xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 37cd4501e873f30c69f738c3efe4730e7c95d60a)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2020 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <linux/rtnetlink.h>
14 #include <linux/sockios.h>
15 #include <linux/ethtool.h>
16 #include <fcntl.h>
17 
18 #include <rte_malloc.h>
19 #include <rte_ethdev_driver.h>
20 #include <rte_ethdev_pci.h>
21 #include <rte_pci.h>
22 #include <rte_bus_pci.h>
23 #include <rte_common.h>
24 #include <rte_kvargs.h>
25 #include <rte_rwlock.h>
26 #include <rte_spinlock.h>
27 #include <rte_string_fns.h>
28 #include <rte_alarm.h>
29 #include <rte_eal_paging.h>
30 
31 #include <mlx5_glue.h>
32 #include <mlx5_devx_cmds.h>
33 #include <mlx5_common.h>
34 #include <mlx5_common_mp.h>
35 #include <mlx5_common_mr.h>
36 #include <mlx5_malloc.h>
37 
38 #include "mlx5_defs.h"
39 #include "mlx5.h"
40 #include "mlx5_common_os.h"
41 #include "mlx5_utils.h"
42 #include "mlx5_rxtx.h"
43 #include "mlx5_autoconf.h"
44 #include "mlx5_mr.h"
45 #include "mlx5_flow.h"
46 #include "rte_pmd_mlx5.h"
47 #include "mlx5_verbs.h"
48 #include "mlx5_nl.h"
49 #include "mlx5_devx.h"
50 
51 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
52 
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /**
71  * Set the completion channel file descriptor interrupt as non-blocking.
72  *
73  * @param[in] rxq_obj
74  *   Pointer to RQ channel object, which includes the channel fd
75  *
76  * @param[out] fd
77  *   The file descriptor (representing the intetrrupt) used in this channel.
78  *
79  * @return
80  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
81  */
82 int
83 mlx5_os_set_nonblock_channel_fd(int fd)
84 {
85 	int flags;
86 
87 	flags = fcntl(fd, F_GETFL);
88 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
89 }
90 
91 /**
92  * Get mlx5 device attributes. The glue function query_device_ex() is called
93  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
94  * device attributes from the glue out parameter.
95  *
96  * @param dev
97  *   Pointer to ibv context.
98  *
99  * @param device_attr
100  *   Pointer to mlx5 device attributes.
101  *
102  * @return
103  *   0 on success, non zero error number otherwise
104  */
105 int
106 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
107 {
108 	int err;
109 	struct ibv_device_attr_ex attr_ex;
110 	memset(device_attr, 0, sizeof(*device_attr));
111 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
112 	if (err)
113 		return err;
114 
115 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
116 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
117 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
118 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
119 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
120 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
121 	device_attr->max_rwq_indirection_table_size =
122 		attr_ex.rss_caps.max_rwq_indirection_table_size;
123 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
124 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
125 
126 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
127 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
128 	if (err)
129 		return err;
130 
131 	device_attr->flags = dv_attr.flags;
132 	device_attr->comp_mask = dv_attr.comp_mask;
133 #ifdef HAVE_IBV_MLX5_MOD_SWP
134 	device_attr->sw_parsing_offloads =
135 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
136 #endif
137 	device_attr->min_single_stride_log_num_of_bytes =
138 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
139 	device_attr->max_single_stride_log_num_of_bytes =
140 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
141 	device_attr->min_single_wqe_log_num_of_strides =
142 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
143 	device_attr->max_single_wqe_log_num_of_strides =
144 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
145 	device_attr->stride_supported_qpts =
146 		dv_attr.striding_rq_caps.supported_qpts;
147 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
148 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
149 #endif
150 
151 	return err;
152 }
153 
154 /**
155  * Verbs callback to allocate a memory. This function should allocate the space
156  * according to the size provided residing inside a huge page.
157  * Please note that all allocation must respect the alignment from libmlx5
158  * (i.e. currently rte_mem_page_size()).
159  *
160  * @param[in] size
161  *   The size in bytes of the memory to allocate.
162  * @param[in] data
163  *   A pointer to the callback data.
164  *
165  * @return
166  *   Allocated buffer, NULL otherwise and rte_errno is set.
167  */
168 static void *
169 mlx5_alloc_verbs_buf(size_t size, void *data)
170 {
171 	struct mlx5_priv *priv = data;
172 	void *ret;
173 	unsigned int socket = SOCKET_ID_ANY;
174 	size_t alignment = rte_mem_page_size();
175 	if (alignment == (size_t)-1) {
176 		DRV_LOG(ERR, "Failed to get mem page size");
177 		rte_errno = ENOMEM;
178 		return NULL;
179 	}
180 
181 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
182 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
183 
184 		socket = ctrl->socket;
185 	} else if (priv->verbs_alloc_ctx.type ==
186 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
187 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
188 
189 		socket = ctrl->socket;
190 	}
191 	MLX5_ASSERT(data != NULL);
192 	ret = mlx5_malloc(0, size, alignment, socket);
193 	if (!ret && size)
194 		rte_errno = ENOMEM;
195 	return ret;
196 }
197 
198 /**
199  * Verbs callback to free a memory.
200  *
201  * @param[in] ptr
202  *   A pointer to the memory to free.
203  * @param[in] data
204  *   A pointer to the callback data.
205  */
206 static void
207 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
208 {
209 	MLX5_ASSERT(data != NULL);
210 	mlx5_free(ptr);
211 }
212 
213 /**
214  * Initialize DR related data within private structure.
215  * Routine checks the reference counter and does actual
216  * resources creation/initialization only if counter is zero.
217  *
218  * @param[in] priv
219  *   Pointer to the private device data structure.
220  *
221  * @return
222  *   Zero on success, positive error code otherwise.
223  */
224 static int
225 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
226 {
227 	struct mlx5_dev_ctx_shared *sh = priv->sh;
228 	char s[MLX5_HLIST_NAMESIZE];
229 	int err;
230 
231 	MLX5_ASSERT(sh && sh->refcnt);
232 	if (sh->refcnt > 1)
233 		return 0;
234 	err = mlx5_alloc_table_hash_list(priv);
235 	if (err)
236 		return err;
237 	/* Create tags hash list table. */
238 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
239 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
240 	if (!sh->tag_table) {
241 		DRV_LOG(ERR, "tags with hash creation failed.");
242 		err = ENOMEM;
243 		goto error;
244 	}
245 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
246 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ);
247 	if (!sh->modify_cmds) {
248 		DRV_LOG(ERR, "hdr modify hash creation failed");
249 		err = ENOMEM;
250 		goto error;
251 	}
252 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
253 	sh->encaps_decaps = mlx5_hlist_create(s,
254 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ);
255 	if (!sh->encaps_decaps) {
256 		DRV_LOG(ERR, "encap decap hash creation failed");
257 		err = ENOMEM;
258 		goto error;
259 	}
260 #ifdef HAVE_MLX5DV_DR
261 	void *domain;
262 
263 	/* Reference counter is zero, we should initialize structures. */
264 	domain = mlx5_glue->dr_create_domain(sh->ctx,
265 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
266 	if (!domain) {
267 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
268 		err = errno;
269 		goto error;
270 	}
271 	sh->rx_domain = domain;
272 	domain = mlx5_glue->dr_create_domain(sh->ctx,
273 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
274 	if (!domain) {
275 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
276 		err = errno;
277 		goto error;
278 	}
279 	pthread_mutex_init(&sh->dv_mutex, NULL);
280 	sh->tx_domain = domain;
281 #ifdef HAVE_MLX5DV_DR_ESWITCH
282 	if (priv->config.dv_esw_en) {
283 		domain  = mlx5_glue->dr_create_domain
284 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
285 		if (!domain) {
286 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
287 			err = errno;
288 			goto error;
289 		}
290 		sh->fdb_domain = domain;
291 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
292 	}
293 #endif
294 	if (!sh->tunnel_hub)
295 		err = mlx5_alloc_tunnel_hub(sh);
296 	if (err) {
297 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
298 		goto error;
299 	}
300 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
301 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
302 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
303 		if (sh->fdb_domain)
304 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
305 	}
306 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
307 #endif /* HAVE_MLX5DV_DR */
308 	return 0;
309 error:
310 	/* Rollback the created objects. */
311 	if (sh->rx_domain) {
312 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
313 		sh->rx_domain = NULL;
314 	}
315 	if (sh->tx_domain) {
316 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
317 		sh->tx_domain = NULL;
318 	}
319 	if (sh->fdb_domain) {
320 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
321 		sh->fdb_domain = NULL;
322 	}
323 	if (sh->esw_drop_action) {
324 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
325 		sh->esw_drop_action = NULL;
326 	}
327 	if (sh->pop_vlan_action) {
328 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
329 		sh->pop_vlan_action = NULL;
330 	}
331 	if (sh->encaps_decaps) {
332 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
333 		sh->encaps_decaps = NULL;
334 	}
335 	if (sh->modify_cmds) {
336 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
337 		sh->modify_cmds = NULL;
338 	}
339 	if (sh->tag_table) {
340 		/* tags should be destroyed with flow before. */
341 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
342 		sh->tag_table = NULL;
343 	}
344 	if (sh->tunnel_hub) {
345 		mlx5_release_tunnel_hub(sh, priv->dev_port);
346 		sh->tunnel_hub = NULL;
347 	}
348 	mlx5_free_table_hash_list(priv);
349 	return err;
350 }
351 
352 /**
353  * Destroy DR related data within private structure.
354  *
355  * @param[in] priv
356  *   Pointer to the private device data structure.
357  */
358 void
359 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
360 {
361 	struct mlx5_dev_ctx_shared *sh = priv->sh;
362 
363 	MLX5_ASSERT(sh && sh->refcnt);
364 	if (sh->refcnt > 1)
365 		return;
366 #ifdef HAVE_MLX5DV_DR
367 	if (sh->rx_domain) {
368 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
369 		sh->rx_domain = NULL;
370 	}
371 	if (sh->tx_domain) {
372 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
373 		sh->tx_domain = NULL;
374 	}
375 #ifdef HAVE_MLX5DV_DR_ESWITCH
376 	if (sh->fdb_domain) {
377 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
378 		sh->fdb_domain = NULL;
379 	}
380 	if (sh->esw_drop_action) {
381 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
382 		sh->esw_drop_action = NULL;
383 	}
384 #endif
385 	if (sh->pop_vlan_action) {
386 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
387 		sh->pop_vlan_action = NULL;
388 	}
389 	pthread_mutex_destroy(&sh->dv_mutex);
390 #endif /* HAVE_MLX5DV_DR */
391 	if (sh->encaps_decaps) {
392 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
393 		sh->encaps_decaps = NULL;
394 	}
395 	if (sh->modify_cmds) {
396 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
397 		sh->modify_cmds = NULL;
398 	}
399 	if (sh->tag_table) {
400 		/* tags should be destroyed with flow before. */
401 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
402 		sh->tag_table = NULL;
403 	}
404 	if (sh->tunnel_hub) {
405 		mlx5_release_tunnel_hub(sh, priv->dev_port);
406 		sh->tunnel_hub = NULL;
407 	}
408 	mlx5_free_table_hash_list(priv);
409 }
410 
411 /**
412  * Initialize shared data between primary and secondary process.
413  *
414  * A memzone is reserved by primary process and secondary processes attach to
415  * the memzone.
416  *
417  * @return
418  *   0 on success, a negative errno value otherwise and rte_errno is set.
419  */
420 static int
421 mlx5_init_shared_data(void)
422 {
423 	const struct rte_memzone *mz;
424 	int ret = 0;
425 
426 	rte_spinlock_lock(&mlx5_shared_data_lock);
427 	if (mlx5_shared_data == NULL) {
428 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
429 			/* Allocate shared memory. */
430 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
431 						 sizeof(*mlx5_shared_data),
432 						 SOCKET_ID_ANY, 0);
433 			if (mz == NULL) {
434 				DRV_LOG(ERR,
435 					"Cannot allocate mlx5 shared data");
436 				ret = -rte_errno;
437 				goto error;
438 			}
439 			mlx5_shared_data = mz->addr;
440 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
441 			rte_spinlock_init(&mlx5_shared_data->lock);
442 		} else {
443 			/* Lookup allocated shared memory. */
444 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
445 			if (mz == NULL) {
446 				DRV_LOG(ERR,
447 					"Cannot attach mlx5 shared data");
448 				ret = -rte_errno;
449 				goto error;
450 			}
451 			mlx5_shared_data = mz->addr;
452 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
453 		}
454 	}
455 error:
456 	rte_spinlock_unlock(&mlx5_shared_data_lock);
457 	return ret;
458 }
459 
460 /**
461  * PMD global initialization.
462  *
463  * Independent from individual device, this function initializes global
464  * per-PMD data structures distinguishing primary and secondary processes.
465  * Hence, each initialization is called once per a process.
466  *
467  * @return
468  *   0 on success, a negative errno value otherwise and rte_errno is set.
469  */
470 static int
471 mlx5_init_once(void)
472 {
473 	struct mlx5_shared_data *sd;
474 	struct mlx5_local_data *ld = &mlx5_local_data;
475 	int ret = 0;
476 
477 	if (mlx5_init_shared_data())
478 		return -rte_errno;
479 	sd = mlx5_shared_data;
480 	MLX5_ASSERT(sd);
481 	rte_spinlock_lock(&sd->lock);
482 	switch (rte_eal_process_type()) {
483 	case RTE_PROC_PRIMARY:
484 		if (sd->init_done)
485 			break;
486 		LIST_INIT(&sd->mem_event_cb_list);
487 		rte_rwlock_init(&sd->mem_event_rwlock);
488 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
489 						mlx5_mr_mem_event_cb, NULL);
490 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
491 					   mlx5_mp_os_primary_handle);
492 		if (ret)
493 			goto out;
494 		sd->init_done = true;
495 		break;
496 	case RTE_PROC_SECONDARY:
497 		if (ld->init_done)
498 			break;
499 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
500 					     mlx5_mp_os_secondary_handle);
501 		if (ret)
502 			goto out;
503 		++sd->secondary_cnt;
504 		ld->init_done = true;
505 		break;
506 	default:
507 		break;
508 	}
509 out:
510 	rte_spinlock_unlock(&sd->lock);
511 	return ret;
512 }
513 
514 /**
515  * Create the Tx queue DevX/Verbs object.
516  *
517  * @param dev
518  *   Pointer to Ethernet device.
519  * @param idx
520  *   Queue index in DPDK Tx queue array.
521  *
522  * @return
523  *   0 on success, a negative errno value otherwise and rte_errno is set.
524  */
525 static int
526 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
527 {
528 	struct mlx5_priv *priv = dev->data->dev_private;
529 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
530 	struct mlx5_txq_ctrl *txq_ctrl =
531 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
532 
533 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
534 		return mlx5_txq_devx_obj_new(dev, idx);
535 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
536 	if (!priv->config.dv_esw_en)
537 		return mlx5_txq_devx_obj_new(dev, idx);
538 #endif
539 	return mlx5_txq_ibv_obj_new(dev, idx);
540 }
541 
542 /**
543  * Release an Tx DevX/verbs queue object.
544  *
545  * @param txq_obj
546  *   DevX/Verbs Tx queue object.
547  */
548 static void
549 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
550 {
551 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
552 		mlx5_txq_devx_obj_release(txq_obj);
553 		return;
554 	}
555 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
556 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
557 		mlx5_txq_devx_obj_release(txq_obj);
558 		return;
559 	}
560 #endif
561 	mlx5_txq_ibv_obj_release(txq_obj);
562 }
563 
564 /**
565  * DV flow counter mode detect and config.
566  *
567  * @param dev
568  *   Pointer to rte_eth_dev structure.
569  *
570  */
571 static void
572 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
573 {
574 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
575 	struct mlx5_priv *priv = dev->data->dev_private;
576 	struct mlx5_dev_ctx_shared *sh = priv->sh;
577 	bool fallback;
578 
579 #ifndef HAVE_IBV_DEVX_ASYNC
580 	fallback = true;
581 #else
582 	fallback = false;
583 	if (!priv->config.devx || !priv->config.dv_flow_en ||
584 	    !priv->config.hca_attr.flow_counters_dump ||
585 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
586 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
587 		fallback = true;
588 #endif
589 	if (fallback)
590 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
591 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
592 			priv->config.hca_attr.flow_counters_dump,
593 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
594 	/* Initialize fallback mode only on the port initializes sh. */
595 	if (sh->refcnt == 1)
596 		sh->cmng.counter_fallback = fallback;
597 	else if (fallback != sh->cmng.counter_fallback)
598 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
599 			"with others:%d.", PORT_ID(priv), fallback);
600 #endif
601 }
602 
603 /**
604  * Spawn an Ethernet device from Verbs information.
605  *
606  * @param dpdk_dev
607  *   Backing DPDK device.
608  * @param spawn
609  *   Verbs device parameters (name, port, switch_info) to spawn.
610  * @param config
611  *   Device configuration parameters.
612  *
613  * @return
614  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
615  *   is set. The following errors are defined:
616  *
617  *   EBUSY: device is not supposed to be spawned.
618  *   EEXIST: device is already spawned
619  */
620 static struct rte_eth_dev *
621 mlx5_dev_spawn(struct rte_device *dpdk_dev,
622 	       struct mlx5_dev_spawn_data *spawn,
623 	       struct mlx5_dev_config *config)
624 {
625 	const struct mlx5_switch_info *switch_info = &spawn->info;
626 	struct mlx5_dev_ctx_shared *sh = NULL;
627 	struct ibv_port_attr port_attr;
628 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
629 	struct rte_eth_dev *eth_dev = NULL;
630 	struct mlx5_priv *priv = NULL;
631 	int err = 0;
632 	unsigned int hw_padding = 0;
633 	unsigned int mps;
634 	unsigned int cqe_comp;
635 	unsigned int cqe_pad = 0;
636 	unsigned int tunnel_en = 0;
637 	unsigned int mpls_en = 0;
638 	unsigned int swp = 0;
639 	unsigned int mprq = 0;
640 	unsigned int mprq_min_stride_size_n = 0;
641 	unsigned int mprq_max_stride_size_n = 0;
642 	unsigned int mprq_min_stride_num_n = 0;
643 	unsigned int mprq_max_stride_num_n = 0;
644 	struct rte_ether_addr mac;
645 	char name[RTE_ETH_NAME_MAX_LEN];
646 	int own_domain_id = 0;
647 	uint16_t port_id;
648 	unsigned int i;
649 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
650 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
651 #endif
652 
653 	/* Determine if this port representor is supposed to be spawned. */
654 	if (switch_info->representor && dpdk_dev->devargs) {
655 		struct rte_eth_devargs eth_da;
656 
657 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
658 		if (err) {
659 			rte_errno = -err;
660 			DRV_LOG(ERR, "failed to process device arguments: %s",
661 				strerror(rte_errno));
662 			return NULL;
663 		}
664 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
665 			if (eth_da.representor_ports[i] ==
666 			    (uint16_t)switch_info->port_name)
667 				break;
668 		if (i == eth_da.nb_representor_ports) {
669 			rte_errno = EBUSY;
670 			return NULL;
671 		}
672 	}
673 	/* Build device name. */
674 	if (spawn->pf_bond <  0) {
675 		/* Single device. */
676 		if (!switch_info->representor)
677 			strlcpy(name, dpdk_dev->name, sizeof(name));
678 		else
679 			snprintf(name, sizeof(name), "%s_representor_%u",
680 				 dpdk_dev->name, switch_info->port_name);
681 	} else {
682 		/* Bonding device. */
683 		if (!switch_info->representor)
684 			snprintf(name, sizeof(name), "%s_%s",
685 				 dpdk_dev->name,
686 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
687 		else
688 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
689 				 dpdk_dev->name,
690 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
691 				 switch_info->port_name);
692 	}
693 	/* check if the device is already spawned */
694 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
695 		rte_errno = EEXIST;
696 		return NULL;
697 	}
698 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
699 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
700 		struct mlx5_mp_id mp_id;
701 
702 		eth_dev = rte_eth_dev_attach_secondary(name);
703 		if (eth_dev == NULL) {
704 			DRV_LOG(ERR, "can not attach rte ethdev");
705 			rte_errno = ENOMEM;
706 			return NULL;
707 		}
708 		eth_dev->device = dpdk_dev;
709 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
710 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
711 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
712 		err = mlx5_proc_priv_init(eth_dev);
713 		if (err)
714 			return NULL;
715 		mp_id.port_id = eth_dev->data->port_id;
716 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
717 		/* Receive command fd from primary process */
718 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
719 		if (err < 0)
720 			goto err_secondary;
721 		/* Remap UAR for Tx queues. */
722 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
723 		if (err)
724 			goto err_secondary;
725 		/*
726 		 * Ethdev pointer is still required as input since
727 		 * the primary device is not accessible from the
728 		 * secondary process.
729 		 */
730 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
731 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
732 		return eth_dev;
733 err_secondary:
734 		mlx5_dev_close(eth_dev);
735 		return NULL;
736 	}
737 	/*
738 	 * Some parameters ("tx_db_nc" in particularly) are needed in
739 	 * advance to create dv/verbs device context. We proceed the
740 	 * devargs here to get ones, and later proceed devargs again
741 	 * to override some hardware settings.
742 	 */
743 	err = mlx5_args(config, dpdk_dev->devargs);
744 	if (err) {
745 		err = rte_errno;
746 		DRV_LOG(ERR, "failed to process device arguments: %s",
747 			strerror(rte_errno));
748 		goto error;
749 	}
750 	if (config->dv_miss_info) {
751 		if (switch_info->master || switch_info->representor)
752 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
753 	}
754 	mlx5_malloc_mem_select(config->sys_mem_en);
755 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
756 	if (!sh)
757 		return NULL;
758 	config->devx = sh->devx;
759 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
760 	config->dest_tir = 1;
761 #endif
762 #ifdef HAVE_IBV_MLX5_MOD_SWP
763 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
764 #endif
765 	/*
766 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
767 	 * as all ConnectX-5 devices.
768 	 */
769 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
770 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
771 #endif
772 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
773 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
774 #endif
775 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
776 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
777 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
778 			DRV_LOG(DEBUG, "enhanced MPW is supported");
779 			mps = MLX5_MPW_ENHANCED;
780 		} else {
781 			DRV_LOG(DEBUG, "MPW is supported");
782 			mps = MLX5_MPW;
783 		}
784 	} else {
785 		DRV_LOG(DEBUG, "MPW isn't supported");
786 		mps = MLX5_MPW_DISABLED;
787 	}
788 #ifdef HAVE_IBV_MLX5_MOD_SWP
789 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
790 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
791 	DRV_LOG(DEBUG, "SWP support: %u", swp);
792 #endif
793 	config->swp = !!swp;
794 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
795 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
796 		struct mlx5dv_striding_rq_caps mprq_caps =
797 			dv_attr.striding_rq_caps;
798 
799 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
800 			mprq_caps.min_single_stride_log_num_of_bytes);
801 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
802 			mprq_caps.max_single_stride_log_num_of_bytes);
803 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
804 			mprq_caps.min_single_wqe_log_num_of_strides);
805 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
806 			mprq_caps.max_single_wqe_log_num_of_strides);
807 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
808 			mprq_caps.supported_qpts);
809 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
810 		mprq = 1;
811 		mprq_min_stride_size_n =
812 			mprq_caps.min_single_stride_log_num_of_bytes;
813 		mprq_max_stride_size_n =
814 			mprq_caps.max_single_stride_log_num_of_bytes;
815 		mprq_min_stride_num_n =
816 			mprq_caps.min_single_wqe_log_num_of_strides;
817 		mprq_max_stride_num_n =
818 			mprq_caps.max_single_wqe_log_num_of_strides;
819 	}
820 #endif
821 	if (RTE_CACHE_LINE_SIZE == 128 &&
822 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
823 		cqe_comp = 0;
824 	else
825 		cqe_comp = 1;
826 	config->cqe_comp = cqe_comp;
827 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
828 	/* Whether device supports 128B Rx CQE padding. */
829 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
830 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
831 #endif
832 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
833 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
834 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
835 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
836 			     (dv_attr.tunnel_offloads_caps &
837 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
838 			     (dv_attr.tunnel_offloads_caps &
839 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
840 	}
841 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
842 		tunnel_en ? "" : "not ");
843 #else
844 	DRV_LOG(WARNING,
845 		"tunnel offloading disabled due to old OFED/rdma-core version");
846 #endif
847 	config->tunnel_en = tunnel_en;
848 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
849 	mpls_en = ((dv_attr.tunnel_offloads_caps &
850 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
851 		   (dv_attr.tunnel_offloads_caps &
852 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
853 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
854 		mpls_en ? "" : "not ");
855 #else
856 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
857 		" old OFED/rdma-core version or firmware configuration");
858 #endif
859 	config->mpls_en = mpls_en;
860 	/* Check port status. */
861 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
862 	if (err) {
863 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
864 		goto error;
865 	}
866 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
867 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
868 		err = EINVAL;
869 		goto error;
870 	}
871 	if (port_attr.state != IBV_PORT_ACTIVE)
872 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
873 			mlx5_glue->port_state_str(port_attr.state),
874 			port_attr.state);
875 	/* Allocate private eth device data. */
876 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
877 			   sizeof(*priv),
878 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
879 	if (priv == NULL) {
880 		DRV_LOG(ERR, "priv allocation failure");
881 		err = ENOMEM;
882 		goto error;
883 	}
884 	priv->sh = sh;
885 	priv->dev_port = spawn->phys_port;
886 	priv->pci_dev = spawn->pci_dev;
887 	priv->mtu = RTE_ETHER_MTU;
888 	priv->mp_id.port_id = port_id;
889 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
890 	/* Some internal functions rely on Netlink sockets, open them now. */
891 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
892 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
893 	priv->representor = !!switch_info->representor;
894 	priv->master = !!switch_info->master;
895 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
896 	priv->vport_meta_tag = 0;
897 	priv->vport_meta_mask = 0;
898 	priv->pf_bond = spawn->pf_bond;
899 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
900 	/*
901 	 * The DevX port query API is implemented. E-Switch may use
902 	 * either vport or reg_c[0] metadata register to match on
903 	 * vport index. The engaged part of metadata register is
904 	 * defined by mask.
905 	 */
906 	if (switch_info->representor || switch_info->master) {
907 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
908 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
909 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
910 						 &devx_port);
911 		if (err) {
912 			DRV_LOG(WARNING,
913 				"can't query devx port %d on device %s",
914 				spawn->phys_port,
915 				mlx5_os_get_dev_device_name(spawn->phys_dev));
916 			devx_port.comp_mask = 0;
917 		}
918 	}
919 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
920 		priv->vport_meta_tag = devx_port.reg_c_0.value;
921 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
922 		if (!priv->vport_meta_mask) {
923 			DRV_LOG(ERR, "vport zero mask for port %d"
924 				     " on bonding device %s",
925 				     spawn->phys_port,
926 				     mlx5_os_get_dev_device_name
927 							(spawn->phys_dev));
928 			err = ENOTSUP;
929 			goto error;
930 		}
931 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
932 			DRV_LOG(ERR, "invalid vport tag for port %d"
933 				     " on bonding device %s",
934 				     spawn->phys_port,
935 				     mlx5_os_get_dev_device_name
936 							(spawn->phys_dev));
937 			err = ENOTSUP;
938 			goto error;
939 		}
940 	}
941 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
942 		priv->vport_id = devx_port.vport_num;
943 	} else if (spawn->pf_bond >= 0) {
944 		DRV_LOG(ERR, "can't deduce vport index for port %d"
945 			     " on bonding device %s",
946 			     spawn->phys_port,
947 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
948 		err = ENOTSUP;
949 		goto error;
950 	} else {
951 		/* Suppose vport index in compatible way. */
952 		priv->vport_id = switch_info->representor ?
953 				 switch_info->port_name + 1 : -1;
954 	}
955 #else
956 	/*
957 	 * Kernel/rdma_core support single E-Switch per PF configurations
958 	 * only and vport_id field contains the vport index for
959 	 * associated VF, which is deduced from representor port name.
960 	 * For example, let's have the IB device port 10, it has
961 	 * attached network device eth0, which has port name attribute
962 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
963 	 * as 3 (2+1). This assigning schema should be changed if the
964 	 * multiple E-Switch instances per PF configurations or/and PCI
965 	 * subfunctions are added.
966 	 */
967 	priv->vport_id = switch_info->representor ?
968 			 switch_info->port_name + 1 : -1;
969 #endif
970 	/* representor_id field keeps the unmodified VF index. */
971 	priv->representor_id = switch_info->representor ?
972 			       switch_info->port_name : -1;
973 	/*
974 	 * Look for sibling devices in order to reuse their switch domain
975 	 * if any, otherwise allocate one.
976 	 */
977 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
978 		const struct mlx5_priv *opriv =
979 			rte_eth_devices[port_id].data->dev_private;
980 
981 		if (!opriv ||
982 		    opriv->sh != priv->sh ||
983 			opriv->domain_id ==
984 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
985 			continue;
986 		priv->domain_id = opriv->domain_id;
987 		break;
988 	}
989 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
990 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
991 		if (err) {
992 			err = rte_errno;
993 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
994 				strerror(rte_errno));
995 			goto error;
996 		}
997 		own_domain_id = 1;
998 	}
999 	/* Override some values set by hardware configuration. */
1000 	mlx5_args(config, dpdk_dev->devargs);
1001 	err = mlx5_dev_check_sibling_config(priv, config);
1002 	if (err)
1003 		goto error;
1004 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1005 			    IBV_DEVICE_RAW_IP_CSUM);
1006 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1007 		(config->hw_csum ? "" : "not "));
1008 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1009 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1010 	DRV_LOG(DEBUG, "counters are not supported");
1011 #endif
1012 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1013 	if (config->dv_flow_en) {
1014 		DRV_LOG(WARNING, "DV flow is not supported");
1015 		config->dv_flow_en = 0;
1016 	}
1017 #endif
1018 	config->ind_table_max_size =
1019 		sh->device_attr.max_rwq_indirection_table_size;
1020 	/*
1021 	 * Remove this check once DPDK supports larger/variable
1022 	 * indirection tables.
1023 	 */
1024 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1025 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1026 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1027 		config->ind_table_max_size);
1028 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1029 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1030 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1031 		(config->hw_vlan_strip ? "" : "not "));
1032 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1033 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1034 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1035 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1036 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1037 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1038 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1039 #endif
1040 	if (config->hw_padding && !hw_padding) {
1041 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1042 		config->hw_padding = 0;
1043 	} else if (config->hw_padding) {
1044 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1045 	}
1046 	config->tso = (sh->device_attr.max_tso > 0 &&
1047 		      (sh->device_attr.tso_supported_qpts &
1048 		       (1 << IBV_QPT_RAW_PACKET)));
1049 	if (config->tso)
1050 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1051 	/*
1052 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1053 	 * by default.
1054 	 */
1055 	if (config->mps == MLX5_ARG_UNSET)
1056 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1057 							  MLX5_MPW_DISABLED;
1058 	else
1059 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1060 	DRV_LOG(INFO, "%sMPS is %s",
1061 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1062 		config->mps == MLX5_MPW ? "legacy " : "",
1063 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1064 	if (config->cqe_comp && !cqe_comp) {
1065 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1066 		config->cqe_comp = 0;
1067 	}
1068 	if (config->cqe_pad && !cqe_pad) {
1069 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1070 		config->cqe_pad = 0;
1071 	} else if (config->cqe_pad) {
1072 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1073 	}
1074 	if (config->devx) {
1075 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1076 		if (err) {
1077 			err = -err;
1078 			goto error;
1079 		}
1080 		/* Check relax ordering support. */
1081 		if (config->hca_attr.relaxed_ordering_write &&
1082 		    config->hca_attr.relaxed_ordering_read  &&
1083 		    !haswell_broadwell_cpu)
1084 			sh->cmng.relaxed_ordering = 1;
1085 		/* Check for LRO support. */
1086 		if (config->dest_tir && config->hca_attr.lro_cap &&
1087 		    config->dv_flow_en) {
1088 			/* TBD check tunnel lro caps. */
1089 			config->lro.supported = config->hca_attr.lro_cap;
1090 			DRV_LOG(DEBUG, "Device supports LRO");
1091 			/*
1092 			 * If LRO timeout is not configured by application,
1093 			 * use the minimal supported value.
1094 			 */
1095 			if (!config->lro.timeout)
1096 				config->lro.timeout =
1097 				config->hca_attr.lro_timer_supported_periods[0];
1098 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1099 				config->lro.timeout);
1100 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1101 				"required for coalescing is %d bytes",
1102 				config->hca_attr.lro_min_mss_size);
1103 		}
1104 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1105 		if (config->hca_attr.qos.sup &&
1106 		    config->hca_attr.qos.srtcm_sup &&
1107 		    config->dv_flow_en) {
1108 			uint8_t reg_c_mask =
1109 				config->hca_attr.qos.flow_meter_reg_c_ids;
1110 			/*
1111 			 * Meter needs two REG_C's for color match and pre-sfx
1112 			 * flow match. Here get the REG_C for color match.
1113 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1114 			 */
1115 			reg_c_mask &= 0xfc;
1116 			if (__builtin_popcount(reg_c_mask) < 1) {
1117 				priv->mtr_en = 0;
1118 				DRV_LOG(WARNING, "No available register for"
1119 					" meter.");
1120 			} else {
1121 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1122 						      REG_C_0;
1123 				priv->mtr_en = 1;
1124 				priv->mtr_reg_share =
1125 				      config->hca_attr.qos.flow_meter_reg_share;
1126 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1127 					priv->mtr_color_reg);
1128 			}
1129 		}
1130 #endif
1131 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1132 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1133 		    config->dv_flow_en) {
1134 			priv->sampler_en = 1;
1135 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1136 		} else {
1137 			priv->sampler_en = 0;
1138 			if (!config->hca_attr.log_max_ft_sampler_num)
1139 				DRV_LOG(WARNING, "No available register for"
1140 						" Sampler.");
1141 			else
1142 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1143 		}
1144 #endif
1145 	}
1146 	if (config->tx_pp) {
1147 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1148 			config->hca_attr.dev_freq_khz);
1149 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1150 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1151 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1152 			config->hca_attr.cross_channel ? "" : "not ");
1153 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1154 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1155 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1156 			config->hca_attr.non_wire_sq ? "" : "not ");
1157 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1158 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1159 			config->hca_attr.log_max_static_sq_wq);
1160 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1161 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1162 		if (!config->devx) {
1163 			DRV_LOG(ERR, "DevX is required for packet pacing");
1164 			err = ENODEV;
1165 			goto error;
1166 		}
1167 		if (!config->hca_attr.qos.packet_pacing) {
1168 			DRV_LOG(ERR, "Packet pacing is not supported");
1169 			err = ENODEV;
1170 			goto error;
1171 		}
1172 		if (!config->hca_attr.cross_channel) {
1173 			DRV_LOG(ERR, "Cross channel operations are"
1174 				     " required for packet pacing");
1175 			err = ENODEV;
1176 			goto error;
1177 		}
1178 		if (!config->hca_attr.wqe_index_ignore) {
1179 			DRV_LOG(ERR, "WQE index ignore feature is"
1180 				     " required for packet pacing");
1181 			err = ENODEV;
1182 			goto error;
1183 		}
1184 		if (!config->hca_attr.non_wire_sq) {
1185 			DRV_LOG(ERR, "Non-wire SQ feature is"
1186 				     " required for packet pacing");
1187 			err = ENODEV;
1188 			goto error;
1189 		}
1190 		if (!config->hca_attr.log_max_static_sq_wq) {
1191 			DRV_LOG(ERR, "Static WQE SQ feature is"
1192 				     " required for packet pacing");
1193 			err = ENODEV;
1194 			goto error;
1195 		}
1196 		if (!config->hca_attr.qos.wqe_rate_pp) {
1197 			DRV_LOG(ERR, "WQE rate mode is required"
1198 				     " for packet pacing");
1199 			err = ENODEV;
1200 			goto error;
1201 		}
1202 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1203 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1204 			     " can't create queues for packet pacing");
1205 		err = ENODEV;
1206 		goto error;
1207 #endif
1208 	}
1209 	if (config->devx) {
1210 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1211 
1212 		err = config->hca_attr.access_register_user ?
1213 			mlx5_devx_cmd_register_read
1214 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1215 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1216 		if (!err) {
1217 			uint32_t ts_mode;
1218 
1219 			/* MTUTC register is read successfully. */
1220 			ts_mode = MLX5_GET(register_mtutc, reg,
1221 					   time_stamp_mode);
1222 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1223 				config->rt_timestamp = 1;
1224 		} else {
1225 			/* Kernel does not support register reading. */
1226 			if (config->hca_attr.dev_freq_khz ==
1227 						 (NS_PER_S / MS_PER_S))
1228 				config->rt_timestamp = 1;
1229 		}
1230 	}
1231 	/*
1232 	 * If HW has bug working with tunnel packet decapsulation and
1233 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1234 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1235 	 */
1236 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1237 		config->hw_fcs_strip = 0;
1238 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1239 		(config->hw_fcs_strip ? "" : "not "));
1240 	if (config->mprq.enabled && mprq) {
1241 		if (config->mprq.stride_num_n &&
1242 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1243 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1244 			config->mprq.stride_num_n =
1245 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1246 						mprq_min_stride_num_n),
1247 					mprq_max_stride_num_n);
1248 			DRV_LOG(WARNING,
1249 				"the number of strides"
1250 				" for Multi-Packet RQ is out of range,"
1251 				" setting default value (%u)",
1252 				1 << config->mprq.stride_num_n);
1253 		}
1254 		if (config->mprq.stride_size_n &&
1255 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1256 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1257 			config->mprq.stride_size_n =
1258 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1259 						mprq_min_stride_size_n),
1260 					mprq_max_stride_size_n);
1261 			DRV_LOG(WARNING,
1262 				"the size of a stride"
1263 				" for Multi-Packet RQ is out of range,"
1264 				" setting default value (%u)",
1265 				1 << config->mprq.stride_size_n);
1266 		}
1267 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1268 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1269 	} else if (config->mprq.enabled && !mprq) {
1270 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1271 		config->mprq.enabled = 0;
1272 	}
1273 	if (config->max_dump_files_num == 0)
1274 		config->max_dump_files_num = 128;
1275 	eth_dev = rte_eth_dev_allocate(name);
1276 	if (eth_dev == NULL) {
1277 		DRV_LOG(ERR, "can not allocate rte ethdev");
1278 		err = ENOMEM;
1279 		goto error;
1280 	}
1281 	if (priv->representor) {
1282 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1283 		eth_dev->data->representor_id = priv->representor_id;
1284 	}
1285 	/*
1286 	 * Store associated network device interface index. This index
1287 	 * is permanent throughout the lifetime of device. So, we may store
1288 	 * the ifindex here and use the cached value further.
1289 	 */
1290 	MLX5_ASSERT(spawn->ifindex);
1291 	priv->if_index = spawn->ifindex;
1292 	if (priv->pf_bond >= 0 && priv->master) {
1293 		/* Get bond interface info */
1294 		err = mlx5_sysfs_bond_info(priv->if_index,
1295 				     &priv->bond_ifindex,
1296 				     priv->bond_name);
1297 		if (err)
1298 			DRV_LOG(ERR, "unable to get bond info: %s",
1299 				strerror(rte_errno));
1300 		else
1301 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1302 				priv->if_index, priv->bond_ifindex,
1303 				priv->bond_name);
1304 	}
1305 	eth_dev->data->dev_private = priv;
1306 	priv->dev_data = eth_dev->data;
1307 	eth_dev->data->mac_addrs = priv->mac;
1308 	eth_dev->device = dpdk_dev;
1309 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1310 	/* Configure the first MAC address by default. */
1311 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1312 		DRV_LOG(ERR,
1313 			"port %u cannot get MAC address, is mlx5_en"
1314 			" loaded? (errno: %s)",
1315 			eth_dev->data->port_id, strerror(rte_errno));
1316 		err = ENODEV;
1317 		goto error;
1318 	}
1319 	DRV_LOG(INFO,
1320 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1321 		eth_dev->data->port_id,
1322 		mac.addr_bytes[0], mac.addr_bytes[1],
1323 		mac.addr_bytes[2], mac.addr_bytes[3],
1324 		mac.addr_bytes[4], mac.addr_bytes[5]);
1325 #ifdef RTE_LIBRTE_MLX5_DEBUG
1326 	{
1327 		char ifname[IF_NAMESIZE];
1328 
1329 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1330 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1331 				eth_dev->data->port_id, ifname);
1332 		else
1333 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1334 				eth_dev->data->port_id);
1335 	}
1336 #endif
1337 	/* Get actual MTU if possible. */
1338 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1339 	if (err) {
1340 		err = rte_errno;
1341 		goto error;
1342 	}
1343 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1344 		priv->mtu);
1345 	/* Initialize burst functions to prevent crashes before link-up. */
1346 	eth_dev->rx_pkt_burst = removed_rx_burst;
1347 	eth_dev->tx_pkt_burst = removed_tx_burst;
1348 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1349 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1350 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1351 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1352 	/* Register MAC address. */
1353 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1354 	if (config->vf && config->vf_nl_en)
1355 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1356 				      mlx5_ifindex(eth_dev),
1357 				      eth_dev->data->mac_addrs,
1358 				      MLX5_MAX_MAC_ADDRESSES);
1359 	priv->flows = 0;
1360 	priv->ctrl_flows = 0;
1361 	TAILQ_INIT(&priv->flow_meters);
1362 	TAILQ_INIT(&priv->flow_meter_profiles);
1363 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1364 	mlx5_glue->dv_set_context_attr(sh->ctx,
1365 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1366 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1367 				.alloc = &mlx5_alloc_verbs_buf,
1368 				.free = &mlx5_free_verbs_buf,
1369 				.data = priv,
1370 			}));
1371 	/* Bring Ethernet device up. */
1372 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1373 		eth_dev->data->port_id);
1374 	mlx5_set_link_up(eth_dev);
1375 	/*
1376 	 * Even though the interrupt handler is not installed yet,
1377 	 * interrupts will still trigger on the async_fd from
1378 	 * Verbs context returned by ibv_open_device().
1379 	 */
1380 	mlx5_link_update(eth_dev, 0);
1381 #ifdef HAVE_MLX5DV_DR_ESWITCH
1382 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1383 	      (switch_info->representor || switch_info->master)))
1384 		config->dv_esw_en = 0;
1385 #else
1386 	config->dv_esw_en = 0;
1387 #endif
1388 	/* Detect minimal data bytes to inline. */
1389 	mlx5_set_min_inline(spawn, config);
1390 	/* Store device configuration on private structure. */
1391 	priv->config = *config;
1392 	/* Create context for virtual machine VLAN workaround. */
1393 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1394 	if (config->dv_flow_en) {
1395 		err = mlx5_alloc_shared_dr(priv);
1396 		if (err)
1397 			goto error;
1398 		/*
1399 		 * RSS id is shared with meter flow id. Meter flow id can only
1400 		 * use the 24 MSB of the register.
1401 		 */
1402 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1403 				     MLX5_MTR_COLOR_BITS);
1404 		if (!priv->qrss_id_pool) {
1405 			DRV_LOG(ERR, "can't create flow id pool");
1406 			err = ENOMEM;
1407 			goto error;
1408 		}
1409 	}
1410 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1411 		priv->obj_ops = devx_obj_ops;
1412 		priv->obj_ops.drop_action_create =
1413 						ibv_obj_ops.drop_action_create;
1414 		priv->obj_ops.drop_action_destroy =
1415 						ibv_obj_ops.drop_action_destroy;
1416 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1417 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1418 #else
1419 		if (config->dv_esw_en)
1420 			priv->obj_ops.txq_obj_modify =
1421 						ibv_obj_ops.txq_obj_modify;
1422 #endif
1423 		/* Use specific wrappers for Tx object. */
1424 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1425 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1426 
1427 	} else {
1428 		priv->obj_ops = ibv_obj_ops;
1429 	}
1430 	/* Supported Verbs flow priority number detection. */
1431 	err = mlx5_flow_discover_priorities(eth_dev);
1432 	if (err < 0) {
1433 		err = -err;
1434 		goto error;
1435 	}
1436 	priv->config.flow_prio = err;
1437 	if (!priv->config.dv_esw_en &&
1438 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1439 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1440 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1441 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1442 	}
1443 	mlx5_set_metadata_mask(eth_dev);
1444 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1445 	    !priv->sh->dv_regc0_mask) {
1446 		DRV_LOG(ERR, "metadata mode %u is not supported "
1447 			     "(no metadata reg_c[0] is available)",
1448 			     priv->config.dv_xmeta_en);
1449 			err = ENOTSUP;
1450 			goto error;
1451 	}
1452 	/*
1453 	 * Allocate the buffer for flow creating, just once.
1454 	 * The allocation must be done before any flow creating.
1455 	 */
1456 	mlx5_flow_alloc_intermediate(eth_dev);
1457 	/* Query availability of metadata reg_c's. */
1458 	err = mlx5_flow_discover_mreg_c(eth_dev);
1459 	if (err < 0) {
1460 		err = -err;
1461 		goto error;
1462 	}
1463 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1464 		DRV_LOG(DEBUG,
1465 			"port %u extensive metadata register is not supported",
1466 			eth_dev->data->port_id);
1467 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1468 			DRV_LOG(ERR, "metadata mode %u is not supported "
1469 				     "(no metadata registers available)",
1470 				     priv->config.dv_xmeta_en);
1471 			err = ENOTSUP;
1472 			goto error;
1473 		}
1474 	}
1475 	if (priv->config.dv_flow_en &&
1476 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1477 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1478 	    priv->sh->dv_regc0_mask) {
1479 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1480 						      MLX5_FLOW_MREG_HTABLE_SZ);
1481 		if (!priv->mreg_cp_tbl) {
1482 			err = ENOMEM;
1483 			goto error;
1484 		}
1485 	}
1486 	mlx5_flow_counter_mode_config(eth_dev);
1487 	return eth_dev;
1488 error:
1489 	if (priv) {
1490 		if (priv->mreg_cp_tbl)
1491 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1492 		if (priv->sh)
1493 			mlx5_os_free_shared_dr(priv);
1494 		if (priv->nl_socket_route >= 0)
1495 			close(priv->nl_socket_route);
1496 		if (priv->nl_socket_rdma >= 0)
1497 			close(priv->nl_socket_rdma);
1498 		if (priv->vmwa_context)
1499 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1500 		if (priv->qrss_id_pool)
1501 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1502 		if (own_domain_id)
1503 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1504 		mlx5_free(priv);
1505 		if (eth_dev != NULL)
1506 			eth_dev->data->dev_private = NULL;
1507 	}
1508 	if (eth_dev != NULL) {
1509 		/* mac_addrs must not be freed alone because part of
1510 		 * dev_private
1511 		 **/
1512 		eth_dev->data->mac_addrs = NULL;
1513 		rte_eth_dev_release_port(eth_dev);
1514 	}
1515 	if (sh)
1516 		mlx5_free_shared_dev_ctx(sh);
1517 	MLX5_ASSERT(err > 0);
1518 	rte_errno = err;
1519 	return NULL;
1520 }
1521 
1522 /**
1523  * Comparison callback to sort device data.
1524  *
1525  * This is meant to be used with qsort().
1526  *
1527  * @param a[in]
1528  *   Pointer to pointer to first data object.
1529  * @param b[in]
1530  *   Pointer to pointer to second data object.
1531  *
1532  * @return
1533  *   0 if both objects are equal, less than 0 if the first argument is less
1534  *   than the second, greater than 0 otherwise.
1535  */
1536 static int
1537 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1538 {
1539 	const struct mlx5_switch_info *si_a =
1540 		&((const struct mlx5_dev_spawn_data *)a)->info;
1541 	const struct mlx5_switch_info *si_b =
1542 		&((const struct mlx5_dev_spawn_data *)b)->info;
1543 	int ret;
1544 
1545 	/* Master device first. */
1546 	ret = si_b->master - si_a->master;
1547 	if (ret)
1548 		return ret;
1549 	/* Then representor devices. */
1550 	ret = si_b->representor - si_a->representor;
1551 	if (ret)
1552 		return ret;
1553 	/* Unidentified devices come last in no specific order. */
1554 	if (!si_a->representor)
1555 		return 0;
1556 	/* Order representors by name. */
1557 	return si_a->port_name - si_b->port_name;
1558 }
1559 
1560 /**
1561  * Match PCI information for possible slaves of bonding device.
1562  *
1563  * @param[in] ibv_dev
1564  *   Pointer to Infiniband device structure.
1565  * @param[in] pci_dev
1566  *   Pointer to PCI device structure to match PCI address.
1567  * @param[in] nl_rdma
1568  *   Netlink RDMA group socket handle.
1569  *
1570  * @return
1571  *   negative value if no bonding device found, otherwise
1572  *   positive index of slave PF in bonding.
1573  */
1574 static int
1575 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1576 			   const struct rte_pci_device *pci_dev,
1577 			   int nl_rdma)
1578 {
1579 	char ifname[IF_NAMESIZE + 1];
1580 	unsigned int ifindex;
1581 	unsigned int np, i;
1582 	FILE *file = NULL;
1583 	int pf = -1;
1584 
1585 	/*
1586 	 * Try to get master device name. If something goes
1587 	 * wrong suppose the lack of kernel support and no
1588 	 * bonding devices.
1589 	 */
1590 	if (nl_rdma < 0)
1591 		return -1;
1592 	if (!strstr(ibv_dev->name, "bond"))
1593 		return -1;
1594 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1595 	if (!np)
1596 		return -1;
1597 	/*
1598 	 * The Master device might not be on the predefined
1599 	 * port (not on port index 1, it is not garanted),
1600 	 * we have to scan all Infiniband device port and
1601 	 * find master.
1602 	 */
1603 	for (i = 1; i <= np; ++i) {
1604 		/* Check whether Infiniband port is populated. */
1605 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1606 		if (!ifindex)
1607 			continue;
1608 		if (!if_indextoname(ifindex, ifname))
1609 			continue;
1610 		/* Try to read bonding slave names from sysfs. */
1611 		MKSTR(slaves,
1612 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1613 		file = fopen(slaves, "r");
1614 		if (file)
1615 			break;
1616 	}
1617 	if (!file)
1618 		return -1;
1619 	/* Use safe format to check maximal buffer length. */
1620 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1621 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1622 		char tmp_str[IF_NAMESIZE + 32];
1623 		struct rte_pci_addr pci_addr;
1624 		struct mlx5_switch_info	info;
1625 
1626 		/* Process slave interface names in the loop. */
1627 		snprintf(tmp_str, sizeof(tmp_str),
1628 			 "/sys/class/net/%s", ifname);
1629 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1630 			DRV_LOG(WARNING, "can not get PCI address"
1631 					 " for netdev \"%s\"", ifname);
1632 			continue;
1633 		}
1634 		if (pci_dev->addr.domain != pci_addr.domain ||
1635 		    pci_dev->addr.bus != pci_addr.bus ||
1636 		    pci_dev->addr.devid != pci_addr.devid ||
1637 		    pci_dev->addr.function != pci_addr.function)
1638 			continue;
1639 		/* Slave interface PCI address match found. */
1640 		fclose(file);
1641 		snprintf(tmp_str, sizeof(tmp_str),
1642 			 "/sys/class/net/%s/phys_port_name", ifname);
1643 		file = fopen(tmp_str, "rb");
1644 		if (!file)
1645 			break;
1646 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1647 		if (fscanf(file, "%32s", tmp_str) == 1)
1648 			mlx5_translate_port_name(tmp_str, &info);
1649 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1650 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1651 			pf = info.port_name;
1652 		break;
1653 	}
1654 	if (file)
1655 		fclose(file);
1656 	return pf;
1657 }
1658 
1659 /**
1660  * DPDK callback to register a PCI device.
1661  *
1662  * This function spawns Ethernet devices out of a given PCI device.
1663  *
1664  * @param[in] pci_drv
1665  *   PCI driver structure (mlx5_driver).
1666  * @param[in] pci_dev
1667  *   PCI device information.
1668  *
1669  * @return
1670  *   0 on success, a negative errno value otherwise and rte_errno is set.
1671  */
1672 int
1673 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1674 		  struct rte_pci_device *pci_dev)
1675 {
1676 	struct ibv_device **ibv_list;
1677 	/*
1678 	 * Number of found IB Devices matching with requested PCI BDF.
1679 	 * nd != 1 means there are multiple IB devices over the same
1680 	 * PCI device and we have representors and master.
1681 	 */
1682 	unsigned int nd = 0;
1683 	/*
1684 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1685 	 * we have the single multiport IB device, and there may be
1686 	 * representors attached to some of found ports.
1687 	 */
1688 	unsigned int np = 0;
1689 	/*
1690 	 * Number of DPDK ethernet devices to Spawn - either over
1691 	 * multiple IB devices or multiple ports of single IB device.
1692 	 * Actually this is the number of iterations to spawn.
1693 	 */
1694 	unsigned int ns = 0;
1695 	/*
1696 	 * Bonding device
1697 	 *   < 0 - no bonding device (single one)
1698 	 *  >= 0 - bonding device (value is slave PF index)
1699 	 */
1700 	int bd = -1;
1701 	struct mlx5_dev_spawn_data *list = NULL;
1702 	struct mlx5_dev_config dev_config;
1703 	unsigned int dev_config_vf;
1704 	int ret;
1705 
1706 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1707 		mlx5_pmd_socket_init();
1708 	ret = mlx5_init_once();
1709 	if (ret) {
1710 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1711 			strerror(rte_errno));
1712 		return -rte_errno;
1713 	}
1714 	errno = 0;
1715 	ibv_list = mlx5_glue->get_device_list(&ret);
1716 	if (!ibv_list) {
1717 		rte_errno = errno ? errno : ENOSYS;
1718 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1719 		return -rte_errno;
1720 	}
1721 	/*
1722 	 * First scan the list of all Infiniband devices to find
1723 	 * matching ones, gathering into the list.
1724 	 */
1725 	struct ibv_device *ibv_match[ret + 1];
1726 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1727 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1728 	unsigned int i;
1729 
1730 	while (ret-- > 0) {
1731 		struct rte_pci_addr pci_addr;
1732 
1733 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1734 		bd = mlx5_device_bond_pci_match
1735 				(ibv_list[ret], pci_dev, nl_rdma);
1736 		if (bd >= 0) {
1737 			/*
1738 			 * Bonding device detected. Only one match is allowed,
1739 			 * the bonding is supported over multi-port IB device,
1740 			 * there should be no matches on representor PCI
1741 			 * functions or non VF LAG bonding devices with
1742 			 * specified address.
1743 			 */
1744 			if (nd) {
1745 				DRV_LOG(ERR,
1746 					"multiple PCI match on bonding device"
1747 					"\"%s\" found", ibv_list[ret]->name);
1748 				rte_errno = ENOENT;
1749 				ret = -rte_errno;
1750 				goto exit;
1751 			}
1752 			DRV_LOG(INFO, "PCI information matches for"
1753 				      " slave %d bonding device \"%s\"",
1754 				      bd, ibv_list[ret]->name);
1755 			ibv_match[nd++] = ibv_list[ret];
1756 			break;
1757 		}
1758 		if (mlx5_dev_to_pci_addr
1759 			(ibv_list[ret]->ibdev_path, &pci_addr))
1760 			continue;
1761 		if (pci_dev->addr.domain != pci_addr.domain ||
1762 		    pci_dev->addr.bus != pci_addr.bus ||
1763 		    pci_dev->addr.devid != pci_addr.devid ||
1764 		    pci_dev->addr.function != pci_addr.function)
1765 			continue;
1766 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1767 			ibv_list[ret]->name);
1768 		ibv_match[nd++] = ibv_list[ret];
1769 	}
1770 	ibv_match[nd] = NULL;
1771 	if (!nd) {
1772 		/* No device matches, just complain and bail out. */
1773 		DRV_LOG(WARNING,
1774 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1775 			" are kernel drivers loaded?",
1776 			pci_dev->addr.domain, pci_dev->addr.bus,
1777 			pci_dev->addr.devid, pci_dev->addr.function);
1778 		rte_errno = ENOENT;
1779 		ret = -rte_errno;
1780 		goto exit;
1781 	}
1782 	if (nd == 1) {
1783 		/*
1784 		 * Found single matching device may have multiple ports.
1785 		 * Each port may be representor, we have to check the port
1786 		 * number and check the representors existence.
1787 		 */
1788 		if (nl_rdma >= 0)
1789 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1790 		if (!np)
1791 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1792 					 " ports number", ibv_match[0]->name);
1793 		if (bd >= 0 && !np) {
1794 			DRV_LOG(ERR, "can not get ports"
1795 				     " for bonding device");
1796 			rte_errno = ENOENT;
1797 			ret = -rte_errno;
1798 			goto exit;
1799 		}
1800 	}
1801 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1802 	if (bd >= 0) {
1803 		/*
1804 		 * This may happen if there is VF LAG kernel support and
1805 		 * application is compiled with older rdma_core library.
1806 		 */
1807 		DRV_LOG(ERR,
1808 			"No kernel/verbs support for VF LAG bonding found.");
1809 		rte_errno = ENOTSUP;
1810 		ret = -rte_errno;
1811 		goto exit;
1812 	}
1813 #endif
1814 	/*
1815 	 * Now we can determine the maximal
1816 	 * amount of devices to be spawned.
1817 	 */
1818 	list = mlx5_malloc(MLX5_MEM_ZERO,
1819 			   sizeof(struct mlx5_dev_spawn_data) *
1820 			   (np ? np : nd),
1821 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1822 	if (!list) {
1823 		DRV_LOG(ERR, "spawn data array allocation failure");
1824 		rte_errno = ENOMEM;
1825 		ret = -rte_errno;
1826 		goto exit;
1827 	}
1828 	if (bd >= 0 || np > 1) {
1829 		/*
1830 		 * Single IB device with multiple ports found,
1831 		 * it may be E-Switch master device and representors.
1832 		 * We have to perform identification through the ports.
1833 		 */
1834 		MLX5_ASSERT(nl_rdma >= 0);
1835 		MLX5_ASSERT(ns == 0);
1836 		MLX5_ASSERT(nd == 1);
1837 		MLX5_ASSERT(np);
1838 		for (i = 1; i <= np; ++i) {
1839 			list[ns].max_port = np;
1840 			list[ns].phys_port = i;
1841 			list[ns].phys_dev = ibv_match[0];
1842 			list[ns].eth_dev = NULL;
1843 			list[ns].pci_dev = pci_dev;
1844 			list[ns].pf_bond = bd;
1845 			list[ns].ifindex = mlx5_nl_ifindex
1846 				(nl_rdma,
1847 				mlx5_os_get_dev_device_name
1848 						(list[ns].phys_dev), i);
1849 			if (!list[ns].ifindex) {
1850 				/*
1851 				 * No network interface index found for the
1852 				 * specified port, it means there is no
1853 				 * representor on this port. It's OK,
1854 				 * there can be disabled ports, for example
1855 				 * if sriov_numvfs < sriov_totalvfs.
1856 				 */
1857 				continue;
1858 			}
1859 			ret = -1;
1860 			if (nl_route >= 0)
1861 				ret = mlx5_nl_switch_info
1862 					       (nl_route,
1863 						list[ns].ifindex,
1864 						&list[ns].info);
1865 			if (ret || (!list[ns].info.representor &&
1866 				    !list[ns].info.master)) {
1867 				/*
1868 				 * We failed to recognize representors with
1869 				 * Netlink, let's try to perform the task
1870 				 * with sysfs.
1871 				 */
1872 				ret =  mlx5_sysfs_switch_info
1873 						(list[ns].ifindex,
1874 						 &list[ns].info);
1875 			}
1876 			if (!ret && bd >= 0) {
1877 				switch (list[ns].info.name_type) {
1878 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1879 					if (list[ns].info.port_name == bd)
1880 						ns++;
1881 					break;
1882 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1883 					/* Fallthrough */
1884 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1885 					if (list[ns].info.pf_num == bd)
1886 						ns++;
1887 					break;
1888 				default:
1889 					break;
1890 				}
1891 				continue;
1892 			}
1893 			if (!ret && (list[ns].info.representor ^
1894 				     list[ns].info.master))
1895 				ns++;
1896 		}
1897 		if (!ns) {
1898 			DRV_LOG(ERR,
1899 				"unable to recognize master/representors"
1900 				" on the IB device with multiple ports");
1901 			rte_errno = ENOENT;
1902 			ret = -rte_errno;
1903 			goto exit;
1904 		}
1905 	} else {
1906 		/*
1907 		 * The existence of several matching entries (nd > 1) means
1908 		 * port representors have been instantiated. No existing Verbs
1909 		 * call nor sysfs entries can tell them apart, this can only
1910 		 * be done through Netlink calls assuming kernel drivers are
1911 		 * recent enough to support them.
1912 		 *
1913 		 * In the event of identification failure through Netlink,
1914 		 * try again through sysfs, then:
1915 		 *
1916 		 * 1. A single IB device matches (nd == 1) with single
1917 		 *    port (np=0/1) and is not a representor, assume
1918 		 *    no switch support.
1919 		 *
1920 		 * 2. Otherwise no safe assumptions can be made;
1921 		 *    complain louder and bail out.
1922 		 */
1923 		for (i = 0; i != nd; ++i) {
1924 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1925 			list[ns].max_port = 1;
1926 			list[ns].phys_port = 1;
1927 			list[ns].phys_dev = ibv_match[i];
1928 			list[ns].eth_dev = NULL;
1929 			list[ns].pci_dev = pci_dev;
1930 			list[ns].pf_bond = -1;
1931 			list[ns].ifindex = 0;
1932 			if (nl_rdma >= 0)
1933 				list[ns].ifindex = mlx5_nl_ifindex
1934 				(nl_rdma,
1935 				mlx5_os_get_dev_device_name
1936 						(list[ns].phys_dev), 1);
1937 			if (!list[ns].ifindex) {
1938 				char ifname[IF_NAMESIZE];
1939 
1940 				/*
1941 				 * Netlink failed, it may happen with old
1942 				 * ib_core kernel driver (before 4.16).
1943 				 * We can assume there is old driver because
1944 				 * here we are processing single ports IB
1945 				 * devices. Let's try sysfs to retrieve
1946 				 * the ifindex. The method works for
1947 				 * master device only.
1948 				 */
1949 				if (nd > 1) {
1950 					/*
1951 					 * Multiple devices found, assume
1952 					 * representors, can not distinguish
1953 					 * master/representor and retrieve
1954 					 * ifindex via sysfs.
1955 					 */
1956 					continue;
1957 				}
1958 				ret = mlx5_get_ifname_sysfs
1959 					(ibv_match[i]->ibdev_path, ifname);
1960 				if (!ret)
1961 					list[ns].ifindex =
1962 						if_nametoindex(ifname);
1963 				if (!list[ns].ifindex) {
1964 					/*
1965 					 * No network interface index found
1966 					 * for the specified device, it means
1967 					 * there it is neither representor
1968 					 * nor master.
1969 					 */
1970 					continue;
1971 				}
1972 			}
1973 			ret = -1;
1974 			if (nl_route >= 0)
1975 				ret = mlx5_nl_switch_info
1976 					       (nl_route,
1977 						list[ns].ifindex,
1978 						&list[ns].info);
1979 			if (ret || (!list[ns].info.representor &&
1980 				    !list[ns].info.master)) {
1981 				/*
1982 				 * We failed to recognize representors with
1983 				 * Netlink, let's try to perform the task
1984 				 * with sysfs.
1985 				 */
1986 				ret =  mlx5_sysfs_switch_info
1987 						(list[ns].ifindex,
1988 						 &list[ns].info);
1989 			}
1990 			if (!ret && (list[ns].info.representor ^
1991 				     list[ns].info.master)) {
1992 				ns++;
1993 			} else if ((nd == 1) &&
1994 				   !list[ns].info.representor &&
1995 				   !list[ns].info.master) {
1996 				/*
1997 				 * Single IB device with
1998 				 * one physical port and
1999 				 * attached network device.
2000 				 * May be SRIOV is not enabled
2001 				 * or there is no representors.
2002 				 */
2003 				DRV_LOG(INFO, "no E-Switch support detected");
2004 				ns++;
2005 				break;
2006 			}
2007 		}
2008 		if (!ns) {
2009 			DRV_LOG(ERR,
2010 				"unable to recognize master/representors"
2011 				" on the multiple IB devices");
2012 			rte_errno = ENOENT;
2013 			ret = -rte_errno;
2014 			goto exit;
2015 		}
2016 	}
2017 	MLX5_ASSERT(ns);
2018 	/*
2019 	 * Sort list to probe devices in natural order for users convenience
2020 	 * (i.e. master first, then representors from lowest to highest ID).
2021 	 */
2022 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2023 	/* Device specific configuration. */
2024 	switch (pci_dev->id.device_id) {
2025 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2026 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2027 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2028 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2029 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2030 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2031 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2032 		dev_config_vf = 1;
2033 		break;
2034 	default:
2035 		dev_config_vf = 0;
2036 		break;
2037 	}
2038 	for (i = 0; i != ns; ++i) {
2039 		uint32_t restore;
2040 
2041 		/* Default configuration. */
2042 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2043 		dev_config.vf = dev_config_vf;
2044 		dev_config.mps = MLX5_ARG_UNSET;
2045 		dev_config.dbnc = MLX5_ARG_UNSET;
2046 		dev_config.rx_vec_en = 1;
2047 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2048 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2049 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2050 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2051 		dev_config.vf_nl_en = 1;
2052 		dev_config.mr_ext_memseg_en = 1;
2053 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2054 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2055 		dev_config.dv_esw_en = 1;
2056 		dev_config.dv_flow_en = 1;
2057 		dev_config.decap_en = 1;
2058 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2059 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2060 						 &list[i],
2061 						 &dev_config);
2062 		if (!list[i].eth_dev) {
2063 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2064 				break;
2065 			/* Device is disabled or already spawned. Ignore it. */
2066 			continue;
2067 		}
2068 		restore = list[i].eth_dev->data->dev_flags;
2069 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2070 		/* Restore non-PCI flags cleared by the above call. */
2071 		list[i].eth_dev->data->dev_flags |= restore;
2072 		rte_eth_dev_probing_finish(list[i].eth_dev);
2073 	}
2074 	if (i != ns) {
2075 		DRV_LOG(ERR,
2076 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2077 			" encountering an error: %s",
2078 			pci_dev->addr.domain, pci_dev->addr.bus,
2079 			pci_dev->addr.devid, pci_dev->addr.function,
2080 			strerror(rte_errno));
2081 		ret = -rte_errno;
2082 		/* Roll back. */
2083 		while (i--) {
2084 			if (!list[i].eth_dev)
2085 				continue;
2086 			mlx5_dev_close(list[i].eth_dev);
2087 			/* mac_addrs must not be freed because in dev_private */
2088 			list[i].eth_dev->data->mac_addrs = NULL;
2089 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2090 		}
2091 		/* Restore original error. */
2092 		rte_errno = -ret;
2093 	} else {
2094 		ret = 0;
2095 	}
2096 exit:
2097 	/*
2098 	 * Do the routine cleanup:
2099 	 * - close opened Netlink sockets
2100 	 * - free allocated spawn data array
2101 	 * - free the Infiniband device list
2102 	 */
2103 	if (nl_rdma >= 0)
2104 		close(nl_rdma);
2105 	if (nl_route >= 0)
2106 		close(nl_route);
2107 	if (list)
2108 		mlx5_free(list);
2109 	MLX5_ASSERT(ibv_list);
2110 	mlx5_glue->free_device_list(ibv_list);
2111 	return ret;
2112 }
2113 
2114 static int
2115 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2116 {
2117 	char *env;
2118 	int value;
2119 
2120 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2121 	/* Get environment variable to store. */
2122 	env = getenv(MLX5_SHUT_UP_BF);
2123 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2124 	if (config->dbnc == MLX5_ARG_UNSET)
2125 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2126 	else
2127 		setenv(MLX5_SHUT_UP_BF,
2128 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2129 	return value;
2130 }
2131 
2132 static void
2133 mlx5_restore_doorbell_mapping_env(int value)
2134 {
2135 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2136 	/* Restore the original environment variable state. */
2137 	if (value == MLX5_ARG_UNSET)
2138 		unsetenv(MLX5_SHUT_UP_BF);
2139 	else
2140 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2141 }
2142 
2143 /**
2144  * Extract pdn of PD object using DV API.
2145  *
2146  * @param[in] pd
2147  *   Pointer to the verbs PD object.
2148  * @param[out] pdn
2149  *   Pointer to the PD object number variable.
2150  *
2151  * @return
2152  *   0 on success, error value otherwise.
2153  */
2154 int
2155 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2156 {
2157 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2158 	struct mlx5dv_obj obj;
2159 	struct mlx5dv_pd pd_info;
2160 	int ret = 0;
2161 
2162 	obj.pd.in = pd;
2163 	obj.pd.out = &pd_info;
2164 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2165 	if (ret) {
2166 		DRV_LOG(DEBUG, "Fail to get PD object info");
2167 		return ret;
2168 	}
2169 	*pdn = pd_info.pdn;
2170 	return 0;
2171 #else
2172 	(void)pd;
2173 	(void)pdn;
2174 	return -ENOTSUP;
2175 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2176 }
2177 
2178 /**
2179  * Function API to open IB device.
2180  *
2181  * This function calls the Linux glue APIs to open a device.
2182  *
2183  * @param[in] spawn
2184  *   Pointer to the IB device attributes (name, port, etc).
2185  * @param[out] config
2186  *   Pointer to device configuration structure.
2187  * @param[out] sh
2188  *   Pointer to shared context structure.
2189  *
2190  * @return
2191  *   0 on success, a positive error value otherwise.
2192  */
2193 int
2194 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2195 		     const struct mlx5_dev_config *config,
2196 		     struct mlx5_dev_ctx_shared *sh)
2197 {
2198 	int dbmap_env;
2199 	int err = 0;
2200 
2201 	sh->numa_node = spawn->pci_dev->device.numa_node;
2202 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2203 	/*
2204 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2205 	 * before the device creation. The rdma_core library
2206 	 * checks the variable at device creation and
2207 	 * stores the result internally.
2208 	 */
2209 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2210 	/* Try to open IB device with DV first, then usual Verbs. */
2211 	errno = 0;
2212 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2213 	if (sh->ctx) {
2214 		sh->devx = 1;
2215 		DRV_LOG(DEBUG, "DevX is supported");
2216 		/* The device is created, no need for environment. */
2217 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2218 	} else {
2219 		/* The environment variable is still configured. */
2220 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2221 		err = errno ? errno : ENODEV;
2222 		/*
2223 		 * The environment variable is not needed anymore,
2224 		 * all device creation attempts are completed.
2225 		 */
2226 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2227 		if (!sh->ctx)
2228 			return err;
2229 		DRV_LOG(DEBUG, "DevX is NOT supported");
2230 		err = 0;
2231 	}
2232 	return err;
2233 }
2234 
2235 /**
2236  * Install shared asynchronous device events handler.
2237  * This function is implemented to support event sharing
2238  * between multiple ports of single IB device.
2239  *
2240  * @param sh
2241  *   Pointer to mlx5_dev_ctx_shared object.
2242  */
2243 void
2244 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2245 {
2246 	int ret;
2247 	int flags;
2248 
2249 	sh->intr_handle.fd = -1;
2250 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2251 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2252 		    F_SETFL, flags | O_NONBLOCK);
2253 	if (ret) {
2254 		DRV_LOG(INFO, "failed to change file descriptor async event"
2255 			" queue");
2256 	} else {
2257 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2258 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2259 		if (rte_intr_callback_register(&sh->intr_handle,
2260 					mlx5_dev_interrupt_handler, sh)) {
2261 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2262 			sh->intr_handle.fd = -1;
2263 		}
2264 	}
2265 	if (sh->devx) {
2266 #ifdef HAVE_IBV_DEVX_ASYNC
2267 		sh->intr_handle_devx.fd = -1;
2268 		sh->devx_comp =
2269 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2270 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2271 		if (!devx_comp) {
2272 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2273 			return;
2274 		}
2275 		flags = fcntl(devx_comp->fd, F_GETFL);
2276 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2277 		if (ret) {
2278 			DRV_LOG(INFO, "failed to change file descriptor"
2279 				" devx comp");
2280 			return;
2281 		}
2282 		sh->intr_handle_devx.fd = devx_comp->fd;
2283 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2284 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2285 					mlx5_dev_interrupt_handler_devx, sh)) {
2286 			DRV_LOG(INFO, "Fail to install the devx shared"
2287 				" interrupt.");
2288 			sh->intr_handle_devx.fd = -1;
2289 		}
2290 #endif /* HAVE_IBV_DEVX_ASYNC */
2291 	}
2292 }
2293 
2294 /**
2295  * Uninstall shared asynchronous device events handler.
2296  * This function is implemented to support event sharing
2297  * between multiple ports of single IB device.
2298  *
2299  * @param dev
2300  *   Pointer to mlx5_dev_ctx_shared object.
2301  */
2302 void
2303 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2304 {
2305 	if (sh->intr_handle.fd >= 0)
2306 		mlx5_intr_callback_unregister(&sh->intr_handle,
2307 					      mlx5_dev_interrupt_handler, sh);
2308 #ifdef HAVE_IBV_DEVX_ASYNC
2309 	if (sh->intr_handle_devx.fd >= 0)
2310 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2311 				  mlx5_dev_interrupt_handler_devx, sh);
2312 	if (sh->devx_comp)
2313 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2314 #endif
2315 }
2316 
2317 /**
2318  * Read statistics by a named counter.
2319  *
2320  * @param[in] priv
2321  *   Pointer to the private device data structure.
2322  * @param[in] ctr_name
2323  *   Pointer to the name of the statistic counter to read
2324  * @param[out] stat
2325  *   Pointer to read statistic value.
2326  * @return
2327  *   0 on success and stat is valud, 1 if failed to read the value
2328  *   rte_errno is set.
2329  *
2330  */
2331 int
2332 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2333 		      uint64_t *stat)
2334 {
2335 	int fd;
2336 
2337 	if (priv->sh) {
2338 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2339 		      priv->sh->ibdev_path,
2340 		      priv->dev_port,
2341 		      ctr_name);
2342 		fd = open(path, O_RDONLY);
2343 		/*
2344 		 * in switchdev the file location is not per port
2345 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2346 		 */
2347 		if (fd == -1) {
2348 			MKSTR(path1, "%s/hw_counters/%s",
2349 			      priv->sh->ibdev_path,
2350 			      ctr_name);
2351 			fd = open(path1, O_RDONLY);
2352 		}
2353 		if (fd != -1) {
2354 			char buf[21] = {'\0'};
2355 			ssize_t n = read(fd, buf, sizeof(buf));
2356 
2357 			close(fd);
2358 			if (n != -1) {
2359 				*stat = strtoull(buf, NULL, 10);
2360 				return 0;
2361 			}
2362 		}
2363 	}
2364 	*stat = 0;
2365 	return 1;
2366 }
2367 
2368 /**
2369  * Set the reg_mr and dereg_mr call backs
2370  *
2371  * @param reg_mr_cb[out]
2372  *   Pointer to reg_mr func
2373  * @param dereg_mr_cb[out]
2374  *   Pointer to dereg_mr func
2375  *
2376  */
2377 void
2378 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2379 		      mlx5_dereg_mr_t *dereg_mr_cb)
2380 {
2381 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2382 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2383 }
2384 
2385 /**
2386  * Remove a MAC address from device
2387  *
2388  * @param dev
2389  *   Pointer to Ethernet device structure.
2390  * @param index
2391  *   MAC address index.
2392  */
2393 void
2394 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2395 {
2396 	struct mlx5_priv *priv = dev->data->dev_private;
2397 	const int vf = priv->config.vf;
2398 
2399 	if (vf)
2400 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2401 					mlx5_ifindex(dev), priv->mac_own,
2402 					&dev->data->mac_addrs[index], index);
2403 }
2404 
2405 /**
2406  * Adds a MAC address to the device
2407  *
2408  * @param dev
2409  *   Pointer to Ethernet device structure.
2410  * @param mac_addr
2411  *   MAC address to register.
2412  * @param index
2413  *   MAC address index.
2414  *
2415  * @return
2416  *   0 on success, a negative errno value otherwise
2417  */
2418 int
2419 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2420 		     uint32_t index)
2421 {
2422 	struct mlx5_priv *priv = dev->data->dev_private;
2423 	const int vf = priv->config.vf;
2424 	int ret = 0;
2425 
2426 	if (vf)
2427 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2428 					   mlx5_ifindex(dev), priv->mac_own,
2429 					   mac, index);
2430 	return ret;
2431 }
2432 
2433 /**
2434  * Modify a VF MAC address
2435  *
2436  * @param priv
2437  *   Pointer to device private data.
2438  * @param mac_addr
2439  *   MAC address to modify into.
2440  * @param iface_idx
2441  *   Net device interface index
2442  * @param vf_index
2443  *   VF index
2444  *
2445  * @return
2446  *   0 on success, a negative errno value otherwise
2447  */
2448 int
2449 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2450 			   unsigned int iface_idx,
2451 			   struct rte_ether_addr *mac_addr,
2452 			   int vf_index)
2453 {
2454 	return mlx5_nl_vf_mac_addr_modify
2455 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2456 }
2457 
2458 /**
2459  * Set device promiscuous mode
2460  *
2461  * @param dev
2462  *   Pointer to Ethernet device structure.
2463  * @param enable
2464  *   0 - promiscuous is disabled, otherwise - enabled
2465  *
2466  * @return
2467  *   0 on success, a negative error value otherwise
2468  */
2469 int
2470 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2471 {
2472 	struct mlx5_priv *priv = dev->data->dev_private;
2473 
2474 	return mlx5_nl_promisc(priv->nl_socket_route,
2475 			       mlx5_ifindex(dev), !!enable);
2476 }
2477 
2478 /**
2479  * Set device promiscuous mode
2480  *
2481  * @param dev
2482  *   Pointer to Ethernet device structure.
2483  * @param enable
2484  *   0 - all multicase is disabled, otherwise - enabled
2485  *
2486  * @return
2487  *   0 on success, a negative error value otherwise
2488  */
2489 int
2490 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2491 {
2492 	struct mlx5_priv *priv = dev->data->dev_private;
2493 
2494 	return mlx5_nl_allmulti(priv->nl_socket_route,
2495 				mlx5_ifindex(dev), !!enable);
2496 }
2497 
2498 /**
2499  * Flush device MAC addresses
2500  *
2501  * @param dev
2502  *   Pointer to Ethernet device structure.
2503  *
2504  */
2505 void
2506 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2507 {
2508 	struct mlx5_priv *priv = dev->data->dev_private;
2509 
2510 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2511 			       dev->data->mac_addrs,
2512 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2513 }
2514 
2515 const struct eth_dev_ops mlx5_os_dev_ops = {
2516 	.dev_configure = mlx5_dev_configure,
2517 	.dev_start = mlx5_dev_start,
2518 	.dev_stop = mlx5_dev_stop,
2519 	.dev_set_link_down = mlx5_set_link_down,
2520 	.dev_set_link_up = mlx5_set_link_up,
2521 	.dev_close = mlx5_dev_close,
2522 	.promiscuous_enable = mlx5_promiscuous_enable,
2523 	.promiscuous_disable = mlx5_promiscuous_disable,
2524 	.allmulticast_enable = mlx5_allmulticast_enable,
2525 	.allmulticast_disable = mlx5_allmulticast_disable,
2526 	.link_update = mlx5_link_update,
2527 	.stats_get = mlx5_stats_get,
2528 	.stats_reset = mlx5_stats_reset,
2529 	.xstats_get = mlx5_xstats_get,
2530 	.xstats_reset = mlx5_xstats_reset,
2531 	.xstats_get_names = mlx5_xstats_get_names,
2532 	.fw_version_get = mlx5_fw_version_get,
2533 	.dev_infos_get = mlx5_dev_infos_get,
2534 	.read_clock = mlx5_txpp_read_clock,
2535 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2536 	.vlan_filter_set = mlx5_vlan_filter_set,
2537 	.rx_queue_setup = mlx5_rx_queue_setup,
2538 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2539 	.tx_queue_setup = mlx5_tx_queue_setup,
2540 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2541 	.rx_queue_release = mlx5_rx_queue_release,
2542 	.tx_queue_release = mlx5_tx_queue_release,
2543 	.rx_queue_start = mlx5_rx_queue_start,
2544 	.rx_queue_stop = mlx5_rx_queue_stop,
2545 	.tx_queue_start = mlx5_tx_queue_start,
2546 	.tx_queue_stop = mlx5_tx_queue_stop,
2547 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2548 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2549 	.mac_addr_remove = mlx5_mac_addr_remove,
2550 	.mac_addr_add = mlx5_mac_addr_add,
2551 	.mac_addr_set = mlx5_mac_addr_set,
2552 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2553 	.mtu_set = mlx5_dev_set_mtu,
2554 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2555 	.vlan_offload_set = mlx5_vlan_offload_set,
2556 	.reta_update = mlx5_dev_rss_reta_update,
2557 	.reta_query = mlx5_dev_rss_reta_query,
2558 	.rss_hash_update = mlx5_rss_hash_update,
2559 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2560 	.filter_ctrl = mlx5_dev_filter_ctrl,
2561 	.rxq_info_get = mlx5_rxq_info_get,
2562 	.txq_info_get = mlx5_txq_info_get,
2563 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2564 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2565 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2566 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2567 	.is_removed = mlx5_is_removed,
2568 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2569 	.get_module_info = mlx5_get_module_info,
2570 	.get_module_eeprom = mlx5_get_module_eeprom,
2571 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2572 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2573 	.hairpin_bind = mlx5_hairpin_bind,
2574 	.hairpin_unbind = mlx5_hairpin_unbind,
2575 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2576 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2577 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2578 };
2579 
2580 /* Available operations from secondary process. */
2581 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2582 	.stats_get = mlx5_stats_get,
2583 	.stats_reset = mlx5_stats_reset,
2584 	.xstats_get = mlx5_xstats_get,
2585 	.xstats_reset = mlx5_xstats_reset,
2586 	.xstats_get_names = mlx5_xstats_get_names,
2587 	.fw_version_get = mlx5_fw_version_get,
2588 	.dev_infos_get = mlx5_dev_infos_get,
2589 	.read_clock = mlx5_txpp_read_clock,
2590 	.rx_queue_start = mlx5_rx_queue_start,
2591 	.rx_queue_stop = mlx5_rx_queue_stop,
2592 	.tx_queue_start = mlx5_tx_queue_start,
2593 	.tx_queue_stop = mlx5_tx_queue_stop,
2594 	.rxq_info_get = mlx5_rxq_info_get,
2595 	.txq_info_get = mlx5_txq_info_get,
2596 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2597 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2598 	.get_module_info = mlx5_get_module_info,
2599 	.get_module_eeprom = mlx5_get_module_eeprom,
2600 };
2601 
2602 /* Available operations in flow isolated mode. */
2603 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2604 	.dev_configure = mlx5_dev_configure,
2605 	.dev_start = mlx5_dev_start,
2606 	.dev_stop = mlx5_dev_stop,
2607 	.dev_set_link_down = mlx5_set_link_down,
2608 	.dev_set_link_up = mlx5_set_link_up,
2609 	.dev_close = mlx5_dev_close,
2610 	.promiscuous_enable = mlx5_promiscuous_enable,
2611 	.promiscuous_disable = mlx5_promiscuous_disable,
2612 	.allmulticast_enable = mlx5_allmulticast_enable,
2613 	.allmulticast_disable = mlx5_allmulticast_disable,
2614 	.link_update = mlx5_link_update,
2615 	.stats_get = mlx5_stats_get,
2616 	.stats_reset = mlx5_stats_reset,
2617 	.xstats_get = mlx5_xstats_get,
2618 	.xstats_reset = mlx5_xstats_reset,
2619 	.xstats_get_names = mlx5_xstats_get_names,
2620 	.fw_version_get = mlx5_fw_version_get,
2621 	.dev_infos_get = mlx5_dev_infos_get,
2622 	.read_clock = mlx5_txpp_read_clock,
2623 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2624 	.vlan_filter_set = mlx5_vlan_filter_set,
2625 	.rx_queue_setup = mlx5_rx_queue_setup,
2626 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2627 	.tx_queue_setup = mlx5_tx_queue_setup,
2628 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2629 	.rx_queue_release = mlx5_rx_queue_release,
2630 	.tx_queue_release = mlx5_tx_queue_release,
2631 	.rx_queue_start = mlx5_rx_queue_start,
2632 	.rx_queue_stop = mlx5_rx_queue_stop,
2633 	.tx_queue_start = mlx5_tx_queue_start,
2634 	.tx_queue_stop = mlx5_tx_queue_stop,
2635 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2636 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2637 	.mac_addr_remove = mlx5_mac_addr_remove,
2638 	.mac_addr_add = mlx5_mac_addr_add,
2639 	.mac_addr_set = mlx5_mac_addr_set,
2640 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2641 	.mtu_set = mlx5_dev_set_mtu,
2642 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2643 	.vlan_offload_set = mlx5_vlan_offload_set,
2644 	.filter_ctrl = mlx5_dev_filter_ctrl,
2645 	.rxq_info_get = mlx5_rxq_info_get,
2646 	.txq_info_get = mlx5_txq_info_get,
2647 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2648 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2649 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2650 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2651 	.is_removed = mlx5_is_removed,
2652 	.get_module_info = mlx5_get_module_info,
2653 	.get_module_eeprom = mlx5_get_module_eeprom,
2654 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2655 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2656 	.hairpin_bind = mlx5_hairpin_bind,
2657 	.hairpin_unbind = mlx5_hairpin_unbind,
2658 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2659 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2660 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2661 };
2662