xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision b7b78a089c454d42eb654360eeecb1e2f15e6cd8)
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 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
755 			if (eth_da.representor_ports[i] ==
756 			    (uint16_t)switch_info->port_name)
757 				break;
758 		if (i == eth_da.nb_representor_ports) {
759 			rte_errno = EBUSY;
760 			return NULL;
761 		}
762 	}
763 	/* Build device name. */
764 	if (spawn->pf_bond <  0) {
765 		/* Single device. */
766 		if (!switch_info->representor)
767 			strlcpy(name, dpdk_dev->name, sizeof(name));
768 		else
769 			snprintf(name, sizeof(name), "%s_representor_%u",
770 				 dpdk_dev->name, switch_info->port_name);
771 	} else {
772 		/* Bonding device. */
773 		if (!switch_info->representor)
774 			snprintf(name, sizeof(name), "%s_%s",
775 				 dpdk_dev->name,
776 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
777 		else
778 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
779 				 dpdk_dev->name,
780 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
781 				 switch_info->port_name);
782 	}
783 	/* check if the device is already spawned */
784 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
785 		rte_errno = EEXIST;
786 		return NULL;
787 	}
788 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
789 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
790 		struct mlx5_mp_id mp_id;
791 
792 		eth_dev = rte_eth_dev_attach_secondary(name);
793 		if (eth_dev == NULL) {
794 			DRV_LOG(ERR, "can not attach rte ethdev");
795 			rte_errno = ENOMEM;
796 			return NULL;
797 		}
798 		priv = eth_dev->data->dev_private;
799 		if (priv->sh->bond_dev != UINT16_MAX)
800 			/* For bonding port, use primary PCI device. */
801 			eth_dev->device =
802 				rte_eth_devices[priv->sh->bond_dev].device;
803 		else
804 			eth_dev->device = dpdk_dev;
805 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
806 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
807 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
808 		err = mlx5_proc_priv_init(eth_dev);
809 		if (err)
810 			return NULL;
811 		mp_id.port_id = eth_dev->data->port_id;
812 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
813 		/* Receive command fd from primary process */
814 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
815 		if (err < 0)
816 			goto err_secondary;
817 		/* Remap UAR for Tx queues. */
818 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
819 		if (err)
820 			goto err_secondary;
821 		/*
822 		 * Ethdev pointer is still required as input since
823 		 * the primary device is not accessible from the
824 		 * secondary process.
825 		 */
826 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
827 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
828 		return eth_dev;
829 err_secondary:
830 		mlx5_dev_close(eth_dev);
831 		return NULL;
832 	}
833 	/*
834 	 * Some parameters ("tx_db_nc" in particularly) are needed in
835 	 * advance to create dv/verbs device context. We proceed the
836 	 * devargs here to get ones, and later proceed devargs again
837 	 * to override some hardware settings.
838 	 */
839 	err = mlx5_args(config, dpdk_dev->devargs);
840 	if (err) {
841 		err = rte_errno;
842 		DRV_LOG(ERR, "failed to process device arguments: %s",
843 			strerror(rte_errno));
844 		goto error;
845 	}
846 	if (config->dv_miss_info) {
847 		if (switch_info->master || switch_info->representor)
848 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
849 	}
850 	mlx5_malloc_mem_select(config->sys_mem_en);
851 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
852 	if (!sh)
853 		return NULL;
854 	config->devx = sh->devx;
855 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
856 	config->dest_tir = 1;
857 #endif
858 #ifdef HAVE_IBV_MLX5_MOD_SWP
859 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
860 #endif
861 	/*
862 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
863 	 * as all ConnectX-5 devices.
864 	 */
865 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
866 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
867 #endif
868 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
869 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
870 #endif
871 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
872 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
873 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
874 			DRV_LOG(DEBUG, "enhanced MPW is supported");
875 			mps = MLX5_MPW_ENHANCED;
876 		} else {
877 			DRV_LOG(DEBUG, "MPW is supported");
878 			mps = MLX5_MPW;
879 		}
880 	} else {
881 		DRV_LOG(DEBUG, "MPW isn't supported");
882 		mps = MLX5_MPW_DISABLED;
883 	}
884 #ifdef HAVE_IBV_MLX5_MOD_SWP
885 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
886 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
887 	DRV_LOG(DEBUG, "SWP support: %u", swp);
888 #endif
889 	config->swp = !!swp;
890 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
891 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
892 		struct mlx5dv_striding_rq_caps mprq_caps =
893 			dv_attr.striding_rq_caps;
894 
895 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
896 			mprq_caps.min_single_stride_log_num_of_bytes);
897 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
898 			mprq_caps.max_single_stride_log_num_of_bytes);
899 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
900 			mprq_caps.min_single_wqe_log_num_of_strides);
901 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
902 			mprq_caps.max_single_wqe_log_num_of_strides);
903 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
904 			mprq_caps.supported_qpts);
905 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
906 		mprq = 1;
907 		mprq_min_stride_size_n =
908 			mprq_caps.min_single_stride_log_num_of_bytes;
909 		mprq_max_stride_size_n =
910 			mprq_caps.max_single_stride_log_num_of_bytes;
911 		mprq_min_stride_num_n =
912 			mprq_caps.min_single_wqe_log_num_of_strides;
913 		mprq_max_stride_num_n =
914 			mprq_caps.max_single_wqe_log_num_of_strides;
915 	}
916 #endif
917 	/* Rx CQE compression is enabled by default. */
918 	config->cqe_comp = 1;
919 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
920 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
921 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
922 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
923 			     (dv_attr.tunnel_offloads_caps &
924 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
925 			     (dv_attr.tunnel_offloads_caps &
926 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
927 	}
928 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
929 		tunnel_en ? "" : "not ");
930 #else
931 	DRV_LOG(WARNING,
932 		"tunnel offloading disabled due to old OFED/rdma-core version");
933 #endif
934 	config->tunnel_en = tunnel_en;
935 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
936 	mpls_en = ((dv_attr.tunnel_offloads_caps &
937 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
938 		   (dv_attr.tunnel_offloads_caps &
939 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
940 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
941 		mpls_en ? "" : "not ");
942 #else
943 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
944 		" old OFED/rdma-core version or firmware configuration");
945 #endif
946 	config->mpls_en = mpls_en;
947 	/* Check port status. */
948 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
949 	if (err) {
950 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
951 		goto error;
952 	}
953 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
954 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
955 		err = EINVAL;
956 		goto error;
957 	}
958 	if (port_attr.state != IBV_PORT_ACTIVE)
959 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
960 			mlx5_glue->port_state_str(port_attr.state),
961 			port_attr.state);
962 	/* Allocate private eth device data. */
963 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
964 			   sizeof(*priv),
965 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
966 	if (priv == NULL) {
967 		DRV_LOG(ERR, "priv allocation failure");
968 		err = ENOMEM;
969 		goto error;
970 	}
971 	priv->sh = sh;
972 	priv->dev_port = spawn->phys_port;
973 	priv->pci_dev = spawn->pci_dev;
974 	priv->mtu = RTE_ETHER_MTU;
975 	/* Some internal functions rely on Netlink sockets, open them now. */
976 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
977 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
978 	priv->representor = !!switch_info->representor;
979 	priv->master = !!switch_info->master;
980 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
981 	priv->vport_meta_tag = 0;
982 	priv->vport_meta_mask = 0;
983 	priv->pf_bond = spawn->pf_bond;
984 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
985 	/*
986 	 * The DevX port query API is implemented. E-Switch may use
987 	 * either vport or reg_c[0] metadata register to match on
988 	 * vport index. The engaged part of metadata register is
989 	 * defined by mask.
990 	 */
991 	if (switch_info->representor || switch_info->master) {
992 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
993 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
994 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
995 						 &devx_port);
996 		if (err) {
997 			DRV_LOG(WARNING,
998 				"can't query devx port %d on device %s",
999 				spawn->phys_port,
1000 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1001 			devx_port.comp_mask = 0;
1002 		}
1003 	}
1004 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1005 		priv->vport_meta_tag = devx_port.reg_c_0.value;
1006 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
1007 		if (!priv->vport_meta_mask) {
1008 			DRV_LOG(ERR, "vport zero mask for port %d"
1009 				     " on bonding device %s",
1010 				     spawn->phys_port,
1011 				     mlx5_os_get_dev_device_name
1012 							(spawn->phys_dev));
1013 			err = ENOTSUP;
1014 			goto error;
1015 		}
1016 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1017 			DRV_LOG(ERR, "invalid vport tag for port %d"
1018 				     " on bonding device %s",
1019 				     spawn->phys_port,
1020 				     mlx5_os_get_dev_device_name
1021 							(spawn->phys_dev));
1022 			err = ENOTSUP;
1023 			goto error;
1024 		}
1025 	}
1026 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1027 		priv->vport_id = devx_port.vport_num;
1028 	} else if (spawn->pf_bond >= 0) {
1029 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1030 			     " on bonding device %s",
1031 			     spawn->phys_port,
1032 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1033 		err = ENOTSUP;
1034 		goto error;
1035 	} else {
1036 		/* Suppose vport index in compatible way. */
1037 		priv->vport_id = switch_info->representor ?
1038 				 switch_info->port_name + 1 : -1;
1039 	}
1040 #else
1041 	/*
1042 	 * Kernel/rdma_core support single E-Switch per PF configurations
1043 	 * only and vport_id field contains the vport index for
1044 	 * associated VF, which is deduced from representor port name.
1045 	 * For example, let's have the IB device port 10, it has
1046 	 * attached network device eth0, which has port name attribute
1047 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1048 	 * as 3 (2+1). This assigning schema should be changed if the
1049 	 * multiple E-Switch instances per PF configurations or/and PCI
1050 	 * subfunctions are added.
1051 	 */
1052 	priv->vport_id = switch_info->representor ?
1053 			 switch_info->port_name + 1 : -1;
1054 #endif
1055 	/* representor_id field keeps the unmodified VF index. */
1056 	priv->representor_id = switch_info->representor ?
1057 			       switch_info->port_name : -1;
1058 	/*
1059 	 * Look for sibling devices in order to reuse their switch domain
1060 	 * if any, otherwise allocate one.
1061 	 */
1062 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1063 		const struct mlx5_priv *opriv =
1064 			rte_eth_devices[port_id].data->dev_private;
1065 
1066 		if (!opriv ||
1067 		    opriv->sh != priv->sh ||
1068 			opriv->domain_id ==
1069 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1070 			continue;
1071 		priv->domain_id = opriv->domain_id;
1072 		break;
1073 	}
1074 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1075 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1076 		if (err) {
1077 			err = rte_errno;
1078 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1079 				strerror(rte_errno));
1080 			goto error;
1081 		}
1082 		own_domain_id = 1;
1083 	}
1084 	/* Override some values set by hardware configuration. */
1085 	mlx5_args(config, dpdk_dev->devargs);
1086 	err = mlx5_dev_check_sibling_config(priv, config);
1087 	if (err)
1088 		goto error;
1089 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1090 			    IBV_DEVICE_RAW_IP_CSUM);
1091 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1092 		(config->hw_csum ? "" : "not "));
1093 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1094 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1095 	DRV_LOG(DEBUG, "counters are not supported");
1096 #endif
1097 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1098 	if (config->dv_flow_en) {
1099 		DRV_LOG(WARNING, "DV flow is not supported");
1100 		config->dv_flow_en = 0;
1101 	}
1102 #endif
1103 	config->ind_table_max_size =
1104 		sh->device_attr.max_rwq_indirection_table_size;
1105 	/*
1106 	 * Remove this check once DPDK supports larger/variable
1107 	 * indirection tables.
1108 	 */
1109 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1110 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1111 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1112 		config->ind_table_max_size);
1113 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1114 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1115 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1116 		(config->hw_vlan_strip ? "" : "not "));
1117 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1118 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1119 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1120 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1121 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1122 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1123 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1124 #endif
1125 	if (config->hw_padding && !hw_padding) {
1126 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1127 		config->hw_padding = 0;
1128 	} else if (config->hw_padding) {
1129 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1130 	}
1131 	config->tso = (sh->device_attr.max_tso > 0 &&
1132 		      (sh->device_attr.tso_supported_qpts &
1133 		       (1 << IBV_QPT_RAW_PACKET)));
1134 	if (config->tso)
1135 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1136 	/*
1137 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1138 	 * by default.
1139 	 */
1140 	if (config->mps == MLX5_ARG_UNSET)
1141 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1142 							  MLX5_MPW_DISABLED;
1143 	else
1144 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1145 	DRV_LOG(INFO, "%sMPS is %s",
1146 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1147 		config->mps == MLX5_MPW ? "legacy " : "",
1148 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1149 	if (config->devx) {
1150 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1151 		if (err) {
1152 			err = -err;
1153 			goto error;
1154 		}
1155 		/* Check relax ordering support. */
1156 		if (!haswell_broadwell_cpu) {
1157 			sh->cmng.relaxed_ordering_write =
1158 				config->hca_attr.relaxed_ordering_write;
1159 			sh->cmng.relaxed_ordering_read =
1160 				config->hca_attr.relaxed_ordering_read;
1161 		} else {
1162 			sh->cmng.relaxed_ordering_read = 0;
1163 			sh->cmng.relaxed_ordering_write = 0;
1164 		}
1165 		/* Check for LRO support. */
1166 		if (config->dest_tir && config->hca_attr.lro_cap &&
1167 		    config->dv_flow_en) {
1168 			/* TBD check tunnel lro caps. */
1169 			config->lro.supported = config->hca_attr.lro_cap;
1170 			DRV_LOG(DEBUG, "Device supports LRO");
1171 			/*
1172 			 * If LRO timeout is not configured by application,
1173 			 * use the minimal supported value.
1174 			 */
1175 			if (!config->lro.timeout)
1176 				config->lro.timeout =
1177 				config->hca_attr.lro_timer_supported_periods[0];
1178 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1179 				config->lro.timeout);
1180 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1181 				"required for coalescing is %d bytes",
1182 				config->hca_attr.lro_min_mss_size);
1183 		}
1184 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1185 		if (config->hca_attr.qos.sup &&
1186 		    config->hca_attr.qos.flow_meter_old &&
1187 		    config->dv_flow_en) {
1188 			uint8_t reg_c_mask =
1189 				config->hca_attr.qos.flow_meter_reg_c_ids;
1190 			/*
1191 			 * Meter needs two REG_C's for color match and pre-sfx
1192 			 * flow match. Here get the REG_C for color match.
1193 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1194 			 */
1195 			reg_c_mask &= 0xfc;
1196 			if (__builtin_popcount(reg_c_mask) < 1) {
1197 				priv->mtr_en = 0;
1198 				DRV_LOG(WARNING, "No available register for"
1199 					" meter.");
1200 			} else {
1201 				/*
1202 				 * The meter color register is used by the
1203 				 * flow-hit feature as well.
1204 				 * The flow-hit feature must use REG_C_3
1205 				 * Prefer REG_C_3 if it is available.
1206 				 */
1207 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1208 					priv->mtr_color_reg = REG_C_3;
1209 				else
1210 					priv->mtr_color_reg = ffs(reg_c_mask)
1211 							      - 1 + REG_C_0;
1212 				priv->mtr_en = 1;
1213 				priv->mtr_reg_share =
1214 				      config->hca_attr.qos.flow_meter;
1215 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1216 					priv->mtr_color_reg);
1217 			}
1218 		}
1219 #endif
1220 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1221 		if (config->hca_attr.flow_hit_aso &&
1222 		    priv->mtr_color_reg == REG_C_3) {
1223 			sh->flow_hit_aso_en = 1;
1224 			err = mlx5_flow_aso_age_mng_init(sh);
1225 			if (err) {
1226 				err = -err;
1227 				goto error;
1228 			}
1229 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1230 		}
1231 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1232 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1233 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1234 		    config->dv_flow_en) {
1235 			priv->sampler_en = 1;
1236 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1237 		} else {
1238 			priv->sampler_en = 0;
1239 			if (!config->hca_attr.log_max_ft_sampler_num)
1240 				DRV_LOG(WARNING, "No available register for"
1241 						" Sampler.");
1242 			else
1243 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1244 		}
1245 #endif
1246 	}
1247 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1248 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1249 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1250 		config->cqe_comp = 0;
1251 	}
1252 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1253 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1254 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1255 				 " format isn't supported.");
1256 		config->cqe_comp = 0;
1257 	}
1258 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1259 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1260 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1261 				 " format isn't supported.");
1262 		config->cqe_comp = 0;
1263 	}
1264 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1265 			config->cqe_comp ? "" : "not ");
1266 	if (config->tx_pp) {
1267 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1268 			config->hca_attr.dev_freq_khz);
1269 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1270 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1271 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1272 			config->hca_attr.cross_channel ? "" : "not ");
1273 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1274 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1275 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1276 			config->hca_attr.non_wire_sq ? "" : "not ");
1277 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1278 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1279 			config->hca_attr.log_max_static_sq_wq);
1280 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1281 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1282 		if (!config->devx) {
1283 			DRV_LOG(ERR, "DevX is required for packet pacing");
1284 			err = ENODEV;
1285 			goto error;
1286 		}
1287 		if (!config->hca_attr.qos.packet_pacing) {
1288 			DRV_LOG(ERR, "Packet pacing is not supported");
1289 			err = ENODEV;
1290 			goto error;
1291 		}
1292 		if (!config->hca_attr.cross_channel) {
1293 			DRV_LOG(ERR, "Cross channel operations are"
1294 				     " required for packet pacing");
1295 			err = ENODEV;
1296 			goto error;
1297 		}
1298 		if (!config->hca_attr.wqe_index_ignore) {
1299 			DRV_LOG(ERR, "WQE index ignore feature is"
1300 				     " required for packet pacing");
1301 			err = ENODEV;
1302 			goto error;
1303 		}
1304 		if (!config->hca_attr.non_wire_sq) {
1305 			DRV_LOG(ERR, "Non-wire SQ feature is"
1306 				     " required for packet pacing");
1307 			err = ENODEV;
1308 			goto error;
1309 		}
1310 		if (!config->hca_attr.log_max_static_sq_wq) {
1311 			DRV_LOG(ERR, "Static WQE SQ feature is"
1312 				     " required for packet pacing");
1313 			err = ENODEV;
1314 			goto error;
1315 		}
1316 		if (!config->hca_attr.qos.wqe_rate_pp) {
1317 			DRV_LOG(ERR, "WQE rate mode is required"
1318 				     " for packet pacing");
1319 			err = ENODEV;
1320 			goto error;
1321 		}
1322 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1323 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1324 			     " can't create queues for packet pacing");
1325 		err = ENODEV;
1326 		goto error;
1327 #endif
1328 	}
1329 	if (config->devx) {
1330 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1331 
1332 		err = config->hca_attr.access_register_user ?
1333 			mlx5_devx_cmd_register_read
1334 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1335 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1336 		if (!err) {
1337 			uint32_t ts_mode;
1338 
1339 			/* MTUTC register is read successfully. */
1340 			ts_mode = MLX5_GET(register_mtutc, reg,
1341 					   time_stamp_mode);
1342 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1343 				config->rt_timestamp = 1;
1344 		} else {
1345 			/* Kernel does not support register reading. */
1346 			if (config->hca_attr.dev_freq_khz ==
1347 						 (NS_PER_S / MS_PER_S))
1348 				config->rt_timestamp = 1;
1349 		}
1350 	}
1351 	/*
1352 	 * If HW has bug working with tunnel packet decapsulation and
1353 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1354 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1355 	 */
1356 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1357 		config->hw_fcs_strip = 0;
1358 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1359 		(config->hw_fcs_strip ? "" : "not "));
1360 	if (config->mprq.enabled && mprq) {
1361 		if (config->mprq.stride_num_n &&
1362 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1363 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1364 			config->mprq.stride_num_n =
1365 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1366 						mprq_min_stride_num_n),
1367 					mprq_max_stride_num_n);
1368 			DRV_LOG(WARNING,
1369 				"the number of strides"
1370 				" for Multi-Packet RQ is out of range,"
1371 				" setting default value (%u)",
1372 				1 << config->mprq.stride_num_n);
1373 		}
1374 		if (config->mprq.stride_size_n &&
1375 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1376 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1377 			config->mprq.stride_size_n =
1378 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1379 						mprq_min_stride_size_n),
1380 					mprq_max_stride_size_n);
1381 			DRV_LOG(WARNING,
1382 				"the size of a stride"
1383 				" for Multi-Packet RQ is out of range,"
1384 				" setting default value (%u)",
1385 				1 << config->mprq.stride_size_n);
1386 		}
1387 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1388 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1389 	} else if (config->mprq.enabled && !mprq) {
1390 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1391 		config->mprq.enabled = 0;
1392 	}
1393 	if (config->max_dump_files_num == 0)
1394 		config->max_dump_files_num = 128;
1395 	eth_dev = rte_eth_dev_allocate(name);
1396 	if (eth_dev == NULL) {
1397 		DRV_LOG(ERR, "can not allocate rte ethdev");
1398 		err = ENOMEM;
1399 		goto error;
1400 	}
1401 	if (priv->representor) {
1402 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1403 		eth_dev->data->representor_id = priv->representor_id;
1404 	}
1405 	priv->mp_id.port_id = eth_dev->data->port_id;
1406 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1407 	/*
1408 	 * Store associated network device interface index. This index
1409 	 * is permanent throughout the lifetime of device. So, we may store
1410 	 * the ifindex here and use the cached value further.
1411 	 */
1412 	MLX5_ASSERT(spawn->ifindex);
1413 	priv->if_index = spawn->ifindex;
1414 	if (priv->pf_bond >= 0 && priv->master) {
1415 		/* Get bond interface info */
1416 		err = mlx5_sysfs_bond_info(priv->if_index,
1417 				     &priv->bond_ifindex,
1418 				     priv->bond_name);
1419 		if (err)
1420 			DRV_LOG(ERR, "unable to get bond info: %s",
1421 				strerror(rte_errno));
1422 		else
1423 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1424 				priv->if_index, priv->bond_ifindex,
1425 				priv->bond_name);
1426 	}
1427 	eth_dev->data->dev_private = priv;
1428 	priv->dev_data = eth_dev->data;
1429 	eth_dev->data->mac_addrs = priv->mac;
1430 	if (spawn->pf_bond < 0) {
1431 		eth_dev->device = dpdk_dev;
1432 	} else {
1433 		/* Use primary bond PCI as device. */
1434 		if (sh->bond_dev == UINT16_MAX) {
1435 			sh->bond_dev = eth_dev->data->port_id;
1436 			eth_dev->device = dpdk_dev;
1437 		} else {
1438 			eth_dev->device = rte_eth_devices[sh->bond_dev].device;
1439 		}
1440 	}
1441 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1442 	/* Configure the first MAC address by default. */
1443 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1444 		DRV_LOG(ERR,
1445 			"port %u cannot get MAC address, is mlx5_en"
1446 			" loaded? (errno: %s)",
1447 			eth_dev->data->port_id, strerror(rte_errno));
1448 		err = ENODEV;
1449 		goto error;
1450 	}
1451 	DRV_LOG(INFO,
1452 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1453 		eth_dev->data->port_id,
1454 		mac.addr_bytes[0], mac.addr_bytes[1],
1455 		mac.addr_bytes[2], mac.addr_bytes[3],
1456 		mac.addr_bytes[4], mac.addr_bytes[5]);
1457 #ifdef RTE_LIBRTE_MLX5_DEBUG
1458 	{
1459 		char ifname[MLX5_NAMESIZE];
1460 
1461 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1462 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1463 				eth_dev->data->port_id, ifname);
1464 		else
1465 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1466 				eth_dev->data->port_id);
1467 	}
1468 #endif
1469 	/* Get actual MTU if possible. */
1470 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1471 	if (err) {
1472 		err = rte_errno;
1473 		goto error;
1474 	}
1475 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1476 		priv->mtu);
1477 	/* Initialize burst functions to prevent crashes before link-up. */
1478 	eth_dev->rx_pkt_burst = removed_rx_burst;
1479 	eth_dev->tx_pkt_burst = removed_tx_burst;
1480 	eth_dev->dev_ops = &mlx5_dev_ops;
1481 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1482 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1483 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1484 	/* Register MAC address. */
1485 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1486 	if (config->vf && config->vf_nl_en)
1487 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1488 				      mlx5_ifindex(eth_dev),
1489 				      eth_dev->data->mac_addrs,
1490 				      MLX5_MAX_MAC_ADDRESSES);
1491 	priv->flows = 0;
1492 	priv->ctrl_flows = 0;
1493 	rte_spinlock_init(&priv->flow_list_lock);
1494 	TAILQ_INIT(&priv->flow_meters);
1495 	TAILQ_INIT(&priv->flow_meter_profiles);
1496 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1497 	mlx5_glue->dv_set_context_attr(sh->ctx,
1498 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1499 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1500 				.alloc = &mlx5_alloc_verbs_buf,
1501 				.free = &mlx5_free_verbs_buf,
1502 				.data = sh,
1503 			}));
1504 	/* Bring Ethernet device up. */
1505 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1506 		eth_dev->data->port_id);
1507 	mlx5_set_link_up(eth_dev);
1508 	/*
1509 	 * Even though the interrupt handler is not installed yet,
1510 	 * interrupts will still trigger on the async_fd from
1511 	 * Verbs context returned by ibv_open_device().
1512 	 */
1513 	mlx5_link_update(eth_dev, 0);
1514 #ifdef HAVE_MLX5DV_DR_ESWITCH
1515 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1516 	      (switch_info->representor || switch_info->master)))
1517 		config->dv_esw_en = 0;
1518 #else
1519 	config->dv_esw_en = 0;
1520 #endif
1521 	/* Detect minimal data bytes to inline. */
1522 	mlx5_set_min_inline(spawn, config);
1523 	/* Store device configuration on private structure. */
1524 	priv->config = *config;
1525 	/* Create context for virtual machine VLAN workaround. */
1526 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1527 	if (config->dv_flow_en) {
1528 		err = mlx5_alloc_shared_dr(priv);
1529 		if (err)
1530 			goto error;
1531 	}
1532 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1533 		priv->obj_ops = devx_obj_ops;
1534 		priv->obj_ops.drop_action_create =
1535 						ibv_obj_ops.drop_action_create;
1536 		priv->obj_ops.drop_action_destroy =
1537 						ibv_obj_ops.drop_action_destroy;
1538 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1539 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1540 #else
1541 		if (config->dv_esw_en)
1542 			priv->obj_ops.txq_obj_modify =
1543 						ibv_obj_ops.txq_obj_modify;
1544 #endif
1545 		/* Use specific wrappers for Tx object. */
1546 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1547 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1548 		mlx5_queue_counter_id_prepare(eth_dev);
1549 
1550 	} else {
1551 		priv->obj_ops = ibv_obj_ops;
1552 	}
1553 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1554 	if (!priv->drop_queue.hrxq)
1555 		goto error;
1556 	/* Supported Verbs flow priority number detection. */
1557 	err = mlx5_flow_discover_priorities(eth_dev);
1558 	if (err < 0) {
1559 		err = -err;
1560 		goto error;
1561 	}
1562 	priv->config.flow_prio = err;
1563 	if (!priv->config.dv_esw_en &&
1564 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1565 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1566 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1567 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1568 	}
1569 	mlx5_set_metadata_mask(eth_dev);
1570 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1571 	    !priv->sh->dv_regc0_mask) {
1572 		DRV_LOG(ERR, "metadata mode %u is not supported "
1573 			     "(no metadata reg_c[0] is available)",
1574 			     priv->config.dv_xmeta_en);
1575 			err = ENOTSUP;
1576 			goto error;
1577 	}
1578 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1579 			     mlx5_hrxq_create_cb,
1580 			     mlx5_hrxq_match_cb,
1581 			     mlx5_hrxq_remove_cb);
1582 	/* Query availability of metadata reg_c's. */
1583 	err = mlx5_flow_discover_mreg_c(eth_dev);
1584 	if (err < 0) {
1585 		err = -err;
1586 		goto error;
1587 	}
1588 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1589 		DRV_LOG(DEBUG,
1590 			"port %u extensive metadata register is not supported",
1591 			eth_dev->data->port_id);
1592 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1593 			DRV_LOG(ERR, "metadata mode %u is not supported "
1594 				     "(no metadata registers available)",
1595 				     priv->config.dv_xmeta_en);
1596 			err = ENOTSUP;
1597 			goto error;
1598 		}
1599 	}
1600 	if (priv->config.dv_flow_en &&
1601 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1602 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1603 	    priv->sh->dv_regc0_mask) {
1604 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1605 						      MLX5_FLOW_MREG_HTABLE_SZ,
1606 						      0, 0,
1607 						      flow_dv_mreg_create_cb,
1608 						      flow_dv_mreg_match_cb,
1609 						      flow_dv_mreg_remove_cb);
1610 		if (!priv->mreg_cp_tbl) {
1611 			err = ENOMEM;
1612 			goto error;
1613 		}
1614 		priv->mreg_cp_tbl->ctx = eth_dev;
1615 	}
1616 	rte_spinlock_init(&priv->shared_act_sl);
1617 	mlx5_flow_counter_mode_config(eth_dev);
1618 	if (priv->config.dv_flow_en)
1619 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1620 	return eth_dev;
1621 error:
1622 	if (priv) {
1623 		if (priv->mreg_cp_tbl)
1624 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1625 		if (priv->sh)
1626 			mlx5_os_free_shared_dr(priv);
1627 		if (priv->nl_socket_route >= 0)
1628 			close(priv->nl_socket_route);
1629 		if (priv->nl_socket_rdma >= 0)
1630 			close(priv->nl_socket_rdma);
1631 		if (priv->vmwa_context)
1632 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1633 		if (eth_dev && priv->drop_queue.hrxq)
1634 			mlx5_drop_action_destroy(eth_dev);
1635 		if (own_domain_id)
1636 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1637 		mlx5_cache_list_destroy(&priv->hrxqs);
1638 		mlx5_free(priv);
1639 		if (eth_dev != NULL)
1640 			eth_dev->data->dev_private = NULL;
1641 	}
1642 	if (eth_dev != NULL) {
1643 		/* mac_addrs must not be freed alone because part of
1644 		 * dev_private
1645 		 **/
1646 		eth_dev->data->mac_addrs = NULL;
1647 		rte_eth_dev_release_port(eth_dev);
1648 	}
1649 	if (sh)
1650 		mlx5_free_shared_dev_ctx(sh);
1651 	MLX5_ASSERT(err > 0);
1652 	rte_errno = err;
1653 	return NULL;
1654 }
1655 
1656 /**
1657  * Comparison callback to sort device data.
1658  *
1659  * This is meant to be used with qsort().
1660  *
1661  * @param a[in]
1662  *   Pointer to pointer to first data object.
1663  * @param b[in]
1664  *   Pointer to pointer to second data object.
1665  *
1666  * @return
1667  *   0 if both objects are equal, less than 0 if the first argument is less
1668  *   than the second, greater than 0 otherwise.
1669  */
1670 static int
1671 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1672 {
1673 	const struct mlx5_switch_info *si_a =
1674 		&((const struct mlx5_dev_spawn_data *)a)->info;
1675 	const struct mlx5_switch_info *si_b =
1676 		&((const struct mlx5_dev_spawn_data *)b)->info;
1677 	int ret;
1678 
1679 	/* Master device first. */
1680 	ret = si_b->master - si_a->master;
1681 	if (ret)
1682 		return ret;
1683 	/* Then representor devices. */
1684 	ret = si_b->representor - si_a->representor;
1685 	if (ret)
1686 		return ret;
1687 	/* Unidentified devices come last in no specific order. */
1688 	if (!si_a->representor)
1689 		return 0;
1690 	/* Order representors by name. */
1691 	return si_a->port_name - si_b->port_name;
1692 }
1693 
1694 /**
1695  * Match PCI information for possible slaves of bonding device.
1696  *
1697  * @param[in] ibv_dev
1698  *   Pointer to Infiniband device structure.
1699  * @param[in] pci_dev
1700  *   Pointer to PCI device structure to match PCI address.
1701  * @param[in] nl_rdma
1702  *   Netlink RDMA group socket handle.
1703  *
1704  * @return
1705  *   negative value if no bonding device found, otherwise
1706  *   positive index of slave PF in bonding.
1707  */
1708 static int
1709 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1710 			   const struct rte_pci_device *pci_dev,
1711 			   int nl_rdma)
1712 {
1713 	char ifname[IF_NAMESIZE + 1];
1714 	unsigned int ifindex;
1715 	unsigned int np, i;
1716 	FILE *file = NULL;
1717 	int pf = -1;
1718 
1719 	/*
1720 	 * Try to get master device name. If something goes
1721 	 * wrong suppose the lack of kernel support and no
1722 	 * bonding devices.
1723 	 */
1724 	if (nl_rdma < 0)
1725 		return -1;
1726 	if (!strstr(ibv_dev->name, "bond"))
1727 		return -1;
1728 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1729 	if (!np)
1730 		return -1;
1731 	/*
1732 	 * The Master device might not be on the predefined
1733 	 * port (not on port index 1, it is not garanted),
1734 	 * we have to scan all Infiniband device port and
1735 	 * find master.
1736 	 */
1737 	for (i = 1; i <= np; ++i) {
1738 		/* Check whether Infiniband port is populated. */
1739 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1740 		if (!ifindex)
1741 			continue;
1742 		if (!if_indextoname(ifindex, ifname))
1743 			continue;
1744 		/* Try to read bonding slave names from sysfs. */
1745 		MKSTR(slaves,
1746 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1747 		file = fopen(slaves, "r");
1748 		if (file)
1749 			break;
1750 	}
1751 	if (!file)
1752 		return -1;
1753 	/* Use safe format to check maximal buffer length. */
1754 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1755 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1756 		char tmp_str[IF_NAMESIZE + 32];
1757 		struct rte_pci_addr pci_addr;
1758 		struct mlx5_switch_info	info;
1759 
1760 		/* Process slave interface names in the loop. */
1761 		snprintf(tmp_str, sizeof(tmp_str),
1762 			 "/sys/class/net/%s", ifname);
1763 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1764 			DRV_LOG(WARNING, "can not get PCI address"
1765 					 " for netdev \"%s\"", ifname);
1766 			continue;
1767 		}
1768 		if (pci_dev->addr.domain != pci_addr.domain ||
1769 		    pci_dev->addr.bus != pci_addr.bus ||
1770 		    pci_dev->addr.devid != pci_addr.devid ||
1771 		    pci_dev->addr.function != pci_addr.function)
1772 			continue;
1773 		/* Slave interface PCI address match found. */
1774 		fclose(file);
1775 		snprintf(tmp_str, sizeof(tmp_str),
1776 			 "/sys/class/net/%s/phys_port_name", ifname);
1777 		file = fopen(tmp_str, "rb");
1778 		if (!file)
1779 			break;
1780 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1781 		if (fscanf(file, "%32s", tmp_str) == 1)
1782 			mlx5_translate_port_name(tmp_str, &info);
1783 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1784 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1785 			pf = info.port_name;
1786 		break;
1787 	}
1788 	if (file)
1789 		fclose(file);
1790 	return pf;
1791 }
1792 
1793 /**
1794  * DPDK callback to register a PCI device.
1795  *
1796  * This function spawns Ethernet devices out of a given PCI device.
1797  *
1798  * @param[in] pci_drv
1799  *   PCI driver structure (mlx5_driver).
1800  * @param[in] pci_dev
1801  *   PCI device information.
1802  *
1803  * @return
1804  *   0 on success, a negative errno value otherwise and rte_errno is set.
1805  */
1806 int
1807 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1808 		  struct rte_pci_device *pci_dev)
1809 {
1810 	struct ibv_device **ibv_list;
1811 	/*
1812 	 * Number of found IB Devices matching with requested PCI BDF.
1813 	 * nd != 1 means there are multiple IB devices over the same
1814 	 * PCI device and we have representors and master.
1815 	 */
1816 	unsigned int nd = 0;
1817 	/*
1818 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1819 	 * we have the single multiport IB device, and there may be
1820 	 * representors attached to some of found ports.
1821 	 */
1822 	unsigned int np = 0;
1823 	/*
1824 	 * Number of DPDK ethernet devices to Spawn - either over
1825 	 * multiple IB devices or multiple ports of single IB device.
1826 	 * Actually this is the number of iterations to spawn.
1827 	 */
1828 	unsigned int ns = 0;
1829 	/*
1830 	 * Bonding device
1831 	 *   < 0 - no bonding device (single one)
1832 	 *  >= 0 - bonding device (value is slave PF index)
1833 	 */
1834 	int bd = -1;
1835 	struct mlx5_dev_spawn_data *list = NULL;
1836 	struct mlx5_dev_config dev_config;
1837 	unsigned int dev_config_vf;
1838 	int ret;
1839 
1840 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1841 		mlx5_pmd_socket_init();
1842 	ret = mlx5_init_once();
1843 	if (ret) {
1844 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1845 			strerror(rte_errno));
1846 		return -rte_errno;
1847 	}
1848 	errno = 0;
1849 	ibv_list = mlx5_glue->get_device_list(&ret);
1850 	if (!ibv_list) {
1851 		rte_errno = errno ? errno : ENOSYS;
1852 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1853 		return -rte_errno;
1854 	}
1855 	/*
1856 	 * First scan the list of all Infiniband devices to find
1857 	 * matching ones, gathering into the list.
1858 	 */
1859 	struct ibv_device *ibv_match[ret + 1];
1860 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1861 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1862 	unsigned int i;
1863 
1864 	while (ret-- > 0) {
1865 		struct rte_pci_addr pci_addr;
1866 
1867 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1868 		bd = mlx5_device_bond_pci_match
1869 				(ibv_list[ret], pci_dev, nl_rdma);
1870 		if (bd >= 0) {
1871 			/*
1872 			 * Bonding device detected. Only one match is allowed,
1873 			 * the bonding is supported over multi-port IB device,
1874 			 * there should be no matches on representor PCI
1875 			 * functions or non VF LAG bonding devices with
1876 			 * specified address.
1877 			 */
1878 			if (nd) {
1879 				DRV_LOG(ERR,
1880 					"multiple PCI match on bonding device"
1881 					"\"%s\" found", ibv_list[ret]->name);
1882 				rte_errno = ENOENT;
1883 				ret = -rte_errno;
1884 				goto exit;
1885 			}
1886 			DRV_LOG(INFO, "PCI information matches for"
1887 				      " slave %d bonding device \"%s\"",
1888 				      bd, ibv_list[ret]->name);
1889 			ibv_match[nd++] = ibv_list[ret];
1890 			break;
1891 		}
1892 		if (mlx5_dev_to_pci_addr
1893 			(ibv_list[ret]->ibdev_path, &pci_addr))
1894 			continue;
1895 		if (pci_dev->addr.domain != pci_addr.domain ||
1896 		    pci_dev->addr.bus != pci_addr.bus ||
1897 		    pci_dev->addr.devid != pci_addr.devid ||
1898 		    pci_dev->addr.function != pci_addr.function)
1899 			continue;
1900 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1901 			ibv_list[ret]->name);
1902 		ibv_match[nd++] = ibv_list[ret];
1903 	}
1904 	ibv_match[nd] = NULL;
1905 	if (!nd) {
1906 		/* No device matches, just complain and bail out. */
1907 		DRV_LOG(WARNING,
1908 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1909 			" are kernel drivers loaded?",
1910 			pci_dev->addr.domain, pci_dev->addr.bus,
1911 			pci_dev->addr.devid, pci_dev->addr.function);
1912 		rte_errno = ENOENT;
1913 		ret = -rte_errno;
1914 		goto exit;
1915 	}
1916 	if (nd == 1) {
1917 		/*
1918 		 * Found single matching device may have multiple ports.
1919 		 * Each port may be representor, we have to check the port
1920 		 * number and check the representors existence.
1921 		 */
1922 		if (nl_rdma >= 0)
1923 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1924 		if (!np)
1925 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1926 					 " ports number", ibv_match[0]->name);
1927 		if (bd >= 0 && !np) {
1928 			DRV_LOG(ERR, "can not get ports"
1929 				     " for bonding device");
1930 			rte_errno = ENOENT;
1931 			ret = -rte_errno;
1932 			goto exit;
1933 		}
1934 	}
1935 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1936 	if (bd >= 0) {
1937 		/*
1938 		 * This may happen if there is VF LAG kernel support and
1939 		 * application is compiled with older rdma_core library.
1940 		 */
1941 		DRV_LOG(ERR,
1942 			"No kernel/verbs support for VF LAG bonding found.");
1943 		rte_errno = ENOTSUP;
1944 		ret = -rte_errno;
1945 		goto exit;
1946 	}
1947 #endif
1948 	/*
1949 	 * Now we can determine the maximal
1950 	 * amount of devices to be spawned.
1951 	 */
1952 	list = mlx5_malloc(MLX5_MEM_ZERO,
1953 			   sizeof(struct mlx5_dev_spawn_data) *
1954 			   (np ? np : nd),
1955 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1956 	if (!list) {
1957 		DRV_LOG(ERR, "spawn data array allocation failure");
1958 		rte_errno = ENOMEM;
1959 		ret = -rte_errno;
1960 		goto exit;
1961 	}
1962 	if (bd >= 0 || np > 1) {
1963 		/*
1964 		 * Single IB device with multiple ports found,
1965 		 * it may be E-Switch master device and representors.
1966 		 * We have to perform identification through the ports.
1967 		 */
1968 		MLX5_ASSERT(nl_rdma >= 0);
1969 		MLX5_ASSERT(ns == 0);
1970 		MLX5_ASSERT(nd == 1);
1971 		MLX5_ASSERT(np);
1972 		for (i = 1; i <= np; ++i) {
1973 			list[ns].max_port = np;
1974 			list[ns].phys_port = i;
1975 			list[ns].phys_dev = ibv_match[0];
1976 			list[ns].eth_dev = NULL;
1977 			list[ns].pci_dev = pci_dev;
1978 			list[ns].pf_bond = bd;
1979 			list[ns].ifindex = mlx5_nl_ifindex
1980 				(nl_rdma,
1981 				mlx5_os_get_dev_device_name
1982 						(list[ns].phys_dev), i);
1983 			if (!list[ns].ifindex) {
1984 				/*
1985 				 * No network interface index found for the
1986 				 * specified port, it means there is no
1987 				 * representor on this port. It's OK,
1988 				 * there can be disabled ports, for example
1989 				 * if sriov_numvfs < sriov_totalvfs.
1990 				 */
1991 				continue;
1992 			}
1993 			ret = -1;
1994 			if (nl_route >= 0)
1995 				ret = mlx5_nl_switch_info
1996 					       (nl_route,
1997 						list[ns].ifindex,
1998 						&list[ns].info);
1999 			if (ret || (!list[ns].info.representor &&
2000 				    !list[ns].info.master)) {
2001 				/*
2002 				 * We failed to recognize representors with
2003 				 * Netlink, let's try to perform the task
2004 				 * with sysfs.
2005 				 */
2006 				ret =  mlx5_sysfs_switch_info
2007 						(list[ns].ifindex,
2008 						 &list[ns].info);
2009 			}
2010 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2011 			if (!ret && bd >= 0) {
2012 				switch (list[ns].info.name_type) {
2013 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2014 					if (list[ns].info.port_name == bd)
2015 						ns++;
2016 					break;
2017 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2018 					/* Fallthrough */
2019 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2020 					if (list[ns].info.pf_num == bd)
2021 						ns++;
2022 					break;
2023 				default:
2024 					break;
2025 				}
2026 				continue;
2027 			}
2028 #endif
2029 			if (!ret && (list[ns].info.representor ^
2030 				     list[ns].info.master))
2031 				ns++;
2032 		}
2033 		if (!ns) {
2034 			DRV_LOG(ERR,
2035 				"unable to recognize master/representors"
2036 				" on the IB device with multiple ports");
2037 			rte_errno = ENOENT;
2038 			ret = -rte_errno;
2039 			goto exit;
2040 		}
2041 	} else {
2042 		/*
2043 		 * The existence of several matching entries (nd > 1) means
2044 		 * port representors have been instantiated. No existing Verbs
2045 		 * call nor sysfs entries can tell them apart, this can only
2046 		 * be done through Netlink calls assuming kernel drivers are
2047 		 * recent enough to support them.
2048 		 *
2049 		 * In the event of identification failure through Netlink,
2050 		 * try again through sysfs, then:
2051 		 *
2052 		 * 1. A single IB device matches (nd == 1) with single
2053 		 *    port (np=0/1) and is not a representor, assume
2054 		 *    no switch support.
2055 		 *
2056 		 * 2. Otherwise no safe assumptions can be made;
2057 		 *    complain louder and bail out.
2058 		 */
2059 		for (i = 0; i != nd; ++i) {
2060 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2061 			list[ns].max_port = 1;
2062 			list[ns].phys_port = 1;
2063 			list[ns].phys_dev = ibv_match[i];
2064 			list[ns].eth_dev = NULL;
2065 			list[ns].pci_dev = pci_dev;
2066 			list[ns].pf_bond = -1;
2067 			list[ns].ifindex = 0;
2068 			if (nl_rdma >= 0)
2069 				list[ns].ifindex = mlx5_nl_ifindex
2070 				(nl_rdma,
2071 				mlx5_os_get_dev_device_name
2072 						(list[ns].phys_dev), 1);
2073 			if (!list[ns].ifindex) {
2074 				char ifname[IF_NAMESIZE];
2075 
2076 				/*
2077 				 * Netlink failed, it may happen with old
2078 				 * ib_core kernel driver (before 4.16).
2079 				 * We can assume there is old driver because
2080 				 * here we are processing single ports IB
2081 				 * devices. Let's try sysfs to retrieve
2082 				 * the ifindex. The method works for
2083 				 * master device only.
2084 				 */
2085 				if (nd > 1) {
2086 					/*
2087 					 * Multiple devices found, assume
2088 					 * representors, can not distinguish
2089 					 * master/representor and retrieve
2090 					 * ifindex via sysfs.
2091 					 */
2092 					continue;
2093 				}
2094 				ret = mlx5_get_ifname_sysfs
2095 					(ibv_match[i]->ibdev_path, ifname);
2096 				if (!ret)
2097 					list[ns].ifindex =
2098 						if_nametoindex(ifname);
2099 				if (!list[ns].ifindex) {
2100 					/*
2101 					 * No network interface index found
2102 					 * for the specified device, it means
2103 					 * there it is neither representor
2104 					 * nor master.
2105 					 */
2106 					continue;
2107 				}
2108 			}
2109 			ret = -1;
2110 			if (nl_route >= 0)
2111 				ret = mlx5_nl_switch_info
2112 					       (nl_route,
2113 						list[ns].ifindex,
2114 						&list[ns].info);
2115 			if (ret || (!list[ns].info.representor &&
2116 				    !list[ns].info.master)) {
2117 				/*
2118 				 * We failed to recognize representors with
2119 				 * Netlink, let's try to perform the task
2120 				 * with sysfs.
2121 				 */
2122 				ret =  mlx5_sysfs_switch_info
2123 						(list[ns].ifindex,
2124 						 &list[ns].info);
2125 			}
2126 			if (!ret && (list[ns].info.representor ^
2127 				     list[ns].info.master)) {
2128 				ns++;
2129 			} else if ((nd == 1) &&
2130 				   !list[ns].info.representor &&
2131 				   !list[ns].info.master) {
2132 				/*
2133 				 * Single IB device with
2134 				 * one physical port and
2135 				 * attached network device.
2136 				 * May be SRIOV is not enabled
2137 				 * or there is no representors.
2138 				 */
2139 				DRV_LOG(INFO, "no E-Switch support detected");
2140 				ns++;
2141 				break;
2142 			}
2143 		}
2144 		if (!ns) {
2145 			DRV_LOG(ERR,
2146 				"unable to recognize master/representors"
2147 				" on the multiple IB devices");
2148 			rte_errno = ENOENT;
2149 			ret = -rte_errno;
2150 			goto exit;
2151 		}
2152 	}
2153 	MLX5_ASSERT(ns);
2154 	/*
2155 	 * Sort list to probe devices in natural order for users convenience
2156 	 * (i.e. master first, then representors from lowest to highest ID).
2157 	 */
2158 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2159 	/* Device specific configuration. */
2160 	switch (pci_dev->id.device_id) {
2161 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2162 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2163 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2164 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2165 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2166 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2167 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2168 		dev_config_vf = 1;
2169 		break;
2170 	default:
2171 		dev_config_vf = 0;
2172 		break;
2173 	}
2174 	for (i = 0; i != ns; ++i) {
2175 		uint32_t restore;
2176 
2177 		/* Default configuration. */
2178 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2179 		dev_config.vf = dev_config_vf;
2180 		dev_config.mps = MLX5_ARG_UNSET;
2181 		dev_config.dbnc = MLX5_ARG_UNSET;
2182 		dev_config.rx_vec_en = 1;
2183 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2184 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2185 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2186 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2187 		dev_config.vf_nl_en = 1;
2188 		dev_config.mr_ext_memseg_en = 1;
2189 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2190 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2191 		dev_config.dv_esw_en = 1;
2192 		dev_config.dv_flow_en = 1;
2193 		dev_config.decap_en = 1;
2194 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2195 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2196 						 &list[i],
2197 						 &dev_config);
2198 		if (!list[i].eth_dev) {
2199 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2200 				break;
2201 			/* Device is disabled or already spawned. Ignore it. */
2202 			continue;
2203 		}
2204 		restore = list[i].eth_dev->data->dev_flags;
2205 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2206 		/* Restore non-PCI flags cleared by the above call. */
2207 		list[i].eth_dev->data->dev_flags |= restore;
2208 		rte_eth_dev_probing_finish(list[i].eth_dev);
2209 	}
2210 	if (i != ns) {
2211 		DRV_LOG(ERR,
2212 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2213 			" encountering an error: %s",
2214 			pci_dev->addr.domain, pci_dev->addr.bus,
2215 			pci_dev->addr.devid, pci_dev->addr.function,
2216 			strerror(rte_errno));
2217 		ret = -rte_errno;
2218 		/* Roll back. */
2219 		while (i--) {
2220 			if (!list[i].eth_dev)
2221 				continue;
2222 			mlx5_dev_close(list[i].eth_dev);
2223 			/* mac_addrs must not be freed because in dev_private */
2224 			list[i].eth_dev->data->mac_addrs = NULL;
2225 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2226 		}
2227 		/* Restore original error. */
2228 		rte_errno = -ret;
2229 	} else {
2230 		ret = 0;
2231 	}
2232 exit:
2233 	/*
2234 	 * Do the routine cleanup:
2235 	 * - close opened Netlink sockets
2236 	 * - free allocated spawn data array
2237 	 * - free the Infiniband device list
2238 	 */
2239 	if (nl_rdma >= 0)
2240 		close(nl_rdma);
2241 	if (nl_route >= 0)
2242 		close(nl_route);
2243 	if (list)
2244 		mlx5_free(list);
2245 	MLX5_ASSERT(ibv_list);
2246 	mlx5_glue->free_device_list(ibv_list);
2247 	return ret;
2248 }
2249 
2250 static int
2251 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2252 {
2253 	char *env;
2254 	int value;
2255 
2256 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2257 	/* Get environment variable to store. */
2258 	env = getenv(MLX5_SHUT_UP_BF);
2259 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2260 	if (config->dbnc == MLX5_ARG_UNSET)
2261 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2262 	else
2263 		setenv(MLX5_SHUT_UP_BF,
2264 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2265 	return value;
2266 }
2267 
2268 static void
2269 mlx5_restore_doorbell_mapping_env(int value)
2270 {
2271 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2272 	/* Restore the original environment variable state. */
2273 	if (value == MLX5_ARG_UNSET)
2274 		unsetenv(MLX5_SHUT_UP_BF);
2275 	else
2276 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2277 }
2278 
2279 /**
2280  * Extract pdn of PD object using DV API.
2281  *
2282  * @param[in] pd
2283  *   Pointer to the verbs PD object.
2284  * @param[out] pdn
2285  *   Pointer to the PD object number variable.
2286  *
2287  * @return
2288  *   0 on success, error value otherwise.
2289  */
2290 int
2291 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2292 {
2293 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2294 	struct mlx5dv_obj obj;
2295 	struct mlx5dv_pd pd_info;
2296 	int ret = 0;
2297 
2298 	obj.pd.in = pd;
2299 	obj.pd.out = &pd_info;
2300 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2301 	if (ret) {
2302 		DRV_LOG(DEBUG, "Fail to get PD object info");
2303 		return ret;
2304 	}
2305 	*pdn = pd_info.pdn;
2306 	return 0;
2307 #else
2308 	(void)pd;
2309 	(void)pdn;
2310 	return -ENOTSUP;
2311 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2312 }
2313 
2314 /**
2315  * Function API to open IB device.
2316  *
2317  * This function calls the Linux glue APIs to open a device.
2318  *
2319  * @param[in] spawn
2320  *   Pointer to the IB device attributes (name, port, etc).
2321  * @param[out] config
2322  *   Pointer to device configuration structure.
2323  * @param[out] sh
2324  *   Pointer to shared context structure.
2325  *
2326  * @return
2327  *   0 on success, a positive error value otherwise.
2328  */
2329 int
2330 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2331 		     const struct mlx5_dev_config *config,
2332 		     struct mlx5_dev_ctx_shared *sh)
2333 {
2334 	int dbmap_env;
2335 	int err = 0;
2336 
2337 	sh->numa_node = spawn->pci_dev->device.numa_node;
2338 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2339 	/*
2340 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2341 	 * before the device creation. The rdma_core library
2342 	 * checks the variable at device creation and
2343 	 * stores the result internally.
2344 	 */
2345 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2346 	/* Try to open IB device with DV first, then usual Verbs. */
2347 	errno = 0;
2348 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2349 	if (sh->ctx) {
2350 		sh->devx = 1;
2351 		DRV_LOG(DEBUG, "DevX is supported");
2352 		/* The device is created, no need for environment. */
2353 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2354 	} else {
2355 		/* The environment variable is still configured. */
2356 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2357 		err = errno ? errno : ENODEV;
2358 		/*
2359 		 * The environment variable is not needed anymore,
2360 		 * all device creation attempts are completed.
2361 		 */
2362 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2363 		if (!sh->ctx)
2364 			return err;
2365 		DRV_LOG(DEBUG, "DevX is NOT supported");
2366 		err = 0;
2367 	}
2368 	if (!err && sh->ctx) {
2369 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2370 		mlx5_glue->dv_set_context_attr(sh->ctx,
2371 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2372 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2373 				.alloc = &mlx5_alloc_verbs_buf,
2374 				.free = &mlx5_free_verbs_buf,
2375 				.data = sh,
2376 			}));
2377 	}
2378 	return err;
2379 }
2380 
2381 /**
2382  * Install shared asynchronous device events handler.
2383  * This function is implemented to support event sharing
2384  * between multiple ports of single IB device.
2385  *
2386  * @param sh
2387  *   Pointer to mlx5_dev_ctx_shared object.
2388  */
2389 void
2390 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2391 {
2392 	int ret;
2393 	int flags;
2394 
2395 	sh->intr_handle.fd = -1;
2396 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2397 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2398 		    F_SETFL, flags | O_NONBLOCK);
2399 	if (ret) {
2400 		DRV_LOG(INFO, "failed to change file descriptor async event"
2401 			" queue");
2402 	} else {
2403 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2404 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2405 		if (rte_intr_callback_register(&sh->intr_handle,
2406 					mlx5_dev_interrupt_handler, sh)) {
2407 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2408 			sh->intr_handle.fd = -1;
2409 		}
2410 	}
2411 	if (sh->devx) {
2412 #ifdef HAVE_IBV_DEVX_ASYNC
2413 		sh->intr_handle_devx.fd = -1;
2414 		sh->devx_comp =
2415 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2416 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2417 		if (!devx_comp) {
2418 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2419 			return;
2420 		}
2421 		flags = fcntl(devx_comp->fd, F_GETFL);
2422 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2423 		if (ret) {
2424 			DRV_LOG(INFO, "failed to change file descriptor"
2425 				" devx comp");
2426 			return;
2427 		}
2428 		sh->intr_handle_devx.fd = devx_comp->fd;
2429 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2430 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2431 					mlx5_dev_interrupt_handler_devx, sh)) {
2432 			DRV_LOG(INFO, "Fail to install the devx shared"
2433 				" interrupt.");
2434 			sh->intr_handle_devx.fd = -1;
2435 		}
2436 #endif /* HAVE_IBV_DEVX_ASYNC */
2437 	}
2438 }
2439 
2440 /**
2441  * Uninstall shared asynchronous device events handler.
2442  * This function is implemented to support event sharing
2443  * between multiple ports of single IB device.
2444  *
2445  * @param dev
2446  *   Pointer to mlx5_dev_ctx_shared object.
2447  */
2448 void
2449 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2450 {
2451 	if (sh->intr_handle.fd >= 0)
2452 		mlx5_intr_callback_unregister(&sh->intr_handle,
2453 					      mlx5_dev_interrupt_handler, sh);
2454 #ifdef HAVE_IBV_DEVX_ASYNC
2455 	if (sh->intr_handle_devx.fd >= 0)
2456 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2457 				  mlx5_dev_interrupt_handler_devx, sh);
2458 	if (sh->devx_comp)
2459 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2460 #endif
2461 }
2462 
2463 /**
2464  * Read statistics by a named counter.
2465  *
2466  * @param[in] priv
2467  *   Pointer to the private device data structure.
2468  * @param[in] ctr_name
2469  *   Pointer to the name of the statistic counter to read
2470  * @param[out] stat
2471  *   Pointer to read statistic value.
2472  * @return
2473  *   0 on success and stat is valud, 1 if failed to read the value
2474  *   rte_errno is set.
2475  *
2476  */
2477 int
2478 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2479 		      uint64_t *stat)
2480 {
2481 	int fd;
2482 
2483 	if (priv->sh) {
2484 		if (priv->q_counters != NULL &&
2485 		    strcmp(ctr_name, "out_of_buffer") == 0)
2486 			return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2487 							   0, (uint32_t *)stat);
2488 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2489 		      priv->sh->ibdev_path,
2490 		      priv->dev_port,
2491 		      ctr_name);
2492 		fd = open(path, O_RDONLY);
2493 		/*
2494 		 * in switchdev the file location is not per port
2495 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2496 		 */
2497 		if (fd == -1) {
2498 			MKSTR(path1, "%s/hw_counters/%s",
2499 			      priv->sh->ibdev_path,
2500 			      ctr_name);
2501 			fd = open(path1, O_RDONLY);
2502 		}
2503 		if (fd != -1) {
2504 			char buf[21] = {'\0'};
2505 			ssize_t n = read(fd, buf, sizeof(buf));
2506 
2507 			close(fd);
2508 			if (n != -1) {
2509 				*stat = strtoull(buf, NULL, 10);
2510 				return 0;
2511 			}
2512 		}
2513 	}
2514 	*stat = 0;
2515 	return 1;
2516 }
2517 
2518 /**
2519  * Set the reg_mr and dereg_mr call backs
2520  *
2521  * @param reg_mr_cb[out]
2522  *   Pointer to reg_mr func
2523  * @param dereg_mr_cb[out]
2524  *   Pointer to dereg_mr func
2525  *
2526  */
2527 void
2528 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2529 		      mlx5_dereg_mr_t *dereg_mr_cb)
2530 {
2531 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2532 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2533 }
2534 
2535 /**
2536  * Remove a MAC address from device
2537  *
2538  * @param dev
2539  *   Pointer to Ethernet device structure.
2540  * @param index
2541  *   MAC address index.
2542  */
2543 void
2544 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2545 {
2546 	struct mlx5_priv *priv = dev->data->dev_private;
2547 	const int vf = priv->config.vf;
2548 
2549 	if (vf)
2550 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2551 					mlx5_ifindex(dev), priv->mac_own,
2552 					&dev->data->mac_addrs[index], index);
2553 }
2554 
2555 /**
2556  * Adds a MAC address to the device
2557  *
2558  * @param dev
2559  *   Pointer to Ethernet device structure.
2560  * @param mac_addr
2561  *   MAC address to register.
2562  * @param index
2563  *   MAC address index.
2564  *
2565  * @return
2566  *   0 on success, a negative errno value otherwise
2567  */
2568 int
2569 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2570 		     uint32_t index)
2571 {
2572 	struct mlx5_priv *priv = dev->data->dev_private;
2573 	const int vf = priv->config.vf;
2574 	int ret = 0;
2575 
2576 	if (vf)
2577 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2578 					   mlx5_ifindex(dev), priv->mac_own,
2579 					   mac, index);
2580 	return ret;
2581 }
2582 
2583 /**
2584  * Modify a VF MAC address
2585  *
2586  * @param priv
2587  *   Pointer to device private data.
2588  * @param mac_addr
2589  *   MAC address to modify into.
2590  * @param iface_idx
2591  *   Net device interface index
2592  * @param vf_index
2593  *   VF index
2594  *
2595  * @return
2596  *   0 on success, a negative errno value otherwise
2597  */
2598 int
2599 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2600 			   unsigned int iface_idx,
2601 			   struct rte_ether_addr *mac_addr,
2602 			   int vf_index)
2603 {
2604 	return mlx5_nl_vf_mac_addr_modify
2605 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2606 }
2607 
2608 /**
2609  * Set device promiscuous mode
2610  *
2611  * @param dev
2612  *   Pointer to Ethernet device structure.
2613  * @param enable
2614  *   0 - promiscuous is disabled, otherwise - enabled
2615  *
2616  * @return
2617  *   0 on success, a negative error value otherwise
2618  */
2619 int
2620 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2621 {
2622 	struct mlx5_priv *priv = dev->data->dev_private;
2623 
2624 	return mlx5_nl_promisc(priv->nl_socket_route,
2625 			       mlx5_ifindex(dev), !!enable);
2626 }
2627 
2628 /**
2629  * Set device promiscuous mode
2630  *
2631  * @param dev
2632  *   Pointer to Ethernet device structure.
2633  * @param enable
2634  *   0 - all multicase is disabled, otherwise - enabled
2635  *
2636  * @return
2637  *   0 on success, a negative error value otherwise
2638  */
2639 int
2640 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2641 {
2642 	struct mlx5_priv *priv = dev->data->dev_private;
2643 
2644 	return mlx5_nl_allmulti(priv->nl_socket_route,
2645 				mlx5_ifindex(dev), !!enable);
2646 }
2647 
2648 /**
2649  * Flush device MAC addresses
2650  *
2651  * @param dev
2652  *   Pointer to Ethernet device structure.
2653  *
2654  */
2655 void
2656 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2657 {
2658 	struct mlx5_priv *priv = dev->data->dev_private;
2659 
2660 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2661 			       dev->data->mac_addrs,
2662 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2663 }
2664