en"
1457 			" loaded? (errno: %s)",
1458 			eth_dev->data->port_id, strerror(rte_errno));
1459 		err = ENODEV;
1460 		goto error;
1461 	}
1462 	DRV_LOG(INFO,
1463 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1464 		eth_dev->data->port_id,
1465 		mac.addr_bytes[0], mac.addr_bytes[1],
1466 		mac.addr_bytes[2], mac.addr_bytes[3],
1467 		mac.addr_bytes[4], mac.addr_bytes[5]);
1468 #ifdef RTE_LIBRTE_MLX5_DEBUG
1469 	{
1470 		char ifname[MLX5_NAMESIZE];
1471 
1472 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1473 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1474 				eth_dev->data->port_id, ifname);
1475 		else
1476 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1477 				eth_dev->data->port_id);
1478 	}
1479 #endif
1480 	/* Get actual MTU if possible. */
1481 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1482 	if (err) {
1483 		err = rte_errno;
1484 		goto error;
1485 	}
1486 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1487 		priv->mtu);
1488 	/* Initialize burst functions to prevent crashes before link-up. */
1489 	eth_dev->rx_pkt_burst = removed_rx_burst;
1490 	eth_dev->tx_pkt_burst = removed_tx_burst;
1491 	eth_dev->dev_ops = &mlx5_dev_ops;
1492 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1493 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1494 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1495 	/* Register MAC address. */
1496 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1497 	if (config->vf && config->vf_nl_en)
1498 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1499 				      mlx5_ifindex(eth_dev),
1500 				      eth_dev->data->mac_addrs,
1501 				      MLX5_MAX_MAC_ADDRESSES);
1502 	priv->flows = 0;
1503 	priv->ctrl_flows = 0;
1504 	rte_spinlock_init(&priv->flow_list_lock);
1505 	TAILQ_INIT(&priv->flow_meters);
1506 	TAILQ_INIT(&priv->flow_meter_profiles);
1507 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1508 	mlx5_glue->dv_set_context_attr(sh->ctx,
1509 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1510 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1511 				.alloc = &mlx5_alloc_verbs_buf,
1512 				.free = &mlx5_free_verbs_buf,
1513 				.data = sh,
1514 			}));
1515 	/* Bring Ethernet device up. */
1516 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1517 		eth_dev->data->port_id);
1518 	mlx5_set_link_up(eth_dev);
1519 	/*
1520 	 * Even though the interrupt handler is not installed yet,
1521 	 * interrupts will still trigger on the async_fd from
1522 	 * Verbs context returned by ibv_open_device().
1523 	 */
1524 	mlx5_link_update(eth_dev, 0);
1525 #ifdef HAVE_MLX5DV_DR_ESWITCH
1526 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1527 	      (switch_info->representor || switch_info->master)))
1528 		config->dv_esw_en = 0;
1529 #else
1530 	config->dv_esw_en = 0;
1531 #endif
1532 	/* Detect minimal data bytes to inline. */
1533 	mlx5_set_min_inline(spawn, config);
1534 	/* Store device configuration on private structure. */
1535 	priv->config = *config;
1536 	/* Create context for virtual machine VLAN workaround. */
1537 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1538 	if (config->dv_flow_en) {
1539 		err = mlx5_alloc_shared_dr(priv);
1540 		if (err)
1541 			goto error;
1542 	}
1543 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1544 		priv->obj_ops = devx_obj_ops;
1545 		priv->obj_ops.drop_action_create =
1546 						ibv_obj_ops.drop_action_create;
1547 		priv->obj_ops.drop_action_destroy =
1548 						ibv_obj_ops.drop_action_destroy;
1549 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1550 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1551 #else
1552 		if (config->dv_esw_en)
1553 			priv->obj_ops.txq_obj_modify =
1554 						ibv_obj_ops.txq_obj_modify;
1555 #endif
1556 		/* Use specific wrappers for Tx object. */
1557 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1558 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1559 		mlx5_queue_counter_id_prepare(eth_dev);
1560 
1561 	} else {
1562 		priv->obj_ops = ibv_obj_ops;
1563 	}
1564 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1565 	if (!priv->drop_queue.hrxq)
1566 		goto error;
1567 	/* Supported Verbs flow priority number detection. */
1568 	err = mlx5_flow_discover_priorities(eth_dev);
1569 	if (err < 0) {
1570 		err = -err;
1571 		goto error;
1572 	}
1573 	priv->config.flow_prio = err;
1574 	if (!priv->config.dv_esw_en &&
1575 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1576 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1577 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1578 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1579 	}
1580 	mlx5_set_metadata_mask(eth_dev);
1581 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1582 	    !priv->sh->dv_regc0_mask) {
1583 		DRV_LOG(ERR, "metadata mode %u is not supported "
1584 			     "(no metadata reg_c[0] is available)",
1585 			     priv->config.dv_xmeta_en);
1586 			err = ENOTSUP;
1587 			goto error;
1588 	}
1589 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1590 			     mlx5_hrxq_create_cb,
1591 			     mlx5_hrxq_match_cb,
1592 			     mlx5_hrxq_remove_cb);
1593 	/* Query availability of metadata reg_c's. */
1594 	err = mlx5_flow_discover_mreg_c(eth_dev);
1595 	if (err < 0) {
1596 		err = -err;
1597 		goto error;
1598 	}
1599 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1600 		DRV_LOG(DEBUG,
1601 			"port %u extensive metadata register is not supported",
1602 			eth_dev->data->port_id);
1603 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1604 			DRV_LOG(ERR, "metadata mode %u is not supported "
1605 				     "(no metadata registers available)",
1606 				     priv->config.dv_xmeta_en);
1607 			err = ENOTSUP;
1608 			goto error;
1609 		}
1610 	}
1611 	if (priv->config.dv_flow_en &&
1612 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1613 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1614 	    priv->sh->dv_regc0_mask) {
1615 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1616 						      MLX5_FLOW_MREG_HTABLE_SZ,
1617 						      0, 0,
1618 						      flow_dv_mreg_create_cb,
1619 						      flow_dv_mreg_match_cb,
1620 						      flow_dv_mreg_remove_cb);
1621 		if (!priv->mreg_cp_tbl) {
1622 			err = ENOMEM;
1623 			goto error;
1624 		}
1625 		priv->mreg_cp_tbl->ctx = eth_dev;
1626 	}
1627 	rte_spinlock_init(&priv->shared_act_sl);
1628 	mlx5_flow_counter_mode_config(eth_dev);
1629 	if (priv->config.dv_flow_en)
1630 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1631 	return eth_dev;
1632 error:
1633 	if (priv) {
1634 		if (priv->mreg_cp_tbl)
1635 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1636 		if (priv->sh)
1637 			mlx5_os_free_shared_dr(priv);
1638 		if (priv->nl_socket_route >= 0)
1639 			close(priv->nl_socket_route);
1640 		if (priv->nl_socket_rdma >= 0)
1641 			close(priv->nl_socket_rdma);
1642 		if (priv->vmwa_context)
1643 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1644 		if (eth_dev && priv->drop_queue.hrxq)
1645 			mlx5_drop_action_destroy(eth_dev);
1646 		if (own_domain_id)
1647 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1648 		mlx5_cache_list_destroy(&priv->hrxqs);
1649 		mlx5_free(priv);
1650 		if (eth_dev != NULL)
1651 			eth_dev->data->dev_private = NULL;
1652 	}
1653 	if (eth_dev != NULL) {
1654 		/* mac_addrs must not be freed alone because part of
1655 		 * dev_private
1656 		 **/
1657 		eth_dev->data->mac_addrs = NULL;
1658 		rte_eth_dev_release_port(eth_dev);
1659 	}
1660 	if (sh)
1661 		mlx5_free_shared_dev_ctx(sh);
1662 	MLX5_ASSERT(err > 0);
1663 	rte_errno = err;
1664 	return NULL;
1665 }
1666 
1667 /**
1668  * Comparison callback to sort device data.
1669  *
1670  * This is meant to be used with qsort().
1671  *
1672  * @param a[in]
1673  *   Pointer to pointer to first data object.
1674  * @param b[in]
1675  *   Pointer to pointer to second data object.
1676  *
1677  * @return
1678  *   0 if both objects are equal, less than 0 if the first argument is less
1679  *   than the second, greater than 0 otherwise.
1680  */
1681 static int
1682 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1683 {
1684 	const struct mlx5_switch_info *si_a =
1685 		&((const struct mlx5_dev_spawn_data *)a)->info;
1686 	const struct mlx5_switch_info *si_b =
1687 		&((const struct mlx5_dev_spawn_data *)b)->info;
1688 	int ret;
1689 
1690 	/* Master device first. */
1691 	ret = si_b->master - si_a->master;
1692 	if (ret)
1693 		return ret;
1694 	/* Then representor devices. */
1695 	ret = si_b->representor - si_a->representor;
1696 	if (ret)
1697 		return ret;
1698 	/* Unidentified devices come last in no specific order. */
1699 	if (!si_a->representor)
1700 		return 0;
1701 	/* Order representors by name. */
1702 	return si_a->port_name - si_b->port_name;
1703 }
1704 
1705 /**
1706  * Match PCI information for possible slaves of bonding device.
1707  *
1708  * @param[in] ibv_dev
1709  *   Pointer to Infiniband device structure.
1710  * @param[in] pci_dev
1711  *   Pointer to PCI device structure to match PCI address.
1712  * @param[in] nl_rdma
1713  *   Netlink RDMA group socket handle.
1714  *
1715  * @return
1716  *   negative value if no bonding device found, otherwise
1717  *   positive index of slave PF in bonding.
1718  */
1719 static int
1720 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1721 			   const struct rte_pci_device *pci_dev,
1722 			   int nl_rdma)
1723 {
1724 	char ifname[IF_NAMESIZE + 1];
1725 	unsigned int ifindex;
1726 	unsigned int np, i;
1727 	FILE *file = NULL;
1728 	int pf = -1;
1729 
1730 	/*
1731 	 * Try to get master device name. If something goes
1732 	 * wrong suppose the lack of kernel support and no
1733 	 * bonding devices.
1734 	 */
1735 	if (nl_rdma < 0)
1736 		return -1;
1737 	if (!strstr(ibv_dev->name, "bond"))
1738 		return -1;
1739 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1740 	if (!np)
1741 		return -1;
1742 	/*
1743 	 * The Master device might not be on the predefined
1744 	 * port (not on port index 1, it is not garanted),
1745 	 * we have to scan all Infiniband device port and
1746 	 * find master.
1747 	 */
1748 	for (i = 1; i <= np; ++i) {
1749 		/* Check whether Infiniband port is populated. */
1750 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1751 		if (!ifindex)
1752 			continue;
1753 		if (!if_indextoname(ifindex, ifname))
1754 			continue;
1755 		/* Try to read bonding slave names from sysfs. */
1756 		MKSTR(slaves,
1757 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1758 		file = fopen(slaves, "r");
1759 		if (file)
1760 			break;
1761 	}
1762 	if (!file)
1763 		return -1;
1764 	/* Use safe format to check maximal buffer length. */
1765 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1766 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1767 		char tmp_str[IF_NAMESIZE + 32];
1768 		struct rte_pci_addr pci_addr;
1769 		struct mlx5_switch_info	info;
1770 
1771 		/* Process slave interface names in the loop. */
1772 		snprintf(tmp_str, sizeof(tmp_str),
1773 			 "/sys/class/net/%s", ifname);
1774 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1775 			DRV_LOG(WARNING, "can not get PCI address"
1776 					 " for netdev \"%s\"", ifname);
1777 			continue;
1778 		}
1779 		if (pci_dev->addr.domain != pci_addr.domain ||
1780 		    pci_dev->addr.bus != pci_addr.bus ||
1781 		    pci_dev->addr.devid != pci_addr.devid ||
1782 		    pci_dev->addr.function != pci_addr.function)
1783 			continue;
1784 		/* Slave interface PCI address match found. */
1785 		fclose(file);
1786 		snprintf(tmp_str, sizeof(tmp_str),
1787 			 "/sys/class/net/%s/phys_port_name", ifname);
1788 		file = fopen(tmp_str, "rb");
1789 		if (!file)
1790 			break;
1791 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1792 		if (fscanf(file, "%32s", tmp_str) == 1)
1793 			mlx5_translate_port_name(tmp_str, &info);
1794 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1795 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1796 			pf = info.port_name;
1797 		break;
1798 	}
1799 	if (file)
1800 		fclose(file);
1801 	return pf;
1802 }
1803 
1804 /**
1805  * DPDK callback to register a PCI device.
1806  *
1807  * This function spawns Ethernet devices out of a given PCI device.
1808  *
1809  * @param[in] pci_drv
1810  *   PCI driver structure (mlx5_driver).
1811  * @param[in] pci_dev
1812  *   PCI device information.
1813  *
1814  * @return
1815  *   0 on success, a negative errno value otherwise and rte_errno is set.
1816  */
1817 int
1818 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1819 		  struct rte_pci_device *pci_dev)
1820 {
1821 	struct ibv_device **ibv_list;
1822 	/*
1823 	 * Number of found IB Devices matching with requested PCI BDF.
1824 	 * nd != 1 means there are multiple IB devices over the same
1825 	 * PCI device and we have representors and master.
1826 	 */
1827 	unsigned int nd = 0;
1828 	/*
1829 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1830 	 * we have the single multiport IB device, and there may be
1831 	 * representors attached to some of found ports.
1832 	 */
1833 	unsigned int np = 0;
1834 	/*
1835 	 * Number of DPDK ethernet devices to Spawn - either over
1836 	 * multiple IB devices or multiple ports of single IB device.
1837 	 * Actually this is the number of iterations to spawn.
1838 	 */
1839 	unsigned int ns = 0;
1840 	/*
1841 	 * Bonding device
1842 	 *   < 0 - no bonding device (single one)
1843 	 *  >= 0 - bonding device (value is slave PF index)
1844 	 */
1845 	int bd = -1;
1846 	struct mlx5_dev_spawn_data *list = NULL;
1847 	struct mlx5_dev_config dev_config;
1848 	unsigned int dev_config_vf;
1849 	int ret;
1850 
1851 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1852 		mlx5_pmd_socket_init();
1853 	ret = mlx5_init_once();
1854 	if (ret) {
1855 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1856 			strerror(rte_errno));
1857 		return -rte_errno;
1858 	}
1859 	errno = 0;
1860 	ibv_list = mlx5_glue->get_device_list(&ret);
1861 	if (!ibv_list) {
1862 		rte_errno = errno ? errno : ENOSYS;
1863 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1864 		return -rte_errno;
1865 	}
1866 	/*
1867 	 * First scan the list of all Infiniband devices to find
1868 	 * matching ones, gathering into the list.
1869 	 */
1870 	struct ibv_device *ibv_match[ret + 1];
1871 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1872 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1873 	unsigned int i;
1874 
1875 	while (ret-- > 0) {
1876 		struct rte_pci_addr pci_addr;
1877 
1878 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1879 		bd = mlx5_device_bond_pci_match
1880 				(ibv_list[ret], pci_dev, nl_rdma);
1881 		if (bd >= 0) {
1882 			/*
1883 			 * Bonding device detected. Only one match is allowed,
1884 			 * the bonding is supported over multi-port IB device,
1885 			 * there should be no matches on representor PCI
1886 			 * functions or non VF LAG bonding devices with
1887 			 * specified address.
1888 			 */
1889 			if (nd) {
1890 				DRV_LOG(ERR,
1891 					"multiple PCI match on bonding device"
1892 					"\"%s\" found", ibv_list[ret]->name);
1893 				rte_errno = ENOENT;
1894 				ret = -rte_errno;
1895 				goto exit;
1896 			}
1897 			DRV_LOG(INFO, "PCI information matches for"
1898 				      " slave %d bonding device \"%s\"",
1899 				      bd, ibv_list[ret]->name);
1900 			ibv_match[nd++] = ibv_list[ret];
1901 			break;
1902 		}
1903 		if (mlx5_dev_to_pci_addr
1904 			(ibv_list[ret]->ibdev_path, &pci_addr))
1905 			continue;
1906 		if (pci_dev->addr.domain != pci_addr.domain ||
1907 		    pci_dev->addr.bus != pci_addr.bus ||
1908 		    pci_dev->addr.devid != pci_addr.devid ||
1909 		    pci_dev->addr.function != pci_addr.function)
1910 			continue;
1911 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1912 			ibv_list[ret]->name);
1913 		ibv_match[nd++] = ibv_list[ret];
1914 	}
1915 	ibv_match[nd] = NULL;
1916 	if (!nd) {
1917 		/* No device matches, just complain and bail out. */
1918 		DRV_LOG(WARNING,
1919 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1920 			" are kernel drivers loaded?",
1921 			pci_dev->addr.domain, pci_dev->addr.bus,
1922 			pci_dev->addr.devid, pci_dev->addr.function);
1923 		rte_errno = ENOENT;
1924 		ret = -rte_errno;
1925 		goto exit;
1926 	}
1927 	if (nd == 1) {
1928 		/*
1929 		 * Found single matching device may have multiple ports.
1930 		 * Each port may be representor, we have to check the port
1931 		 * number and check the representors existence.
1932 		 */
1933 		if (nl_rdma >= 0)
1934 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1935 		if (!np)
1936 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1937 					 " ports number", ibv_match[0]->name);
1938 		if (bd >= 0 && !np) {
1939 			DRV_LOG(ERR, "can not get ports"
1940 				     " for bonding device");
1941 			rte_errno = ENOENT;
1942 			ret = -rte_errno;
1943 			goto exit;
1944 		}
1945 	}
1946 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1947 	if (bd >= 0) {
1948 		/*
1949 		 * This may happen if there is VF LAG kernel support and
1950 		 * application is compiled with older rdma_core library.
1951 		 */
1952 		DRV_LOG(ERR,
1953 			"No kernel/verbs support for VF LAG bonding found.");
1954 		rte_errno = ENOTSUP;
1955 		ret = -rte_errno;
1956 		goto exit;
1957 	}
1958 #endif
1959 	/*
1960 	 * Now we can determine the maximal
1961 	 * amount of devices to be spawned.
1962 	 */
1963 	list = mlx5_malloc(MLX5_MEM_ZERO,
1964 			   sizeof(struct mlx5_dev_spawn_data) *
1965 			   (np ? np : nd),
1966 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1967 	if (!list) {
1968 		DRV_LOG(ERR, "spawn data array allocation failure");
1969 		rte_errno = ENOMEM;
1970 		ret = -rte_errno;
1971 		goto exit;
1972 	}
1973 	if (bd >= 0 || np > 1) {
1974 		/*
1975 		 * Single IB device with multiple ports found,
1976 		 * it may be E-Switch master device and representors.
1977 		 * We have to perform identification through the ports.
1978 		 */
1979 		MLX5_ASSERT(nl_rdma >= 0);
1980 		MLX5_ASSERT(ns == 0);
1981 		MLX5_ASSERT(nd == 1);
1982 		MLX5_ASSERT(np);
1983 		for (i = 1; i <= np; ++i) {
1984 			list[ns].max_port = np;
1985 			list[ns].phys_port = i;
1986 			list[ns].phys_dev = ibv_match[0];
1987 			list[ns].eth_dev = NULL;
1988 			list[ns].pci_dev = pci_dev;
1989 			list[ns].pf_bond = bd;
1990 			list[ns].ifindex = mlx5_nl_ifindex
1991 				(nl_rdma,
1992 				mlx5_os_get_dev_device_name
1993 						(list[ns].phys_dev), i);
1994 			if (!list[ns].ifindex) {
1995 				/*
1996 				 * No network interface index found for the
1997 				 * specified port, it means there is no
1998 				 * representor on this port. It's OK,
1999 				 * there can be disabled ports, for example
2000 				 * if sriov_numvfs < sriov_totalvfs.
2001 				 */
2002 				continue;
2003 			}
2004 			ret = -1;
2005 			if (nl_route >= 0)
2006 				ret = mlx5_nl_switch_info
2007 					       (nl_route,
2008 						list[ns].ifindex,
2009 						&list[ns].info);
2010 			if (ret || (!list[ns].info.representor &&
2011 				    !list[ns].info.master)) {
2012 				/*
2013 				 * We failed to recognize representors with
2014 				 * Netlink, let's try to perform the task
2015 				 * with sysfs.
2016 				 */
2017 				ret =  mlx5_sysfs_switch_info
2018 						(list[ns].ifindex,
2019 						 &list[ns].info);
2020 			}
2021 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2022 			if (!ret && bd >= 0) {
2023 				switch (list[ns].info.name_type) {
2024 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2025 					if (list[ns].info.port_name == bd)
2026 						ns++;
2027 					break;
2028 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2029 					/* Fallthrough */
2030 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2031 					if (list[ns].info.pf_num == bd)
2032 						ns++;
2033 					break;
2034 				default:
2035 					break;
2036 				}
2037 				continue;
2038 			}
2039 #endif
2040 			if (!ret && (list[ns].info.representor ^
2041 				     list[ns].info.master))
2042 				ns++;
2043 		}
2044 		if (!ns) {
2045 			DRV_LOG(ERR,
2046 				"unable to recognize master/representors"
2047 				" on the IB device with multiple ports");
2048 			rte_errno = ENOENT;
2049 			ret = -rte_errno;
2050 			goto exit;
2051 		}
2052 	} else {
2053 		/*
2054 		 * The existence of several matching entries (nd > 1) means
2055 		 * port representors have been instantiated. No existing Verbs
2056 		 * call nor sysfs entries can tell them apart, this can only
2057 		 * be done through Netlink calls assuming kernel drivers are
2058 		 * recent enough to support them.
2059 		 *
2060 		 * In the event of identification failure through Netlink,
2061 		 * try again through sysfs, then:
2062 		 *
2063 		 * 1. A single IB device matches (nd == 1) with single
2064 		 *    port (np=0/1) and is not a representor, assume
2065 		 *    no switch support.
2066 		 *
2067 		 * 2. Otherwise no safe assumptions can be made;
2068 		 *    complain louder and bail out.
2069 		 */
2070 		for (i = 0; i != nd; ++i) {
2071 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2072 			list[ns].max_port = 1;
2073 			list[ns].phys_port = 1;
2074 			list[ns].phys_dev = ibv_match[i];
2075 			list[ns].eth_dev = NULL;
2076 			list[ns].pci_dev = pci_dev;
2077 			list[ns].pf_bond = -1;
2078 			list[ns].ifindex = 0;
2079 			if (nl_rdma >= 0)
2080 				list[ns].ifindex = mlx5_nl_ifindex
2081 				(nl_rdma,
2082 				mlx5_os_get_dev_device_name
2083 						(list[ns].phys_dev), 1);
2084 			if (!list[ns].ifindex) {
2085 				char ifname[IF_NAMESIZE];
2086 
2087 				/*
2088 				 * Netlink failed, it may happen with old
2089 				 * ib_core kernel driver (before 4.16).
2090 				 * We can assume there is old driver because
2091 				 * here we are processing single ports IB
2092 				 * devices. Let's try sysfs to retrieve
2093 				 * the ifindex. The method works for
2094 				 * master device only.
2095 				 */
2096 				if (nd > 1) {
2097 					/*
2098 					 * Multiple devices found, assume
2099 					 * representors, can not distinguish
2100 					 * master/representor and retrieve
2101 					 * ifindex via sysfs.
2102 					 */
2103 					continue;
2104 				}
2105 				ret = mlx5_get_ifname_sysfs
2106 					(ibv_match[i]->ibdev_path, ifname);
2107 				if (!ret)
2108 					list[ns].ifindex =
2109 						if_nametoindex(ifname);
2110 				if (!list[ns].ifindex) {
2111 					/*
2112 					 * No network interface index found
2113 					 * for the specified device, it means
2114 					 * there it is neither representor
2115 					 * nor master.
2116 					 */
2117 					continue;
2118 				}
2119 			}
2120 			ret = -1;
2121 			if (nl_route >= 0)
2122 				ret = mlx5_nl_switch_info
2123 					       (nl_route,
2124 						list[ns].ifindex,
2125 						&list[ns].info);
2126 			if (ret || (!list[ns].info.representor &&
2127 				    !list[ns].info.master)) {
2128 				/*
2129 				 * We failed to recognize representors with
2130 				 * Netlink, let's try to perform the task
2131 				 * with sysfs.
2132 				 */
2133 				ret =  mlx5_sysfs_switch_info
2134 						(list[ns].ifindex,
2135 						 &list[ns].info);
2136 			}
2137 			if (!ret && (list[ns].info.representor ^
2138 				     list[ns].info.master)) {
2139 				ns++;
2140 			} else if ((nd == 1) &&
2141 				   !list[ns].info.representor &&
2142 				   !list[ns].info.master) {
2143 				/*
2144 				 * Single IB device with
2145 				 * one physical port and
2146 				 * attached network device.
2147 				 * May be SRIOV is not enabled
2148 				 * or there is no representors.
2149 				 */
2150 				DRV_LOG(INFO, "no E-Switch support detected");
2151 				ns++;
2152 				break;
2153 			}
2154 		}
2155 		if (!ns) {
2156 			DRV_LOG(ERR,
2157 				"unable to recognize master/representors"
2158 				" on the multiple IB devices");
2159 			rte_errno = ENOENT;
2160 			ret = -rte_errno;
2161 			goto exit;
2162 		}
2163 	}
2164 	MLX5_ASSERT(ns);
2165 	/*
2166 	 * Sort list to probe devices in natural order for users convenience
2167 	 * (i.e. master first, then representors from lowest to highest ID).
2168 	 */
2169 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2170 	/* Device specific configuration. */
2171 	switch (pci_dev->id.device_id) {
2172 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2173 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2174 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2175 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2176 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2177 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2178 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2179 		dev_config_vf = 1;
2180 		break;
2181 	default:
2182 		dev_config_vf = 0;
2183 		break;
2184 	}
2185 	for (i = 0; i != ns; ++i) {
2186 		uint32_t restore;
2187 
2188 		/* Default configuration. */
2189 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2190 		dev_config.vf = dev_config_vf;
2191 		dev_config.mps = MLX5_ARG_UNSET;
2192 		dev_config.dbnc = MLX5_ARG_UNSET;
2193 		dev_config.rx_vec_en = 1;
2194 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2195 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2196 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2197 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2198 		dev_config.vf_nl_en = 1;
2199 		dev_config.mr_ext_memseg_en = 1;
2200 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2201 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2202 		dev_config.dv_esw_en = 1;
2203 		dev_config.dv_flow_en = 1;
2204 		dev_config.decap_en = 1;
2205 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2206 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2207 						 &list[i],
2208 						 &dev_config);
2209 		if (!list[i].eth_dev) {
2210 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2211 				break;
2212 			/* Device is disabled or already spawned. Ignore it. */
2213 			continue;
2214 		}
2215 		restore = list[i].eth_dev->data->dev_flags;
2216 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2217 		/* Restore non-PCI flags cleared by the above call. */
2218 		list[i].eth_dev->data->dev_flags |= restore;
2219 		rte_eth_dev_probing_finish(list[i].eth_dev);
2220 	}
2221 	if (i != ns) {
2222 		DRV_LOG(ERR,
2223 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2224 			" encountering an error: %s",
2225 			pci_dev->addr.domain, pci_dev->addr.bus,
2226 			pci_dev->addr.devid, pci_dev->addr.function,
2227 			strerror(rte_errno));
2228 		ret = -rte_errno;
2229 		/* Roll back. */
2230 		while (i--) {
2231 			if (!list[i].eth_dev)
2232 				continue;
2233 			mlx5_dev_close(list[i].eth_dev);
2234 			/* mac_addrs must not be freed because in dev_private */
2235 			list[i].eth_dev->data->mac_addrs = NULL;
2236 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2237 		}
2238 		/* Restore original error. */
2239 		rte_errno = -ret;
2240 	} else {
2241 		ret = 0;
2242 	}
2243 exit:
2244 	/*
2245 	 * Do the routine cleanup:
2246 	 * - close opened Netlink sockets
2247 	 * - free allocated spawn data array
2248 	 * - free the Infiniband device list
2249 	 */
2250 	if (nl_rdma >= 0)
2251 		close(nl_rdma);
2252 	if (nl_route >= 0)
2253 		close(nl_route);
2254 	if (list)
2255 		mlx5_free(list);
2256 	MLX5_ASSERT(ibv_list);
2257 	mlx5_glue->free_device_list(ibv_list);
2258 	return ret;
2259 }
2260 
2261 static int
2262 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2263 {
2264 	char *env;
2265 	int value;
2266 
2267 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2268 	/* Get environment variable to store. */
2269 	env = getenv(MLX5_SHUT_UP_BF);
2270 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2271 	if (config->dbnc == MLX5_ARG_UNSET)
2272 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2273 	else
2274 		setenv(MLX5_SHUT_UP_BF,
2275 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2276 	return value;
2277 }
2278 
2279 static void
2280 mlx5_restore_doorbell_mapping_env(int value)
2281 {
2282 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2283 	/* Restore the original environment variable state. */
2284 	if (value == MLX5_ARG_UNSET)
2285 		unsetenv(MLX5_SHUT_UP_BF);
2286 	else
2287 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2288 }
2289 
2290 /**
2291  * Extract pdn of PD object using DV API.
2292  *
2293  * @param[in] pd
2294  *   Pointer to the verbs PD object.
2295  * @param[out] pdn
2296  *   Pointer to the PD object number variable.
2297  *
2298  * @return
2299  *   0 on success, error value otherwise.
2300  */
2301 int
2302 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2303 {
2304 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2305 	struct mlx5dv_obj obj;
2306 	struct mlx5dv_pd pd_info;
2307 	int ret = 0;
2308 
2309 	obj.pd.in = pd;
2310 	obj.pd.out = &pd_info;
2311 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2312 	if (ret) {
2313 		DRV_LOG(DEBUG, "Fail to get PD object info");
2314 		return ret;
2315 	}
2316 	*pdn = pd_info.pdn;
2317 	return 0;
2318 #else
2319 	(void)pd;
2320 	(void)pdn;
2321 	return -ENOTSUP;
2322 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2323 }
2324 
2325 /**
2326  * Function API to open IB device.
2327  *
2328  * This function calls the Linux glue APIs to open a device.
2329  *
2330  * @param[in] spawn
2331  *   Pointer to the IB device attributes (name, port, etc).
2332  * @param[out] config
2333  *   Pointer to device configuration structure.
2334  * @param[out] sh
2335  *   Pointer to shared context structure.
2336  *
2337  * @return
2338  *   0 on success, a positive error value otherwise.
2339  */
2340 int
2341 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2342 		     const struct mlx5_dev_config *config,
2343 		     struct mlx5_dev_ctx_shared *sh)
2344 {
2345 	int dbmap_env;
2346 	int err = 0;
2347 
2348 	sh->numa_node = spawn->pci_dev->device.numa_node;
2349 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2350 	/*
2351 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2352 	 * before the device creation. The rdma_core library
2353 	 * checks the variable at device creation and
2354 	 * stores the result internally.
2355 	 */
2356 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2357 	/* Try to open IB device with DV first, then usual Verbs. */
2358 	errno = 0;
2359 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2360 	if (sh->ctx) {
2361 		sh->devx = 1;
2362 		DRV_LOG(DEBUG, "DevX is supported");
2363 		/* The device is created, no need for environment. */
2364 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2365 	} else {
2366 		/* The environment variable is still configured. */
2367 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2368 		err = errno ? errno : ENODEV;
2369 		/*
2370 		 * The environment variable is not needed anymore,
2371 		 * all device creation attempts are completed.
2372 		 */
2373 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2374 		if (!sh->ctx)
2375 			return err;
2376 		DRV_LOG(DEBUG, "DevX is NOT supported");
2377 		err = 0;
2378 	}
2379 	if (!err && sh->ctx) {
2380 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2381 		mlx5_glue->dv_set_context_attr(sh->ctx,
2382 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2383 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2384 				.alloc = &mlx5_alloc_verbs_buf,
2385 				.free = &mlx5_free_verbs_buf,
2386 				.data = sh,
2387 			}));
2388 	}
2389 	return err;
2390 }
2391 
2392 /**
2393  * Install shared asynchronous device events handler.
2394  * This function is implemented to support event sharing
2395  * between multiple ports of single IB device.
2396  *
2397  * @param sh
2398  *   Pointer to mlx5_dev_ctx_shared object.
2399  */
2400 void
2401 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2402 {
2403 	int ret;
2404 	int flags;
2405 
2406 	sh->intr_handle.fd = -1;
2407 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2408 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2409 		    F_SETFL, flags | O_NONBLOCK);
2410 	if (ret) {
2411 		DRV_LOG(INFO, "failed to change file descriptor async event"
2412 			" queue");
2413 	} else {
2414 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2415 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2416 		if (rte_intr_callback_register(&sh->intr_handle,
2417 					mlx5_dev_interrupt_handler, sh)) {
2418 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2419 			sh->intr_handle.fd = -1;
2420 		}
2421 	}
2422 	if (sh->devx) {
2423 #ifdef HAVE_IBV_DEVX_ASYNC
2424 		sh->intr_handle_devx.fd = -1;
2425 		sh->devx_comp =
2426 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2427 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2428 		if (!devx_comp) {
2429 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2430 			return;
2431 		}
2432 		flags = fcntl(devx_comp->fd, F_GETFL);
2433 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2434 		if (ret) {
2435 			DRV_LOG(INFO, "failed to change file descriptor"
2436 				" devx comp");
2437 			return;
2438 		}
2439 		sh->intr_handle_devx.fd = devx_comp->fd;
2440 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2441 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2442 					mlx5_dev_interrupt_handler_devx, sh)) {
2443 			DRV_LOG(INFO, "Fail to install the devx shared"
2444 				" interrupt.");
2445 			sh->intr_handle_devx.fd = -1;
2446 		}
2447 #endif /* HAVE_IBV_DEVX_ASYNC */
2448 	}
2449 }
2450 
2451 /**
2452  * Uninstall shared asynchronous device events handler.
2453  * This function is implemented to support event sharing
2454  * between multiple ports of single IB device.
2455  *
2456  * @param dev
2457  *   Pointer to mlx5_dev_ctx_shared object.
2458  */
2459 void
2460 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2461 {
2462 	if (sh->intr_handle.fd >= 0)
2463 		mlx5_intr_callback_unregister(&sh->intr_handle,
2464 					      mlx5_dev_interrupt_handler, sh);
2465 #ifdef HAVE_IBV_DEVX_ASYNC
2466 	if (sh->intr_handle_devx.fd >= 0)
2467 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2468 				  mlx5_dev_interrupt_handler_devx, sh);
2469 	if (sh->devx_comp)
2470 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2471 #endif
2472 }
2473 
2474 /**
2475  * Read statistics by a named counter.
2476  *
2477  * @param[in] priv
2478  *   Pointer to the private device data structure.
2479  * @param[in] ctr_name
2480  *   Pointer to the name of the statistic counter to read
2481  * @param[out] stat
2482  *   Pointer to read statistic value.
2483  * @return
2484  *   0 on success and stat is valud, 1 if failed to read the value
2485  *   rte_errno is set.
2486  *
2487  */
2488 int
2489 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2490 		      uint64_t *stat)
2491 {
2492 	int fd;
2493 
2494 	if (priv->sh) {
2495 		if (priv->q_counters != NULL &&
2496 		    strcmp(ctr_name, "out_of_buffer") == 0)
2497 			return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2498 							   0, (uint32_t *)stat);
2499 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2500 		      priv->sh->ibdev_path,
2501 		      priv->dev_port,
2502 		      ctr_name);
2503 		fd = open(path, O_RDONLY);
2504 		/*
2505 		 * in switchdev the file location is not per port
2506 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2507 		 */
2508 		if (fd == -1) {
2509 			MKSTR(path1, "%s/hw_counters/%s",
2510 			      priv->sh->ibdev_path,
2511 			      ctr_name);
2512 			fd = open(path1, O_RDONLY);
2513 		}
2514 		if (fd != -1) {
2515 			char buf[21] = {'\0'};
2516 			ssize_t n = read(fd, buf, sizeof(buf));
2517 
2518 			close(fd);
2519 			if (n != -1) {
2520 				*stat = strtoull(buf, NULL, 10);
2521 				return 0;
2522 			}
2523 		}
2524 	}
2525 	*stat = 0;
2526 	return 1;
2527 }
2528 
2529 /**
2530  * Set the reg_mr and dereg_mr call backs
2531  *
2532  * @param reg_mr_cb[out]
2533  *   Pointer to reg_mr func
2534  * @param dereg_mr_cb[out]
2535  *   Pointer to dereg_mr func
2536  *
2537  */
2538 void
2539 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2540 		      mlx5_dereg_mr_t *dereg_mr_cb)
2541 {
2542 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2543 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2544 }
2545 
2546 /**
2547  * Remove a MAC address from device
2548  *
2549  * @param dev
2550  *   Pointer to Ethernet device structure.
2551  * @param index
2552  *   MAC address index.
2553  */
2554 void
2555 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2556 {
2557 	struct mlx5_priv *priv = dev->data->dev_private;
2558 	const int vf = priv->config.vf;
2559 
2560 	if (vf)
2561 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2562 					mlx5_ifindex(dev), priv->mac_own,
2563 					&dev->data->mac_addrs[index], index);
2564 }
2565 
2566 /**
2567  * Adds a MAC address to the device
2568  *
2569  * @param dev
2570  *   Pointer to Ethernet device structure.
2571  * @param mac_addr
2572  *   MAC address to register.
2573  * @param index
2574  *   MAC address index.
2575  *
2576  * @return
2577  *   0 on success, a negative errno value otherwise
2578  */
2579 int
2580 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2581 		     uint32_t index)
2582 {
2583 	struct mlx5_priv *priv = dev->data->dev_private;
2584 	const int vf = priv->config.vf;
2585 	int ret = 0;
2586 
2587 	if (vf)
2588 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2589 					   mlx5_ifindex(dev), priv->mac_own,
2590 					   mac, index);
2591 	return ret;
2592 }
2593 
2594 /**
2595  * Modify a VF MAC address
2596  *
2597  * @param priv
2598  *   Pointer to device private data.
2599  * @param mac_addr
2600  *   MAC address to modify into.
2601  * @param iface_idx
2602  *   Net device interface index
2603  * @param vf_index
2604  *   VF index
2605  *
2606  * @return
2607  *   0 on success, a negative errno value otherwise
2608  */
2609 int
2610 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2611 			   unsigned int iface_idx,
2612 			   struct rte_ether_addr *mac_addr,
2613 			   int vf_index)
2614 {
2615 	return mlx5_nl_vf_mac_addr_modify
2616 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2617 }
2618 
2619 /**
2620  * Set device promiscuous mode
2621  *
2622  * @param dev
2623  *   Pointer to Ethernet device structure.
2624  * @param enable
2625  *   0 - promiscuous is disabled, otherwise - enabled
2626  *
2627  * @return
2628  *   0 on success, a negative error value otherwise
2629  */
2630 int
2631 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2632 {
2633 	struct mlx5_priv *priv = dev->data->dev_private;
2634 
2635 	return mlx5_nl_promisc(priv->nl_socket_route,
2636 			       mlx5_ifindex(dev), !!enable);
2637 }
2638 
2639 /**
2640  * Set device promiscuous mode
2641  *
2642  * @param dev
2643  *   Pointer to Ethernet device structure.
2644  * @param enable
2645  *   0 - all multicase is disabled, otherwise - enabled
2646  *
2647  * @return
2648  *   0 on success, a negative error value otherwise
2649  */
2650 int
2651 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2652 {
2653 	struct mlx5_priv *priv = dev->data->dev_private;
2654 
2655 	return mlx5_nl_allmulti(priv->nl_socket_route,
2656 				mlx5_ifindex(dev), !!enable);
2657 }
2658 
2659 /**
2660  * Flush device MAC addresses
2661  *
2662  * @param dev
2663  *   Pointer to Ethernet device structure.
2664  *
2665  */
2666 void
2667 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2668 {
2669 	struct mlx5_priv *priv = dev->data->dev_private;
2670 
2671 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2672 			       dev->data->mac_addrs,
2673 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2674 }
2675