xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 6f04fa45211e3ca4fb9a7e9ceabc682fa47e88bf)
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 /**
649  * Spawn an Ethernet device from Verbs information.
650  *
651  * @param dpdk_dev
652  *   Backing DPDK device.
653  * @param spawn
654  *   Verbs device parameters (name, port, switch_info) to spawn.
655  * @param config
656  *   Device configuration parameters.
657  *
658  * @return
659  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
660  *   is set. The following errors are defined:
661  *
662  *   EBUSY: device is not supposed to be spawned.
663  *   EEXIST: device is already spawned
664  */
665 static struct rte_eth_dev *
666 mlx5_dev_spawn(struct rte_device *dpdk_dev,
667 	       struct mlx5_dev_spawn_data *spawn,
668 	       struct mlx5_dev_config *config)
669 {
670 	const struct mlx5_switch_info *switch_info = &spawn->info;
671 	struct mlx5_dev_ctx_shared *sh = NULL;
672 	struct ibv_port_attr port_attr;
673 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
674 	struct rte_eth_dev *eth_dev = NULL;
675 	struct mlx5_priv *priv = NULL;
676 	int err = 0;
677 	unsigned int hw_padding = 0;
678 	unsigned int mps;
679 	unsigned int tunnel_en = 0;
680 	unsigned int mpls_en = 0;
681 	unsigned int swp = 0;
682 	unsigned int mprq = 0;
683 	unsigned int mprq_min_stride_size_n = 0;
684 	unsigned int mprq_max_stride_size_n = 0;
685 	unsigned int mprq_min_stride_num_n = 0;
686 	unsigned int mprq_max_stride_num_n = 0;
687 	struct rte_ether_addr mac;
688 	char name[RTE_ETH_NAME_MAX_LEN];
689 	int own_domain_id = 0;
690 	uint16_t port_id;
691 	unsigned int i;
692 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
693 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
694 #endif
695 
696 	/* Determine if this port representor is supposed to be spawned. */
697 	if (switch_info->representor && dpdk_dev->devargs) {
698 		struct rte_eth_devargs eth_da;
699 
700 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
701 		if (err) {
702 			rte_errno = -err;
703 			DRV_LOG(ERR, "failed to process device arguments: %s",
704 				strerror(rte_errno));
705 			return NULL;
706 		}
707 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
708 			if (eth_da.representor_ports[i] ==
709 			    (uint16_t)switch_info->port_name)
710 				break;
711 		if (i == eth_da.nb_representor_ports) {
712 			rte_errno = EBUSY;
713 			return NULL;
714 		}
715 	}
716 	/* Build device name. */
717 	if (spawn->pf_bond <  0) {
718 		/* Single device. */
719 		if (!switch_info->representor)
720 			strlcpy(name, dpdk_dev->name, sizeof(name));
721 		else
722 			snprintf(name, sizeof(name), "%s_representor_%u",
723 				 dpdk_dev->name, switch_info->port_name);
724 	} else {
725 		/* Bonding device. */
726 		if (!switch_info->representor)
727 			snprintf(name, sizeof(name), "%s_%s",
728 				 dpdk_dev->name,
729 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
730 		else
731 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
732 				 dpdk_dev->name,
733 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
734 				 switch_info->port_name);
735 	}
736 	/* check if the device is already spawned */
737 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
738 		rte_errno = EEXIST;
739 		return NULL;
740 	}
741 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
742 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
743 		struct mlx5_mp_id mp_id;
744 
745 		eth_dev = rte_eth_dev_attach_secondary(name);
746 		if (eth_dev == NULL) {
747 			DRV_LOG(ERR, "can not attach rte ethdev");
748 			rte_errno = ENOMEM;
749 			return NULL;
750 		}
751 		priv = eth_dev->data->dev_private;
752 		if (priv->sh->bond_dev != UINT16_MAX)
753 			/* For bonding port, use primary PCI device. */
754 			eth_dev->device =
755 				rte_eth_devices[priv->sh->bond_dev].device;
756 		else
757 			eth_dev->device = dpdk_dev;
758 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
759 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
760 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
761 		err = mlx5_proc_priv_init(eth_dev);
762 		if (err)
763 			return NULL;
764 		mp_id.port_id = eth_dev->data->port_id;
765 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
766 		/* Receive command fd from primary process */
767 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
768 		if (err < 0)
769 			goto err_secondary;
770 		/* Remap UAR for Tx queues. */
771 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
772 		if (err)
773 			goto err_secondary;
774 		/*
775 		 * Ethdev pointer is still required as input since
776 		 * the primary device is not accessible from the
777 		 * secondary process.
778 		 */
779 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
780 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
781 		return eth_dev;
782 err_secondary:
783 		mlx5_dev_close(eth_dev);
784 		return NULL;
785 	}
786 	/*
787 	 * Some parameters ("tx_db_nc" in particularly) are needed in
788 	 * advance to create dv/verbs device context. We proceed the
789 	 * devargs here to get ones, and later proceed devargs again
790 	 * to override some hardware settings.
791 	 */
792 	err = mlx5_args(config, dpdk_dev->devargs);
793 	if (err) {
794 		err = rte_errno;
795 		DRV_LOG(ERR, "failed to process device arguments: %s",
796 			strerror(rte_errno));
797 		goto error;
798 	}
799 	if (config->dv_miss_info) {
800 		if (switch_info->master || switch_info->representor)
801 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
802 	}
803 	mlx5_malloc_mem_select(config->sys_mem_en);
804 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
805 	if (!sh)
806 		return NULL;
807 	config->devx = sh->devx;
808 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
809 	config->dest_tir = 1;
810 #endif
811 #ifdef HAVE_IBV_MLX5_MOD_SWP
812 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
813 #endif
814 	/*
815 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
816 	 * as all ConnectX-5 devices.
817 	 */
818 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
819 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
820 #endif
821 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
822 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
823 #endif
824 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
825 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
826 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
827 			DRV_LOG(DEBUG, "enhanced MPW is supported");
828 			mps = MLX5_MPW_ENHANCED;
829 		} else {
830 			DRV_LOG(DEBUG, "MPW is supported");
831 			mps = MLX5_MPW;
832 		}
833 	} else {
834 		DRV_LOG(DEBUG, "MPW isn't supported");
835 		mps = MLX5_MPW_DISABLED;
836 	}
837 #ifdef HAVE_IBV_MLX5_MOD_SWP
838 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
839 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
840 	DRV_LOG(DEBUG, "SWP support: %u", swp);
841 #endif
842 	config->swp = !!swp;
843 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
844 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
845 		struct mlx5dv_striding_rq_caps mprq_caps =
846 			dv_attr.striding_rq_caps;
847 
848 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
849 			mprq_caps.min_single_stride_log_num_of_bytes);
850 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
851 			mprq_caps.max_single_stride_log_num_of_bytes);
852 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
853 			mprq_caps.min_single_wqe_log_num_of_strides);
854 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
855 			mprq_caps.max_single_wqe_log_num_of_strides);
856 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
857 			mprq_caps.supported_qpts);
858 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
859 		mprq = 1;
860 		mprq_min_stride_size_n =
861 			mprq_caps.min_single_stride_log_num_of_bytes;
862 		mprq_max_stride_size_n =
863 			mprq_caps.max_single_stride_log_num_of_bytes;
864 		mprq_min_stride_num_n =
865 			mprq_caps.min_single_wqe_log_num_of_strides;
866 		mprq_max_stride_num_n =
867 			mprq_caps.max_single_wqe_log_num_of_strides;
868 	}
869 #endif
870 	/* Rx CQE compression is enabled by default. */
871 	config->cqe_comp = 1;
872 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
873 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
874 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
875 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
876 			     (dv_attr.tunnel_offloads_caps &
877 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
878 			     (dv_attr.tunnel_offloads_caps &
879 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
880 	}
881 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
882 		tunnel_en ? "" : "not ");
883 #else
884 	DRV_LOG(WARNING,
885 		"tunnel offloading disabled due to old OFED/rdma-core version");
886 #endif
887 	config->tunnel_en = tunnel_en;
888 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
889 	mpls_en = ((dv_attr.tunnel_offloads_caps &
890 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
891 		   (dv_attr.tunnel_offloads_caps &
892 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
893 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
894 		mpls_en ? "" : "not ");
895 #else
896 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
897 		" old OFED/rdma-core version or firmware configuration");
898 #endif
899 	config->mpls_en = mpls_en;
900 	/* Check port status. */
901 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
902 	if (err) {
903 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
904 		goto error;
905 	}
906 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
907 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
908 		err = EINVAL;
909 		goto error;
910 	}
911 	if (port_attr.state != IBV_PORT_ACTIVE)
912 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
913 			mlx5_glue->port_state_str(port_attr.state),
914 			port_attr.state);
915 	/* Allocate private eth device data. */
916 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
917 			   sizeof(*priv),
918 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
919 	if (priv == NULL) {
920 		DRV_LOG(ERR, "priv allocation failure");
921 		err = ENOMEM;
922 		goto error;
923 	}
924 	priv->sh = sh;
925 	priv->dev_port = spawn->phys_port;
926 	priv->pci_dev = spawn->pci_dev;
927 	priv->mtu = RTE_ETHER_MTU;
928 	/* Some internal functions rely on Netlink sockets, open them now. */
929 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
930 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
931 	priv->representor = !!switch_info->representor;
932 	priv->master = !!switch_info->master;
933 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
934 	priv->vport_meta_tag = 0;
935 	priv->vport_meta_mask = 0;
936 	priv->pf_bond = spawn->pf_bond;
937 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
938 	/*
939 	 * The DevX port query API is implemented. E-Switch may use
940 	 * either vport or reg_c[0] metadata register to match on
941 	 * vport index. The engaged part of metadata register is
942 	 * defined by mask.
943 	 */
944 	if (switch_info->representor || switch_info->master) {
945 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
946 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
947 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
948 						 &devx_port);
949 		if (err) {
950 			DRV_LOG(WARNING,
951 				"can't query devx port %d on device %s",
952 				spawn->phys_port,
953 				mlx5_os_get_dev_device_name(spawn->phys_dev));
954 			devx_port.comp_mask = 0;
955 		}
956 	}
957 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
958 		priv->vport_meta_tag = devx_port.reg_c_0.value;
959 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
960 		if (!priv->vport_meta_mask) {
961 			DRV_LOG(ERR, "vport zero mask for port %d"
962 				     " on bonding device %s",
963 				     spawn->phys_port,
964 				     mlx5_os_get_dev_device_name
965 							(spawn->phys_dev));
966 			err = ENOTSUP;
967 			goto error;
968 		}
969 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
970 			DRV_LOG(ERR, "invalid vport tag for port %d"
971 				     " on bonding device %s",
972 				     spawn->phys_port,
973 				     mlx5_os_get_dev_device_name
974 							(spawn->phys_dev));
975 			err = ENOTSUP;
976 			goto error;
977 		}
978 	}
979 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
980 		priv->vport_id = devx_port.vport_num;
981 	} else if (spawn->pf_bond >= 0) {
982 		DRV_LOG(ERR, "can't deduce vport index for port %d"
983 			     " on bonding device %s",
984 			     spawn->phys_port,
985 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
986 		err = ENOTSUP;
987 		goto error;
988 	} else {
989 		/* Suppose vport index in compatible way. */
990 		priv->vport_id = switch_info->representor ?
991 				 switch_info->port_name + 1 : -1;
992 	}
993 #else
994 	/*
995 	 * Kernel/rdma_core support single E-Switch per PF configurations
996 	 * only and vport_id field contains the vport index for
997 	 * associated VF, which is deduced from representor port name.
998 	 * For example, let's have the IB device port 10, it has
999 	 * attached network device eth0, which has port name attribute
1000 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1001 	 * as 3 (2+1). This assigning schema should be changed if the
1002 	 * multiple E-Switch instances per PF configurations or/and PCI
1003 	 * subfunctions are added.
1004 	 */
1005 	priv->vport_id = switch_info->representor ?
1006 			 switch_info->port_name + 1 : -1;
1007 #endif
1008 	/* representor_id field keeps the unmodified VF index. */
1009 	priv->representor_id = switch_info->representor ?
1010 			       switch_info->port_name : -1;
1011 	/*
1012 	 * Look for sibling devices in order to reuse their switch domain
1013 	 * if any, otherwise allocate one.
1014 	 */
1015 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1016 		const struct mlx5_priv *opriv =
1017 			rte_eth_devices[port_id].data->dev_private;
1018 
1019 		if (!opriv ||
1020 		    opriv->sh != priv->sh ||
1021 			opriv->domain_id ==
1022 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1023 			continue;
1024 		priv->domain_id = opriv->domain_id;
1025 		break;
1026 	}
1027 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1028 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1029 		if (err) {
1030 			err = rte_errno;
1031 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1032 				strerror(rte_errno));
1033 			goto error;
1034 		}
1035 		own_domain_id = 1;
1036 	}
1037 	/* Override some values set by hardware configuration. */
1038 	mlx5_args(config, dpdk_dev->devargs);
1039 	err = mlx5_dev_check_sibling_config(priv, config);
1040 	if (err)
1041 		goto error;
1042 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1043 			    IBV_DEVICE_RAW_IP_CSUM);
1044 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1045 		(config->hw_csum ? "" : "not "));
1046 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1047 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1048 	DRV_LOG(DEBUG, "counters are not supported");
1049 #endif
1050 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1051 	if (config->dv_flow_en) {
1052 		DRV_LOG(WARNING, "DV flow is not supported");
1053 		config->dv_flow_en = 0;
1054 	}
1055 #endif
1056 	config->ind_table_max_size =
1057 		sh->device_attr.max_rwq_indirection_table_size;
1058 	/*
1059 	 * Remove this check once DPDK supports larger/variable
1060 	 * indirection tables.
1061 	 */
1062 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1063 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1064 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1065 		config->ind_table_max_size);
1066 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1067 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1068 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1069 		(config->hw_vlan_strip ? "" : "not "));
1070 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1071 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1072 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1073 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1074 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1075 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1076 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1077 #endif
1078 	if (config->hw_padding && !hw_padding) {
1079 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1080 		config->hw_padding = 0;
1081 	} else if (config->hw_padding) {
1082 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1083 	}
1084 	config->tso = (sh->device_attr.max_tso > 0 &&
1085 		      (sh->device_attr.tso_supported_qpts &
1086 		       (1 << IBV_QPT_RAW_PACKET)));
1087 	if (config->tso)
1088 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1089 	/*
1090 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1091 	 * by default.
1092 	 */
1093 	if (config->mps == MLX5_ARG_UNSET)
1094 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1095 							  MLX5_MPW_DISABLED;
1096 	else
1097 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1098 	DRV_LOG(INFO, "%sMPS is %s",
1099 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1100 		config->mps == MLX5_MPW ? "legacy " : "",
1101 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1102 	if (config->devx) {
1103 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1104 		if (err) {
1105 			err = -err;
1106 			goto error;
1107 		}
1108 		/* Check relax ordering support. */
1109 		if (!haswell_broadwell_cpu) {
1110 			sh->cmng.relaxed_ordering_write =
1111 				config->hca_attr.relaxed_ordering_write;
1112 			sh->cmng.relaxed_ordering_read =
1113 				config->hca_attr.relaxed_ordering_read;
1114 		} else {
1115 			sh->cmng.relaxed_ordering_read = 0;
1116 			sh->cmng.relaxed_ordering_write = 0;
1117 		}
1118 		/* Check for LRO support. */
1119 		if (config->dest_tir && config->hca_attr.lro_cap &&
1120 		    config->dv_flow_en) {
1121 			/* TBD check tunnel lro caps. */
1122 			config->lro.supported = config->hca_attr.lro_cap;
1123 			DRV_LOG(DEBUG, "Device supports LRO");
1124 			/*
1125 			 * If LRO timeout is not configured by application,
1126 			 * use the minimal supported value.
1127 			 */
1128 			if (!config->lro.timeout)
1129 				config->lro.timeout =
1130 				config->hca_attr.lro_timer_supported_periods[0];
1131 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1132 				config->lro.timeout);
1133 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1134 				"required for coalescing is %d bytes",
1135 				config->hca_attr.lro_min_mss_size);
1136 		}
1137 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1138 		if (config->hca_attr.qos.sup &&
1139 		    config->hca_attr.qos.flow_meter_old &&
1140 		    config->dv_flow_en) {
1141 			uint8_t reg_c_mask =
1142 				config->hca_attr.qos.flow_meter_reg_c_ids;
1143 			/*
1144 			 * Meter needs two REG_C's for color match and pre-sfx
1145 			 * flow match. Here get the REG_C for color match.
1146 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1147 			 */
1148 			reg_c_mask &= 0xfc;
1149 			if (__builtin_popcount(reg_c_mask) < 1) {
1150 				priv->mtr_en = 0;
1151 				DRV_LOG(WARNING, "No available register for"
1152 					" meter.");
1153 			} else {
1154 				/*
1155 				 * The meter color register is used by the
1156 				 * flow-hit feature as well.
1157 				 * The flow-hit feature must use REG_C_3
1158 				 * Prefer REG_C_3 if it is available.
1159 				 */
1160 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1161 					priv->mtr_color_reg = REG_C_3;
1162 				else
1163 					priv->mtr_color_reg = ffs(reg_c_mask)
1164 							      - 1 + REG_C_0;
1165 				priv->mtr_en = 1;
1166 				priv->mtr_reg_share =
1167 				      config->hca_attr.qos.flow_meter;
1168 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1169 					priv->mtr_color_reg);
1170 			}
1171 		}
1172 #endif
1173 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1174 		if (config->hca_attr.flow_hit_aso &&
1175 		    priv->mtr_color_reg == REG_C_3) {
1176 			sh->flow_hit_aso_en = 1;
1177 			err = mlx5_flow_aso_age_mng_init(sh);
1178 			if (err) {
1179 				err = -err;
1180 				goto error;
1181 			}
1182 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1183 		}
1184 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1185 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1186 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1187 		    config->dv_flow_en) {
1188 			priv->sampler_en = 1;
1189 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1190 		} else {
1191 			priv->sampler_en = 0;
1192 			if (!config->hca_attr.log_max_ft_sampler_num)
1193 				DRV_LOG(WARNING, "No available register for"
1194 						" Sampler.");
1195 			else
1196 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1197 		}
1198 #endif
1199 	}
1200 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1201 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1202 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1203 		config->cqe_comp = 0;
1204 	}
1205 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1206 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1207 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1208 				 " format isn't supported.");
1209 		config->cqe_comp = 0;
1210 	}
1211 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1212 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1213 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1214 				 " format isn't supported.");
1215 		config->cqe_comp = 0;
1216 	}
1217 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1218 			config->cqe_comp ? "" : "not ");
1219 	if (config->tx_pp) {
1220 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1221 			config->hca_attr.dev_freq_khz);
1222 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1223 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1224 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1225 			config->hca_attr.cross_channel ? "" : "not ");
1226 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1227 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1228 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1229 			config->hca_attr.non_wire_sq ? "" : "not ");
1230 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1231 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1232 			config->hca_attr.log_max_static_sq_wq);
1233 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1234 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1235 		if (!config->devx) {
1236 			DRV_LOG(ERR, "DevX is required for packet pacing");
1237 			err = ENODEV;
1238 			goto error;
1239 		}
1240 		if (!config->hca_attr.qos.packet_pacing) {
1241 			DRV_LOG(ERR, "Packet pacing is not supported");
1242 			err = ENODEV;
1243 			goto error;
1244 		}
1245 		if (!config->hca_attr.cross_channel) {
1246 			DRV_LOG(ERR, "Cross channel operations are"
1247 				     " required for packet pacing");
1248 			err = ENODEV;
1249 			goto error;
1250 		}
1251 		if (!config->hca_attr.wqe_index_ignore) {
1252 			DRV_LOG(ERR, "WQE index ignore feature is"
1253 				     " required for packet pacing");
1254 			err = ENODEV;
1255 			goto error;
1256 		}
1257 		if (!config->hca_attr.non_wire_sq) {
1258 			DRV_LOG(ERR, "Non-wire SQ feature is"
1259 				     " required for packet pacing");
1260 			err = ENODEV;
1261 			goto error;
1262 		}
1263 		if (!config->hca_attr.log_max_static_sq_wq) {
1264 			DRV_LOG(ERR, "Static WQE SQ feature is"
1265 				     " required for packet pacing");
1266 			err = ENODEV;
1267 			goto error;
1268 		}
1269 		if (!config->hca_attr.qos.wqe_rate_pp) {
1270 			DRV_LOG(ERR, "WQE rate mode is required"
1271 				     " for packet pacing");
1272 			err = ENODEV;
1273 			goto error;
1274 		}
1275 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1276 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1277 			     " can't create queues for packet pacing");
1278 		err = ENODEV;
1279 		goto error;
1280 #endif
1281 	}
1282 	if (config->devx) {
1283 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1284 
1285 		err = config->hca_attr.access_register_user ?
1286 			mlx5_devx_cmd_register_read
1287 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1288 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1289 		if (!err) {
1290 			uint32_t ts_mode;
1291 
1292 			/* MTUTC register is read successfully. */
1293 			ts_mode = MLX5_GET(register_mtutc, reg,
1294 					   time_stamp_mode);
1295 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1296 				config->rt_timestamp = 1;
1297 		} else {
1298 			/* Kernel does not support register reading. */
1299 			if (config->hca_attr.dev_freq_khz ==
1300 						 (NS_PER_S / MS_PER_S))
1301 				config->rt_timestamp = 1;
1302 		}
1303 	}
1304 	/*
1305 	 * If HW has bug working with tunnel packet decapsulation and
1306 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1307 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1308 	 */
1309 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1310 		config->hw_fcs_strip = 0;
1311 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1312 		(config->hw_fcs_strip ? "" : "not "));
1313 	if (config->mprq.enabled && mprq) {
1314 		if (config->mprq.stride_num_n &&
1315 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1316 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1317 			config->mprq.stride_num_n =
1318 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1319 						mprq_min_stride_num_n),
1320 					mprq_max_stride_num_n);
1321 			DRV_LOG(WARNING,
1322 				"the number of strides"
1323 				" for Multi-Packet RQ is out of range,"
1324 				" setting default value (%u)",
1325 				1 << config->mprq.stride_num_n);
1326 		}
1327 		if (config->mprq.stride_size_n &&
1328 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1329 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1330 			config->mprq.stride_size_n =
1331 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1332 						mprq_min_stride_size_n),
1333 					mprq_max_stride_size_n);
1334 			DRV_LOG(WARNING,
1335 				"the size of a stride"
1336 				" for Multi-Packet RQ is out of range,"
1337 				" setting default value (%u)",
1338 				1 << config->mprq.stride_size_n);
1339 		}
1340 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1341 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1342 	} else if (config->mprq.enabled && !mprq) {
1343 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1344 		config->mprq.enabled = 0;
1345 	}
1346 	if (config->max_dump_files_num == 0)
1347 		config->max_dump_files_num = 128;
1348 	eth_dev = rte_eth_dev_allocate(name);
1349 	if (eth_dev == NULL) {
1350 		DRV_LOG(ERR, "can not allocate rte ethdev");
1351 		err = ENOMEM;
1352 		goto error;
1353 	}
1354 	if (priv->representor) {
1355 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1356 		eth_dev->data->representor_id = priv->representor_id;
1357 	}
1358 	priv->mp_id.port_id = eth_dev->data->port_id;
1359 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1360 	/*
1361 	 * Store associated network device interface index. This index
1362 	 * is permanent throughout the lifetime of device. So, we may store
1363 	 * the ifindex here and use the cached value further.
1364 	 */
1365 	MLX5_ASSERT(spawn->ifindex);
1366 	priv->if_index = spawn->ifindex;
1367 	if (priv->pf_bond >= 0 && priv->master) {
1368 		/* Get bond interface info */
1369 		err = mlx5_sysfs_bond_info(priv->if_index,
1370 				     &priv->bond_ifindex,
1371 				     priv->bond_name);
1372 		if (err)
1373 			DRV_LOG(ERR, "unable to get bond info: %s",
1374 				strerror(rte_errno));
1375 		else
1376 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1377 				priv->if_index, priv->bond_ifindex,
1378 				priv->bond_name);
1379 	}
1380 	eth_dev->data->dev_private = priv;
1381 	priv->dev_data = eth_dev->data;
1382 	eth_dev->data->mac_addrs = priv->mac;
1383 	if (spawn->pf_bond < 0) {
1384 		eth_dev->device = dpdk_dev;
1385 	} else {
1386 		/* Use primary bond PCI as device. */
1387 		if (sh->bond_dev == UINT16_MAX) {
1388 			sh->bond_dev = eth_dev->data->port_id;
1389 			eth_dev->device = dpdk_dev;
1390 		} else {
1391 			eth_dev->device = rte_eth_devices[sh->bond_dev].device;
1392 		}
1393 	}
1394 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1395 	/* Configure the first MAC address by default. */
1396 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1397 		DRV_LOG(ERR,
1398 			"port %u cannot get MAC address, is mlx5_en"
1399 			" loaded? (errno: %s)",
1400 			eth_dev->data->port_id, strerror(rte_errno));
1401 		err = ENODEV;
1402 		goto error;
1403 	}
1404 	DRV_LOG(INFO,
1405 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1406 		eth_dev->data->port_id,
1407 		mac.addr_bytes[0], mac.addr_bytes[1],
1408 		mac.addr_bytes[2], mac.addr_bytes[3],
1409 		mac.addr_bytes[4], mac.addr_bytes[5]);
1410 #ifdef RTE_LIBRTE_MLX5_DEBUG
1411 	{
1412 		char ifname[MLX5_NAMESIZE];
1413 
1414 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1415 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1416 				eth_dev->data->port_id, ifname);
1417 		else
1418 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1419 				eth_dev->data->port_id);
1420 	}
1421 #endif
1422 	/* Get actual MTU if possible. */
1423 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1424 	if (err) {
1425 		err = rte_errno;
1426 		goto error;
1427 	}
1428 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1429 		priv->mtu);
1430 	/* Initialize burst functions to prevent crashes before link-up. */
1431 	eth_dev->rx_pkt_burst = removed_rx_burst;
1432 	eth_dev->tx_pkt_burst = removed_tx_burst;
1433 	eth_dev->dev_ops = &mlx5_dev_ops;
1434 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1435 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1436 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1437 	/* Register MAC address. */
1438 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1439 	if (config->vf && config->vf_nl_en)
1440 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1441 				      mlx5_ifindex(eth_dev),
1442 				      eth_dev->data->mac_addrs,
1443 				      MLX5_MAX_MAC_ADDRESSES);
1444 	priv->flows = 0;
1445 	priv->ctrl_flows = 0;
1446 	rte_spinlock_init(&priv->flow_list_lock);
1447 	TAILQ_INIT(&priv->flow_meters);
1448 	TAILQ_INIT(&priv->flow_meter_profiles);
1449 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1450 	mlx5_glue->dv_set_context_attr(sh->ctx,
1451 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1452 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1453 				.alloc = &mlx5_alloc_verbs_buf,
1454 				.free = &mlx5_free_verbs_buf,
1455 				.data = sh,
1456 			}));
1457 	/* Bring Ethernet device up. */
1458 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1459 		eth_dev->data->port_id);
1460 	mlx5_set_link_up(eth_dev);
1461 	/*
1462 	 * Even though the interrupt handler is not installed yet,
1463 	 * interrupts will still trigger on the async_fd from
1464 	 * Verbs context returned by ibv_open_device().
1465 	 */
1466 	mlx5_link_update(eth_dev, 0);
1467 #ifdef HAVE_MLX5DV_DR_ESWITCH
1468 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1469 	      (switch_info->representor || switch_info->master)))
1470 		config->dv_esw_en = 0;
1471 #else
1472 	config->dv_esw_en = 0;
1473 #endif
1474 	/* Detect minimal data bytes to inline. */
1475 	mlx5_set_min_inline(spawn, config);
1476 	/* Store device configuration on private structure. */
1477 	priv->config = *config;
1478 	/* Create context for virtual machine VLAN workaround. */
1479 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1480 	if (config->dv_flow_en) {
1481 		err = mlx5_alloc_shared_dr(priv);
1482 		if (err)
1483 			goto error;
1484 	}
1485 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1486 		priv->obj_ops = devx_obj_ops;
1487 		priv->obj_ops.drop_action_create =
1488 						ibv_obj_ops.drop_action_create;
1489 		priv->obj_ops.drop_action_destroy =
1490 						ibv_obj_ops.drop_action_destroy;
1491 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1492 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1493 #else
1494 		if (config->dv_esw_en)
1495 			priv->obj_ops.txq_obj_modify =
1496 						ibv_obj_ops.txq_obj_modify;
1497 #endif
1498 		/* Use specific wrappers for Tx object. */
1499 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1500 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1501 
1502 	} else {
1503 		priv->obj_ops = ibv_obj_ops;
1504 	}
1505 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1506 	if (!priv->drop_queue.hrxq)
1507 		goto error;
1508 	/* Supported Verbs flow priority number detection. */
1509 	err = mlx5_flow_discover_priorities(eth_dev);
1510 	if (err < 0) {
1511 		err = -err;
1512 		goto error;
1513 	}
1514 	priv->config.flow_prio = err;
1515 	if (!priv->config.dv_esw_en &&
1516 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1517 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1518 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1519 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1520 	}
1521 	mlx5_set_metadata_mask(eth_dev);
1522 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1523 	    !priv->sh->dv_regc0_mask) {
1524 		DRV_LOG(ERR, "metadata mode %u is not supported "
1525 			     "(no metadata reg_c[0] is available)",
1526 			     priv->config.dv_xmeta_en);
1527 			err = ENOTSUP;
1528 			goto error;
1529 	}
1530 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1531 			     mlx5_hrxq_create_cb,
1532 			     mlx5_hrxq_match_cb,
1533 			     mlx5_hrxq_remove_cb);
1534 	/* Query availability of metadata reg_c's. */
1535 	err = mlx5_flow_discover_mreg_c(eth_dev);
1536 	if (err < 0) {
1537 		err = -err;
1538 		goto error;
1539 	}
1540 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1541 		DRV_LOG(DEBUG,
1542 			"port %u extensive metadata register is not supported",
1543 			eth_dev->data->port_id);
1544 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1545 			DRV_LOG(ERR, "metadata mode %u is not supported "
1546 				     "(no metadata registers available)",
1547 				     priv->config.dv_xmeta_en);
1548 			err = ENOTSUP;
1549 			goto error;
1550 		}
1551 	}
1552 	if (priv->config.dv_flow_en &&
1553 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1554 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1555 	    priv->sh->dv_regc0_mask) {
1556 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1557 						      MLX5_FLOW_MREG_HTABLE_SZ,
1558 						      0, 0,
1559 						      flow_dv_mreg_create_cb,
1560 						      flow_dv_mreg_match_cb,
1561 						      flow_dv_mreg_remove_cb);
1562 		if (!priv->mreg_cp_tbl) {
1563 			err = ENOMEM;
1564 			goto error;
1565 		}
1566 		priv->mreg_cp_tbl->ctx = eth_dev;
1567 	}
1568 	rte_spinlock_init(&priv->shared_act_sl);
1569 	mlx5_flow_counter_mode_config(eth_dev);
1570 	if (priv->config.dv_flow_en)
1571 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1572 	return eth_dev;
1573 error:
1574 	if (priv) {
1575 		if (priv->mreg_cp_tbl)
1576 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1577 		if (priv->sh)
1578 			mlx5_os_free_shared_dr(priv);
1579 		if (priv->nl_socket_route >= 0)
1580 			close(priv->nl_socket_route);
1581 		if (priv->nl_socket_rdma >= 0)
1582 			close(priv->nl_socket_rdma);
1583 		if (priv->vmwa_context)
1584 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1585 		if (eth_dev && priv->drop_queue.hrxq)
1586 			mlx5_drop_action_destroy(eth_dev);
1587 		if (own_domain_id)
1588 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1589 		mlx5_cache_list_destroy(&priv->hrxqs);
1590 		mlx5_free(priv);
1591 		if (eth_dev != NULL)
1592 			eth_dev->data->dev_private = NULL;
1593 	}
1594 	if (eth_dev != NULL) {
1595 		/* mac_addrs must not be freed alone because part of
1596 		 * dev_private
1597 		 **/
1598 		eth_dev->data->mac_addrs = NULL;
1599 		rte_eth_dev_release_port(eth_dev);
1600 	}
1601 	if (sh)
1602 		mlx5_free_shared_dev_ctx(sh);
1603 	MLX5_ASSERT(err > 0);
1604 	rte_errno = err;
1605 	return NULL;
1606 }
1607 
1608 /**
1609  * Comparison callback to sort device data.
1610  *
1611  * This is meant to be used with qsort().
1612  *
1613  * @param a[in]
1614  *   Pointer to pointer to first data object.
1615  * @param b[in]
1616  *   Pointer to pointer to second data object.
1617  *
1618  * @return
1619  *   0 if both objects are equal, less than 0 if the first argument is less
1620  *   than the second, greater than 0 otherwise.
1621  */
1622 static int
1623 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1624 {
1625 	const struct mlx5_switch_info *si_a =
1626 		&((const struct mlx5_dev_spawn_data *)a)->info;
1627 	const struct mlx5_switch_info *si_b =
1628 		&((const struct mlx5_dev_spawn_data *)b)->info;
1629 	int ret;
1630 
1631 	/* Master device first. */
1632 	ret = si_b->master - si_a->master;
1633 	if (ret)
1634 		return ret;
1635 	/* Then representor devices. */
1636 	ret = si_b->representor - si_a->representor;
1637 	if (ret)
1638 		return ret;
1639 	/* Unidentified devices come last in no specific order. */
1640 	if (!si_a->representor)
1641 		return 0;
1642 	/* Order representors by name. */
1643 	return si_a->port_name - si_b->port_name;
1644 }
1645 
1646 /**
1647  * Match PCI information for possible slaves of bonding device.
1648  *
1649  * @param[in] ibv_dev
1650  *   Pointer to Infiniband device structure.
1651  * @param[in] pci_dev
1652  *   Pointer to PCI device structure to match PCI address.
1653  * @param[in] nl_rdma
1654  *   Netlink RDMA group socket handle.
1655  *
1656  * @return
1657  *   negative value if no bonding device found, otherwise
1658  *   positive index of slave PF in bonding.
1659  */
1660 static int
1661 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1662 			   const struct rte_pci_device *pci_dev,
1663 			   int nl_rdma)
1664 {
1665 	char ifname[IF_NAMESIZE + 1];
1666 	unsigned int ifindex;
1667 	unsigned int np, i;
1668 	FILE *file = NULL;
1669 	int pf = -1;
1670 
1671 	/*
1672 	 * Try to get master device name. If something goes
1673 	 * wrong suppose the lack of kernel support and no
1674 	 * bonding devices.
1675 	 */
1676 	if (nl_rdma < 0)
1677 		return -1;
1678 	if (!strstr(ibv_dev->name, "bond"))
1679 		return -1;
1680 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1681 	if (!np)
1682 		return -1;
1683 	/*
1684 	 * The Master device might not be on the predefined
1685 	 * port (not on port index 1, it is not garanted),
1686 	 * we have to scan all Infiniband device port and
1687 	 * find master.
1688 	 */
1689 	for (i = 1; i <= np; ++i) {
1690 		/* Check whether Infiniband port is populated. */
1691 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1692 		if (!ifindex)
1693 			continue;
1694 		if (!if_indextoname(ifindex, ifname))
1695 			continue;
1696 		/* Try to read bonding slave names from sysfs. */
1697 		MKSTR(slaves,
1698 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1699 		file = fopen(slaves, "r");
1700 		if (file)
1701 			break;
1702 	}
1703 	if (!file)
1704 		return -1;
1705 	/* Use safe format to check maximal buffer length. */
1706 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1707 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1708 		char tmp_str[IF_NAMESIZE + 32];
1709 		struct rte_pci_addr pci_addr;
1710 		struct mlx5_switch_info	info;
1711 
1712 		/* Process slave interface names in the loop. */
1713 		snprintf(tmp_str, sizeof(tmp_str),
1714 			 "/sys/class/net/%s", ifname);
1715 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1716 			DRV_LOG(WARNING, "can not get PCI address"
1717 					 " for netdev \"%s\"", ifname);
1718 			continue;
1719 		}
1720 		if (pci_dev->addr.domain != pci_addr.domain ||
1721 		    pci_dev->addr.bus != pci_addr.bus ||
1722 		    pci_dev->addr.devid != pci_addr.devid ||
1723 		    pci_dev->addr.function != pci_addr.function)
1724 			continue;
1725 		/* Slave interface PCI address match found. */
1726 		fclose(file);
1727 		snprintf(tmp_str, sizeof(tmp_str),
1728 			 "/sys/class/net/%s/phys_port_name", ifname);
1729 		file = fopen(tmp_str, "rb");
1730 		if (!file)
1731 			break;
1732 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1733 		if (fscanf(file, "%32s", tmp_str) == 1)
1734 			mlx5_translate_port_name(tmp_str, &info);
1735 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1736 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1737 			pf = info.port_name;
1738 		break;
1739 	}
1740 	if (file)
1741 		fclose(file);
1742 	return pf;
1743 }
1744 
1745 /**
1746  * DPDK callback to register a PCI device.
1747  *
1748  * This function spawns Ethernet devices out of a given PCI device.
1749  *
1750  * @param[in] pci_drv
1751  *   PCI driver structure (mlx5_driver).
1752  * @param[in] pci_dev
1753  *   PCI device information.
1754  *
1755  * @return
1756  *   0 on success, a negative errno value otherwise and rte_errno is set.
1757  */
1758 int
1759 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1760 		  struct rte_pci_device *pci_dev)
1761 {
1762 	struct ibv_device **ibv_list;
1763 	/*
1764 	 * Number of found IB Devices matching with requested PCI BDF.
1765 	 * nd != 1 means there are multiple IB devices over the same
1766 	 * PCI device and we have representors and master.
1767 	 */
1768 	unsigned int nd = 0;
1769 	/*
1770 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1771 	 * we have the single multiport IB device, and there may be
1772 	 * representors attached to some of found ports.
1773 	 */
1774 	unsigned int np = 0;
1775 	/*
1776 	 * Number of DPDK ethernet devices to Spawn - either over
1777 	 * multiple IB devices or multiple ports of single IB device.
1778 	 * Actually this is the number of iterations to spawn.
1779 	 */
1780 	unsigned int ns = 0;
1781 	/*
1782 	 * Bonding device
1783 	 *   < 0 - no bonding device (single one)
1784 	 *  >= 0 - bonding device (value is slave PF index)
1785 	 */
1786 	int bd = -1;
1787 	struct mlx5_dev_spawn_data *list = NULL;
1788 	struct mlx5_dev_config dev_config;
1789 	unsigned int dev_config_vf;
1790 	int ret;
1791 
1792 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1793 		mlx5_pmd_socket_init();
1794 	ret = mlx5_init_once();
1795 	if (ret) {
1796 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1797 			strerror(rte_errno));
1798 		return -rte_errno;
1799 	}
1800 	errno = 0;
1801 	ibv_list = mlx5_glue->get_device_list(&ret);
1802 	if (!ibv_list) {
1803 		rte_errno = errno ? errno : ENOSYS;
1804 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1805 		return -rte_errno;
1806 	}
1807 	/*
1808 	 * First scan the list of all Infiniband devices to find
1809 	 * matching ones, gathering into the list.
1810 	 */
1811 	struct ibv_device *ibv_match[ret + 1];
1812 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1813 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1814 	unsigned int i;
1815 
1816 	while (ret-- > 0) {
1817 		struct rte_pci_addr pci_addr;
1818 
1819 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1820 		bd = mlx5_device_bond_pci_match
1821 				(ibv_list[ret], pci_dev, nl_rdma);
1822 		if (bd >= 0) {
1823 			/*
1824 			 * Bonding device detected. Only one match is allowed,
1825 			 * the bonding is supported over multi-port IB device,
1826 			 * there should be no matches on representor PCI
1827 			 * functions or non VF LAG bonding devices with
1828 			 * specified address.
1829 			 */
1830 			if (nd) {
1831 				DRV_LOG(ERR,
1832 					"multiple PCI match on bonding device"
1833 					"\"%s\" found", ibv_list[ret]->name);
1834 				rte_errno = ENOENT;
1835 				ret = -rte_errno;
1836 				goto exit;
1837 			}
1838 			DRV_LOG(INFO, "PCI information matches for"
1839 				      " slave %d bonding device \"%s\"",
1840 				      bd, ibv_list[ret]->name);
1841 			ibv_match[nd++] = ibv_list[ret];
1842 			break;
1843 		}
1844 		if (mlx5_dev_to_pci_addr
1845 			(ibv_list[ret]->ibdev_path, &pci_addr))
1846 			continue;
1847 		if (pci_dev->addr.domain != pci_addr.domain ||
1848 		    pci_dev->addr.bus != pci_addr.bus ||
1849 		    pci_dev->addr.devid != pci_addr.devid ||
1850 		    pci_dev->addr.function != pci_addr.function)
1851 			continue;
1852 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1853 			ibv_list[ret]->name);
1854 		ibv_match[nd++] = ibv_list[ret];
1855 	}
1856 	ibv_match[nd] = NULL;
1857 	if (!nd) {
1858 		/* No device matches, just complain and bail out. */
1859 		DRV_LOG(WARNING,
1860 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1861 			" are kernel drivers loaded?",
1862 			pci_dev->addr.domain, pci_dev->addr.bus,
1863 			pci_dev->addr.devid, pci_dev->addr.function);
1864 		rte_errno = ENOENT;
1865 		ret = -rte_errno;
1866 		goto exit;
1867 	}
1868 	if (nd == 1) {
1869 		/*
1870 		 * Found single matching device may have multiple ports.
1871 		 * Each port may be representor, we have to check the port
1872 		 * number and check the representors existence.
1873 		 */
1874 		if (nl_rdma >= 0)
1875 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1876 		if (!np)
1877 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1878 					 " ports number", ibv_match[0]->name);
1879 		if (bd >= 0 && !np) {
1880 			DRV_LOG(ERR, "can not get ports"
1881 				     " for bonding device");
1882 			rte_errno = ENOENT;
1883 			ret = -rte_errno;
1884 			goto exit;
1885 		}
1886 	}
1887 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1888 	if (bd >= 0) {
1889 		/*
1890 		 * This may happen if there is VF LAG kernel support and
1891 		 * application is compiled with older rdma_core library.
1892 		 */
1893 		DRV_LOG(ERR,
1894 			"No kernel/verbs support for VF LAG bonding found.");
1895 		rte_errno = ENOTSUP;
1896 		ret = -rte_errno;
1897 		goto exit;
1898 	}
1899 #endif
1900 	/*
1901 	 * Now we can determine the maximal
1902 	 * amount of devices to be spawned.
1903 	 */
1904 	list = mlx5_malloc(MLX5_MEM_ZERO,
1905 			   sizeof(struct mlx5_dev_spawn_data) *
1906 			   (np ? np : nd),
1907 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1908 	if (!list) {
1909 		DRV_LOG(ERR, "spawn data array allocation failure");
1910 		rte_errno = ENOMEM;
1911 		ret = -rte_errno;
1912 		goto exit;
1913 	}
1914 	if (bd >= 0 || np > 1) {
1915 		/*
1916 		 * Single IB device with multiple ports found,
1917 		 * it may be E-Switch master device and representors.
1918 		 * We have to perform identification through the ports.
1919 		 */
1920 		MLX5_ASSERT(nl_rdma >= 0);
1921 		MLX5_ASSERT(ns == 0);
1922 		MLX5_ASSERT(nd == 1);
1923 		MLX5_ASSERT(np);
1924 		for (i = 1; i <= np; ++i) {
1925 			list[ns].max_port = np;
1926 			list[ns].phys_port = i;
1927 			list[ns].phys_dev = ibv_match[0];
1928 			list[ns].eth_dev = NULL;
1929 			list[ns].pci_dev = pci_dev;
1930 			list[ns].pf_bond = bd;
1931 			list[ns].ifindex = mlx5_nl_ifindex
1932 				(nl_rdma,
1933 				mlx5_os_get_dev_device_name
1934 						(list[ns].phys_dev), i);
1935 			if (!list[ns].ifindex) {
1936 				/*
1937 				 * No network interface index found for the
1938 				 * specified port, it means there is no
1939 				 * representor on this port. It's OK,
1940 				 * there can be disabled ports, for example
1941 				 * if sriov_numvfs < sriov_totalvfs.
1942 				 */
1943 				continue;
1944 			}
1945 			ret = -1;
1946 			if (nl_route >= 0)
1947 				ret = mlx5_nl_switch_info
1948 					       (nl_route,
1949 						list[ns].ifindex,
1950 						&list[ns].info);
1951 			if (ret || (!list[ns].info.representor &&
1952 				    !list[ns].info.master)) {
1953 				/*
1954 				 * We failed to recognize representors with
1955 				 * Netlink, let's try to perform the task
1956 				 * with sysfs.
1957 				 */
1958 				ret =  mlx5_sysfs_switch_info
1959 						(list[ns].ifindex,
1960 						 &list[ns].info);
1961 			}
1962 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1963 			if (!ret && bd >= 0) {
1964 				switch (list[ns].info.name_type) {
1965 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1966 					if (list[ns].info.port_name == bd)
1967 						ns++;
1968 					break;
1969 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1970 					/* Fallthrough */
1971 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1972 					if (list[ns].info.pf_num == bd)
1973 						ns++;
1974 					break;
1975 				default:
1976 					break;
1977 				}
1978 				continue;
1979 			}
1980 #endif
1981 			if (!ret && (list[ns].info.representor ^
1982 				     list[ns].info.master))
1983 				ns++;
1984 		}
1985 		if (!ns) {
1986 			DRV_LOG(ERR,
1987 				"unable to recognize master/representors"
1988 				" on the IB device with multiple ports");
1989 			rte_errno = ENOENT;
1990 			ret = -rte_errno;
1991 			goto exit;
1992 		}
1993 	} else {
1994 		/*
1995 		 * The existence of several matching entries (nd > 1) means
1996 		 * port representors have been instantiated. No existing Verbs
1997 		 * call nor sysfs entries can tell them apart, this can only
1998 		 * be done through Netlink calls assuming kernel drivers are
1999 		 * recent enough to support them.
2000 		 *
2001 		 * In the event of identification failure through Netlink,
2002 		 * try again through sysfs, then:
2003 		 *
2004 		 * 1. A single IB device matches (nd == 1) with single
2005 		 *    port (np=0/1) and is not a representor, assume
2006 		 *    no switch support.
2007 		 *
2008 		 * 2. Otherwise no safe assumptions can be made;
2009 		 *    complain louder and bail out.
2010 		 */
2011 		for (i = 0; i != nd; ++i) {
2012 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2013 			list[ns].max_port = 1;
2014 			list[ns].phys_port = 1;
2015 			list[ns].phys_dev = ibv_match[i];
2016 			list[ns].eth_dev = NULL;
2017 			list[ns].pci_dev = pci_dev;
2018 			list[ns].pf_bond = -1;
2019 			list[ns].ifindex = 0;
2020 			if (nl_rdma >= 0)
2021 				list[ns].ifindex = mlx5_nl_ifindex
2022 				(nl_rdma,
2023 				mlx5_os_get_dev_device_name
2024 						(list[ns].phys_dev), 1);
2025 			if (!list[ns].ifindex) {
2026 				char ifname[IF_NAMESIZE];
2027 
2028 				/*
2029 				 * Netlink failed, it may happen with old
2030 				 * ib_core kernel driver (before 4.16).
2031 				 * We can assume there is old driver because
2032 				 * here we are processing single ports IB
2033 				 * devices. Let's try sysfs to retrieve
2034 				 * the ifindex. The method works for
2035 				 * master device only.
2036 				 */
2037 				if (nd > 1) {
2038 					/*
2039 					 * Multiple devices found, assume
2040 					 * representors, can not distinguish
2041 					 * master/representor and retrieve
2042 					 * ifindex via sysfs.
2043 					 */
2044 					continue;
2045 				}
2046 				ret = mlx5_get_ifname_sysfs
2047 					(ibv_match[i]->ibdev_path, ifname);
2048 				if (!ret)
2049 					list[ns].ifindex =
2050 						if_nametoindex(ifname);
2051 				if (!list[ns].ifindex) {
2052 					/*
2053 					 * No network interface index found
2054 					 * for the specified device, it means
2055 					 * there it is neither representor
2056 					 * nor master.
2057 					 */
2058 					continue;
2059 				}
2060 			}
2061 			ret = -1;
2062 			if (nl_route >= 0)
2063 				ret = mlx5_nl_switch_info
2064 					       (nl_route,
2065 						list[ns].ifindex,
2066 						&list[ns].info);
2067 			if (ret || (!list[ns].info.representor &&
2068 				    !list[ns].info.master)) {
2069 				/*
2070 				 * We failed to recognize representors with
2071 				 * Netlink, let's try to perform the task
2072 				 * with sysfs.
2073 				 */
2074 				ret =  mlx5_sysfs_switch_info
2075 						(list[ns].ifindex,
2076 						 &list[ns].info);
2077 			}
2078 			if (!ret && (list[ns].info.representor ^
2079 				     list[ns].info.master)) {
2080 				ns++;
2081 			} else if ((nd == 1) &&
2082 				   !list[ns].info.representor &&
2083 				   !list[ns].info.master) {
2084 				/*
2085 				 * Single IB device with
2086 				 * one physical port and
2087 				 * attached network device.
2088 				 * May be SRIOV is not enabled
2089 				 * or there is no representors.
2090 				 */
2091 				DRV_LOG(INFO, "no E-Switch support detected");
2092 				ns++;
2093 				break;
2094 			}
2095 		}
2096 		if (!ns) {
2097 			DRV_LOG(ERR,
2098 				"unable to recognize master/representors"
2099 				" on the multiple IB devices");
2100 			rte_errno = ENOENT;
2101 			ret = -rte_errno;
2102 			goto exit;
2103 		}
2104 	}
2105 	MLX5_ASSERT(ns);
2106 	/*
2107 	 * Sort list to probe devices in natural order for users convenience
2108 	 * (i.e. master first, then representors from lowest to highest ID).
2109 	 */
2110 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2111 	/* Device specific configuration. */
2112 	switch (pci_dev->id.device_id) {
2113 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2114 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2115 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2116 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2117 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2118 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2119 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2120 		dev_config_vf = 1;
2121 		break;
2122 	default:
2123 		dev_config_vf = 0;
2124 		break;
2125 	}
2126 	for (i = 0; i != ns; ++i) {
2127 		uint32_t restore;
2128 
2129 		/* Default configuration. */
2130 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2131 		dev_config.vf = dev_config_vf;
2132 		dev_config.mps = MLX5_ARG_UNSET;
2133 		dev_config.dbnc = MLX5_ARG_UNSET;
2134 		dev_config.rx_vec_en = 1;
2135 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2136 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2137 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2138 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2139 		dev_config.vf_nl_en = 1;
2140 		dev_config.mr_ext_memseg_en = 1;
2141 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2142 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2143 		dev_config.dv_esw_en = 1;
2144 		dev_config.dv_flow_en = 1;
2145 		dev_config.decap_en = 1;
2146 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2147 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2148 						 &list[i],
2149 						 &dev_config);
2150 		if (!list[i].eth_dev) {
2151 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2152 				break;
2153 			/* Device is disabled or already spawned. Ignore it. */
2154 			continue;
2155 		}
2156 		restore = list[i].eth_dev->data->dev_flags;
2157 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2158 		/* Restore non-PCI flags cleared by the above call. */
2159 		list[i].eth_dev->data->dev_flags |= restore;
2160 		rte_eth_dev_probing_finish(list[i].eth_dev);
2161 	}
2162 	if (i != ns) {
2163 		DRV_LOG(ERR,
2164 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2165 			" encountering an error: %s",
2166 			pci_dev->addr.domain, pci_dev->addr.bus,
2167 			pci_dev->addr.devid, pci_dev->addr.function,
2168 			strerror(rte_errno));
2169 		ret = -rte_errno;
2170 		/* Roll back. */
2171 		while (i--) {
2172 			if (!list[i].eth_dev)
2173 				continue;
2174 			mlx5_dev_close(list[i].eth_dev);
2175 			/* mac_addrs must not be freed because in dev_private */
2176 			list[i].eth_dev->data->mac_addrs = NULL;
2177 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2178 		}
2179 		/* Restore original error. */
2180 		rte_errno = -ret;
2181 	} else {
2182 		ret = 0;
2183 	}
2184 exit:
2185 	/*
2186 	 * Do the routine cleanup:
2187 	 * - close opened Netlink sockets
2188 	 * - free allocated spawn data array
2189 	 * - free the Infiniband device list
2190 	 */
2191 	if (nl_rdma >= 0)
2192 		close(nl_rdma);
2193 	if (nl_route >= 0)
2194 		close(nl_route);
2195 	if (list)
2196 		mlx5_free(list);
2197 	MLX5_ASSERT(ibv_list);
2198 	mlx5_glue->free_device_list(ibv_list);
2199 	return ret;
2200 }
2201 
2202 static int
2203 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2204 {
2205 	char *env;
2206 	int value;
2207 
2208 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2209 	/* Get environment variable to store. */
2210 	env = getenv(MLX5_SHUT_UP_BF);
2211 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2212 	if (config->dbnc == MLX5_ARG_UNSET)
2213 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2214 	else
2215 		setenv(MLX5_SHUT_UP_BF,
2216 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2217 	return value;
2218 }
2219 
2220 static void
2221 mlx5_restore_doorbell_mapping_env(int value)
2222 {
2223 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2224 	/* Restore the original environment variable state. */
2225 	if (value == MLX5_ARG_UNSET)
2226 		unsetenv(MLX5_SHUT_UP_BF);
2227 	else
2228 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2229 }
2230 
2231 /**
2232  * Extract pdn of PD object using DV API.
2233  *
2234  * @param[in] pd
2235  *   Pointer to the verbs PD object.
2236  * @param[out] pdn
2237  *   Pointer to the PD object number variable.
2238  *
2239  * @return
2240  *   0 on success, error value otherwise.
2241  */
2242 int
2243 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2244 {
2245 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2246 	struct mlx5dv_obj obj;
2247 	struct mlx5dv_pd pd_info;
2248 	int ret = 0;
2249 
2250 	obj.pd.in = pd;
2251 	obj.pd.out = &pd_info;
2252 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2253 	if (ret) {
2254 		DRV_LOG(DEBUG, "Fail to get PD object info");
2255 		return ret;
2256 	}
2257 	*pdn = pd_info.pdn;
2258 	return 0;
2259 #else
2260 	(void)pd;
2261 	(void)pdn;
2262 	return -ENOTSUP;
2263 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2264 }
2265 
2266 /**
2267  * Function API to open IB device.
2268  *
2269  * This function calls the Linux glue APIs to open a device.
2270  *
2271  * @param[in] spawn
2272  *   Pointer to the IB device attributes (name, port, etc).
2273  * @param[out] config
2274  *   Pointer to device configuration structure.
2275  * @param[out] sh
2276  *   Pointer to shared context structure.
2277  *
2278  * @return
2279  *   0 on success, a positive error value otherwise.
2280  */
2281 int
2282 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2283 		     const struct mlx5_dev_config *config,
2284 		     struct mlx5_dev_ctx_shared *sh)
2285 {
2286 	int dbmap_env;
2287 	int err = 0;
2288 
2289 	sh->numa_node = spawn->pci_dev->device.numa_node;
2290 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2291 	/*
2292 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2293 	 * before the device creation. The rdma_core library
2294 	 * checks the variable at device creation and
2295 	 * stores the result internally.
2296 	 */
2297 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2298 	/* Try to open IB device with DV first, then usual Verbs. */
2299 	errno = 0;
2300 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2301 	if (sh->ctx) {
2302 		sh->devx = 1;
2303 		DRV_LOG(DEBUG, "DevX is supported");
2304 		/* The device is created, no need for environment. */
2305 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2306 	} else {
2307 		/* The environment variable is still configured. */
2308 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2309 		err = errno ? errno : ENODEV;
2310 		/*
2311 		 * The environment variable is not needed anymore,
2312 		 * all device creation attempts are completed.
2313 		 */
2314 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2315 		if (!sh->ctx)
2316 			return err;
2317 		DRV_LOG(DEBUG, "DevX is NOT supported");
2318 		err = 0;
2319 	}
2320 	if (!err && sh->ctx) {
2321 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2322 		mlx5_glue->dv_set_context_attr(sh->ctx,
2323 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2324 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2325 				.alloc = &mlx5_alloc_verbs_buf,
2326 				.free = &mlx5_free_verbs_buf,
2327 				.data = sh,
2328 			}));
2329 	}
2330 	return err;
2331 }
2332 
2333 /**
2334  * Install shared asynchronous device events handler.
2335  * This function is implemented to support event sharing
2336  * between multiple ports of single IB device.
2337  *
2338  * @param sh
2339  *   Pointer to mlx5_dev_ctx_shared object.
2340  */
2341 void
2342 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2343 {
2344 	int ret;
2345 	int flags;
2346 
2347 	sh->intr_handle.fd = -1;
2348 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2349 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2350 		    F_SETFL, flags | O_NONBLOCK);
2351 	if (ret) {
2352 		DRV_LOG(INFO, "failed to change file descriptor async event"
2353 			" queue");
2354 	} else {
2355 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2356 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2357 		if (rte_intr_callback_register(&sh->intr_handle,
2358 					mlx5_dev_interrupt_handler, sh)) {
2359 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2360 			sh->intr_handle.fd = -1;
2361 		}
2362 	}
2363 	if (sh->devx) {
2364 #ifdef HAVE_IBV_DEVX_ASYNC
2365 		sh->intr_handle_devx.fd = -1;
2366 		sh->devx_comp =
2367 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2368 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2369 		if (!devx_comp) {
2370 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2371 			return;
2372 		}
2373 		flags = fcntl(devx_comp->fd, F_GETFL);
2374 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2375 		if (ret) {
2376 			DRV_LOG(INFO, "failed to change file descriptor"
2377 				" devx comp");
2378 			return;
2379 		}
2380 		sh->intr_handle_devx.fd = devx_comp->fd;
2381 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2382 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2383 					mlx5_dev_interrupt_handler_devx, sh)) {
2384 			DRV_LOG(INFO, "Fail to install the devx shared"
2385 				" interrupt.");
2386 			sh->intr_handle_devx.fd = -1;
2387 		}
2388 #endif /* HAVE_IBV_DEVX_ASYNC */
2389 	}
2390 }
2391 
2392 /**
2393  * Uninstall 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 dev
2398  *   Pointer to mlx5_dev_ctx_shared object.
2399  */
2400 void
2401 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2402 {
2403 	if (sh->intr_handle.fd >= 0)
2404 		mlx5_intr_callback_unregister(&sh->intr_handle,
2405 					      mlx5_dev_interrupt_handler, sh);
2406 #ifdef HAVE_IBV_DEVX_ASYNC
2407 	if (sh->intr_handle_devx.fd >= 0)
2408 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2409 				  mlx5_dev_interrupt_handler_devx, sh);
2410 	if (sh->devx_comp)
2411 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2412 #endif
2413 }
2414 
2415 /**
2416  * Read statistics by a named counter.
2417  *
2418  * @param[in] priv
2419  *   Pointer to the private device data structure.
2420  * @param[in] ctr_name
2421  *   Pointer to the name of the statistic counter to read
2422  * @param[out] stat
2423  *   Pointer to read statistic value.
2424  * @return
2425  *   0 on success and stat is valud, 1 if failed to read the value
2426  *   rte_errno is set.
2427  *
2428  */
2429 int
2430 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2431 		      uint64_t *stat)
2432 {
2433 	int fd;
2434 
2435 	if (priv->sh) {
2436 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2437 		      priv->sh->ibdev_path,
2438 		      priv->dev_port,
2439 		      ctr_name);
2440 		fd = open(path, O_RDONLY);
2441 		/*
2442 		 * in switchdev the file location is not per port
2443 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2444 		 */
2445 		if (fd == -1) {
2446 			MKSTR(path1, "%s/hw_counters/%s",
2447 			      priv->sh->ibdev_path,
2448 			      ctr_name);
2449 			fd = open(path1, O_RDONLY);
2450 		}
2451 		if (fd != -1) {
2452 			char buf[21] = {'\0'};
2453 			ssize_t n = read(fd, buf, sizeof(buf));
2454 
2455 			close(fd);
2456 			if (n != -1) {
2457 				*stat = strtoull(buf, NULL, 10);
2458 				return 0;
2459 			}
2460 		}
2461 	}
2462 	*stat = 0;
2463 	return 1;
2464 }
2465 
2466 /**
2467  * Set the reg_mr and dereg_mr call backs
2468  *
2469  * @param reg_mr_cb[out]
2470  *   Pointer to reg_mr func
2471  * @param dereg_mr_cb[out]
2472  *   Pointer to dereg_mr func
2473  *
2474  */
2475 void
2476 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2477 		      mlx5_dereg_mr_t *dereg_mr_cb)
2478 {
2479 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2480 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2481 }
2482 
2483 /**
2484  * Remove a MAC address from device
2485  *
2486  * @param dev
2487  *   Pointer to Ethernet device structure.
2488  * @param index
2489  *   MAC address index.
2490  */
2491 void
2492 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2493 {
2494 	struct mlx5_priv *priv = dev->data->dev_private;
2495 	const int vf = priv->config.vf;
2496 
2497 	if (vf)
2498 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2499 					mlx5_ifindex(dev), priv->mac_own,
2500 					&dev->data->mac_addrs[index], index);
2501 }
2502 
2503 /**
2504  * Adds a MAC address to the device
2505  *
2506  * @param dev
2507  *   Pointer to Ethernet device structure.
2508  * @param mac_addr
2509  *   MAC address to register.
2510  * @param index
2511  *   MAC address index.
2512  *
2513  * @return
2514  *   0 on success, a negative errno value otherwise
2515  */
2516 int
2517 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2518 		     uint32_t index)
2519 {
2520 	struct mlx5_priv *priv = dev->data->dev_private;
2521 	const int vf = priv->config.vf;
2522 	int ret = 0;
2523 
2524 	if (vf)
2525 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2526 					   mlx5_ifindex(dev), priv->mac_own,
2527 					   mac, index);
2528 	return ret;
2529 }
2530 
2531 /**
2532  * Modify a VF MAC address
2533  *
2534  * @param priv
2535  *   Pointer to device private data.
2536  * @param mac_addr
2537  *   MAC address to modify into.
2538  * @param iface_idx
2539  *   Net device interface index
2540  * @param vf_index
2541  *   VF index
2542  *
2543  * @return
2544  *   0 on success, a negative errno value otherwise
2545  */
2546 int
2547 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2548 			   unsigned int iface_idx,
2549 			   struct rte_ether_addr *mac_addr,
2550 			   int vf_index)
2551 {
2552 	return mlx5_nl_vf_mac_addr_modify
2553 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2554 }
2555 
2556 /**
2557  * Set device promiscuous mode
2558  *
2559  * @param dev
2560  *   Pointer to Ethernet device structure.
2561  * @param enable
2562  *   0 - promiscuous is disabled, otherwise - enabled
2563  *
2564  * @return
2565  *   0 on success, a negative error value otherwise
2566  */
2567 int
2568 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2569 {
2570 	struct mlx5_priv *priv = dev->data->dev_private;
2571 
2572 	return mlx5_nl_promisc(priv->nl_socket_route,
2573 			       mlx5_ifindex(dev), !!enable);
2574 }
2575 
2576 /**
2577  * Set device promiscuous mode
2578  *
2579  * @param dev
2580  *   Pointer to Ethernet device structure.
2581  * @param enable
2582  *   0 - all multicase is disabled, otherwise - enabled
2583  *
2584  * @return
2585  *   0 on success, a negative error value otherwise
2586  */
2587 int
2588 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2589 {
2590 	struct mlx5_priv *priv = dev->data->dev_private;
2591 
2592 	return mlx5_nl_allmulti(priv->nl_socket_route,
2593 				mlx5_ifindex(dev), !!enable);
2594 }
2595 
2596 /**
2597  * Flush device MAC addresses
2598  *
2599  * @param dev
2600  *   Pointer to Ethernet device structure.
2601  *
2602  */
2603 void
2604 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2605 {
2606 	struct mlx5_priv *priv = dev->data->dev_private;
2607 
2608 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2609 			       dev->data->mac_addrs,
2610 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2611 }
2612