xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 16dbba257c9ceccb30577d189dc6ca7e1042552f)
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 (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
295 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
296 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
297 		if (sh->fdb_domain)
298 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
299 	}
300 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
301 #endif /* HAVE_MLX5DV_DR */
302 	return 0;
303 error:
304 	/* Rollback the created objects. */
305 	if (sh->rx_domain) {
306 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
307 		sh->rx_domain = NULL;
308 	}
309 	if (sh->tx_domain) {
310 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
311 		sh->tx_domain = NULL;
312 	}
313 	if (sh->fdb_domain) {
314 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
315 		sh->fdb_domain = NULL;
316 	}
317 	if (sh->esw_drop_action) {
318 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
319 		sh->esw_drop_action = NULL;
320 	}
321 	if (sh->pop_vlan_action) {
322 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
323 		sh->pop_vlan_action = NULL;
324 	}
325 	if (sh->encaps_decaps) {
326 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
327 		sh->encaps_decaps = NULL;
328 	}
329 	if (sh->modify_cmds) {
330 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
331 		sh->modify_cmds = NULL;
332 	}
333 	if (sh->tag_table) {
334 		/* tags should be destroyed with flow before. */
335 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
336 		sh->tag_table = NULL;
337 	}
338 	mlx5_free_table_hash_list(priv);
339 	return err;
340 }
341 
342 /**
343  * Destroy DR related data within private structure.
344  *
345  * @param[in] priv
346  *   Pointer to the private device data structure.
347  */
348 void
349 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
350 {
351 	struct mlx5_dev_ctx_shared *sh = priv->sh;
352 
353 	MLX5_ASSERT(sh && sh->refcnt);
354 	if (sh->refcnt > 1)
355 		return;
356 #ifdef HAVE_MLX5DV_DR
357 	if (sh->rx_domain) {
358 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
359 		sh->rx_domain = NULL;
360 	}
361 	if (sh->tx_domain) {
362 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
363 		sh->tx_domain = NULL;
364 	}
365 #ifdef HAVE_MLX5DV_DR_ESWITCH
366 	if (sh->fdb_domain) {
367 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
368 		sh->fdb_domain = NULL;
369 	}
370 	if (sh->esw_drop_action) {
371 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
372 		sh->esw_drop_action = NULL;
373 	}
374 #endif
375 	if (sh->pop_vlan_action) {
376 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
377 		sh->pop_vlan_action = NULL;
378 	}
379 	pthread_mutex_destroy(&sh->dv_mutex);
380 #endif /* HAVE_MLX5DV_DR */
381 	if (sh->encaps_decaps) {
382 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
383 		sh->encaps_decaps = NULL;
384 	}
385 	if (sh->modify_cmds) {
386 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
387 		sh->modify_cmds = NULL;
388 	}
389 	if (sh->tag_table) {
390 		/* tags should be destroyed with flow before. */
391 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
392 		sh->tag_table = NULL;
393 	}
394 	mlx5_free_table_hash_list(priv);
395 }
396 
397 /**
398  * Initialize shared data between primary and secondary process.
399  *
400  * A memzone is reserved by primary process and secondary processes attach to
401  * the memzone.
402  *
403  * @return
404  *   0 on success, a negative errno value otherwise and rte_errno is set.
405  */
406 static int
407 mlx5_init_shared_data(void)
408 {
409 	const struct rte_memzone *mz;
410 	int ret = 0;
411 
412 	rte_spinlock_lock(&mlx5_shared_data_lock);
413 	if (mlx5_shared_data == NULL) {
414 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
415 			/* Allocate shared memory. */
416 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
417 						 sizeof(*mlx5_shared_data),
418 						 SOCKET_ID_ANY, 0);
419 			if (mz == NULL) {
420 				DRV_LOG(ERR,
421 					"Cannot allocate mlx5 shared data");
422 				ret = -rte_errno;
423 				goto error;
424 			}
425 			mlx5_shared_data = mz->addr;
426 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
427 			rte_spinlock_init(&mlx5_shared_data->lock);
428 		} else {
429 			/* Lookup allocated shared memory. */
430 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
431 			if (mz == NULL) {
432 				DRV_LOG(ERR,
433 					"Cannot attach mlx5 shared data");
434 				ret = -rte_errno;
435 				goto error;
436 			}
437 			mlx5_shared_data = mz->addr;
438 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
439 		}
440 	}
441 error:
442 	rte_spinlock_unlock(&mlx5_shared_data_lock);
443 	return ret;
444 }
445 
446 /**
447  * PMD global initialization.
448  *
449  * Independent from individual device, this function initializes global
450  * per-PMD data structures distinguishing primary and secondary processes.
451  * Hence, each initialization is called once per a process.
452  *
453  * @return
454  *   0 on success, a negative errno value otherwise and rte_errno is set.
455  */
456 static int
457 mlx5_init_once(void)
458 {
459 	struct mlx5_shared_data *sd;
460 	struct mlx5_local_data *ld = &mlx5_local_data;
461 	int ret = 0;
462 
463 	if (mlx5_init_shared_data())
464 		return -rte_errno;
465 	sd = mlx5_shared_data;
466 	MLX5_ASSERT(sd);
467 	rte_spinlock_lock(&sd->lock);
468 	switch (rte_eal_process_type()) {
469 	case RTE_PROC_PRIMARY:
470 		if (sd->init_done)
471 			break;
472 		LIST_INIT(&sd->mem_event_cb_list);
473 		rte_rwlock_init(&sd->mem_event_rwlock);
474 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
475 						mlx5_mr_mem_event_cb, NULL);
476 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
477 					   mlx5_mp_os_primary_handle);
478 		if (ret)
479 			goto out;
480 		sd->init_done = true;
481 		break;
482 	case RTE_PROC_SECONDARY:
483 		if (ld->init_done)
484 			break;
485 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
486 					     mlx5_mp_os_secondary_handle);
487 		if (ret)
488 			goto out;
489 		++sd->secondary_cnt;
490 		ld->init_done = true;
491 		break;
492 	default:
493 		break;
494 	}
495 out:
496 	rte_spinlock_unlock(&sd->lock);
497 	return ret;
498 }
499 
500 /**
501  * Create the Tx queue DevX/Verbs object.
502  *
503  * @param dev
504  *   Pointer to Ethernet device.
505  * @param idx
506  *   Queue index in DPDK Tx queue array.
507  *
508  * @return
509  *   0 on success, a negative errno value otherwise and rte_errno is set.
510  */
511 static int
512 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
513 {
514 	struct mlx5_priv *priv = dev->data->dev_private;
515 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
516 	struct mlx5_txq_ctrl *txq_ctrl =
517 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
518 
519 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
520 		return mlx5_txq_devx_obj_new(dev, idx);
521 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
522 	if (!priv->config.dv_esw_en)
523 		return mlx5_txq_devx_obj_new(dev, idx);
524 #endif
525 	return mlx5_txq_ibv_obj_new(dev, idx);
526 }
527 
528 /**
529  * Release an Tx DevX/verbs queue object.
530  *
531  * @param txq_obj
532  *   DevX/Verbs Tx queue object.
533  */
534 static void
535 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
536 {
537 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
538 		mlx5_txq_devx_obj_release(txq_obj);
539 		return;
540 	}
541 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
542 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
543 		mlx5_txq_devx_obj_release(txq_obj);
544 		return;
545 	}
546 #endif
547 	mlx5_txq_ibv_obj_release(txq_obj);
548 }
549 
550 /**
551  * DV flow counter mode detect and config.
552  *
553  * @param dev
554  *   Pointer to rte_eth_dev structure.
555  *
556  */
557 static void
558 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
559 {
560 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
561 	struct mlx5_priv *priv = dev->data->dev_private;
562 	struct mlx5_dev_ctx_shared *sh = priv->sh;
563 	bool fallback;
564 
565 #ifndef HAVE_IBV_DEVX_ASYNC
566 	fallback = true;
567 #else
568 	fallback = false;
569 	if (!priv->config.devx || !priv->config.dv_flow_en ||
570 	    !priv->config.hca_attr.flow_counters_dump ||
571 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
572 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
573 		fallback = true;
574 #endif
575 	if (fallback)
576 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
577 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
578 			priv->config.hca_attr.flow_counters_dump,
579 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
580 	/* Initialize fallback mode only on the port initializes sh. */
581 	if (sh->refcnt == 1)
582 		sh->cmng.counter_fallback = fallback;
583 	else if (fallback != sh->cmng.counter_fallback)
584 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
585 			"with others:%d.", PORT_ID(priv), fallback);
586 #endif
587 }
588 
589 /**
590  * Spawn an Ethernet device from Verbs information.
591  *
592  * @param dpdk_dev
593  *   Backing DPDK device.
594  * @param spawn
595  *   Verbs device parameters (name, port, switch_info) to spawn.
596  * @param config
597  *   Device configuration parameters.
598  *
599  * @return
600  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
601  *   is set. The following errors are defined:
602  *
603  *   EBUSY: device is not supposed to be spawned.
604  *   EEXIST: device is already spawned
605  */
606 static struct rte_eth_dev *
607 mlx5_dev_spawn(struct rte_device *dpdk_dev,
608 	       struct mlx5_dev_spawn_data *spawn,
609 	       struct mlx5_dev_config *config)
610 {
611 	const struct mlx5_switch_info *switch_info = &spawn->info;
612 	struct mlx5_dev_ctx_shared *sh = NULL;
613 	struct ibv_port_attr port_attr;
614 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
615 	struct rte_eth_dev *eth_dev = NULL;
616 	struct mlx5_priv *priv = NULL;
617 	int err = 0;
618 	unsigned int hw_padding = 0;
619 	unsigned int mps;
620 	unsigned int cqe_comp;
621 	unsigned int cqe_pad = 0;
622 	unsigned int tunnel_en = 0;
623 	unsigned int mpls_en = 0;
624 	unsigned int swp = 0;
625 	unsigned int mprq = 0;
626 	unsigned int mprq_min_stride_size_n = 0;
627 	unsigned int mprq_max_stride_size_n = 0;
628 	unsigned int mprq_min_stride_num_n = 0;
629 	unsigned int mprq_max_stride_num_n = 0;
630 	struct rte_ether_addr mac;
631 	char name[RTE_ETH_NAME_MAX_LEN];
632 	int own_domain_id = 0;
633 	uint16_t port_id;
634 	unsigned int i;
635 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
636 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
637 #endif
638 
639 	/* Determine if this port representor is supposed to be spawned. */
640 	if (switch_info->representor && dpdk_dev->devargs) {
641 		struct rte_eth_devargs eth_da;
642 
643 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
644 		if (err) {
645 			rte_errno = -err;
646 			DRV_LOG(ERR, "failed to process device arguments: %s",
647 				strerror(rte_errno));
648 			return NULL;
649 		}
650 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
651 			if (eth_da.representor_ports[i] ==
652 			    (uint16_t)switch_info->port_name)
653 				break;
654 		if (i == eth_da.nb_representor_ports) {
655 			rte_errno = EBUSY;
656 			return NULL;
657 		}
658 	}
659 	/* Build device name. */
660 	if (spawn->pf_bond <  0) {
661 		/* Single device. */
662 		if (!switch_info->representor)
663 			strlcpy(name, dpdk_dev->name, sizeof(name));
664 		else
665 			snprintf(name, sizeof(name), "%s_representor_%u",
666 				 dpdk_dev->name, switch_info->port_name);
667 	} else {
668 		/* Bonding device. */
669 		if (!switch_info->representor)
670 			snprintf(name, sizeof(name), "%s_%s",
671 				 dpdk_dev->name,
672 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
673 		else
674 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
675 				 dpdk_dev->name,
676 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
677 				 switch_info->port_name);
678 	}
679 	/* check if the device is already spawned */
680 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
681 		rte_errno = EEXIST;
682 		return NULL;
683 	}
684 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
685 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
686 		struct mlx5_mp_id mp_id;
687 
688 		eth_dev = rte_eth_dev_attach_secondary(name);
689 		if (eth_dev == NULL) {
690 			DRV_LOG(ERR, "can not attach rte ethdev");
691 			rte_errno = ENOMEM;
692 			return NULL;
693 		}
694 		eth_dev->device = dpdk_dev;
695 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
696 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
697 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
698 		err = mlx5_proc_priv_init(eth_dev);
699 		if (err)
700 			return NULL;
701 		mp_id.port_id = eth_dev->data->port_id;
702 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
703 		/* Receive command fd from primary process */
704 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
705 		if (err < 0)
706 			goto err_secondary;
707 		/* Remap UAR for Tx queues. */
708 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
709 		if (err)
710 			goto err_secondary;
711 		/*
712 		 * Ethdev pointer is still required as input since
713 		 * the primary device is not accessible from the
714 		 * secondary process.
715 		 */
716 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
717 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
718 		return eth_dev;
719 err_secondary:
720 		mlx5_dev_close(eth_dev);
721 		return NULL;
722 	}
723 	/*
724 	 * Some parameters ("tx_db_nc" in particularly) are needed in
725 	 * advance to create dv/verbs device context. We proceed the
726 	 * devargs here to get ones, and later proceed devargs again
727 	 * to override some hardware settings.
728 	 */
729 	err = mlx5_args(config, dpdk_dev->devargs);
730 	if (err) {
731 		err = rte_errno;
732 		DRV_LOG(ERR, "failed to process device arguments: %s",
733 			strerror(rte_errno));
734 		goto error;
735 	}
736 	mlx5_malloc_mem_select(config->sys_mem_en);
737 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
738 	if (!sh)
739 		return NULL;
740 	config->devx = sh->devx;
741 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
742 	config->dest_tir = 1;
743 #endif
744 #ifdef HAVE_IBV_MLX5_MOD_SWP
745 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
746 #endif
747 	/*
748 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
749 	 * as all ConnectX-5 devices.
750 	 */
751 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
752 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
753 #endif
754 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
755 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
756 #endif
757 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
758 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
759 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
760 			DRV_LOG(DEBUG, "enhanced MPW is supported");
761 			mps = MLX5_MPW_ENHANCED;
762 		} else {
763 			DRV_LOG(DEBUG, "MPW is supported");
764 			mps = MLX5_MPW;
765 		}
766 	} else {
767 		DRV_LOG(DEBUG, "MPW isn't supported");
768 		mps = MLX5_MPW_DISABLED;
769 	}
770 #ifdef HAVE_IBV_MLX5_MOD_SWP
771 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
772 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
773 	DRV_LOG(DEBUG, "SWP support: %u", swp);
774 #endif
775 	config->swp = !!swp;
776 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
777 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
778 		struct mlx5dv_striding_rq_caps mprq_caps =
779 			dv_attr.striding_rq_caps;
780 
781 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
782 			mprq_caps.min_single_stride_log_num_of_bytes);
783 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
784 			mprq_caps.max_single_stride_log_num_of_bytes);
785 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
786 			mprq_caps.min_single_wqe_log_num_of_strides);
787 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
788 			mprq_caps.max_single_wqe_log_num_of_strides);
789 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
790 			mprq_caps.supported_qpts);
791 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
792 		mprq = 1;
793 		mprq_min_stride_size_n =
794 			mprq_caps.min_single_stride_log_num_of_bytes;
795 		mprq_max_stride_size_n =
796 			mprq_caps.max_single_stride_log_num_of_bytes;
797 		mprq_min_stride_num_n =
798 			mprq_caps.min_single_wqe_log_num_of_strides;
799 		mprq_max_stride_num_n =
800 			mprq_caps.max_single_wqe_log_num_of_strides;
801 	}
802 #endif
803 	if (RTE_CACHE_LINE_SIZE == 128 &&
804 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
805 		cqe_comp = 0;
806 	else
807 		cqe_comp = 1;
808 	config->cqe_comp = cqe_comp;
809 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
810 	/* Whether device supports 128B Rx CQE padding. */
811 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
812 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
813 #endif
814 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
815 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
816 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
817 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
818 			     (dv_attr.tunnel_offloads_caps &
819 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
820 			     (dv_attr.tunnel_offloads_caps &
821 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
822 	}
823 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
824 		tunnel_en ? "" : "not ");
825 #else
826 	DRV_LOG(WARNING,
827 		"tunnel offloading disabled due to old OFED/rdma-core version");
828 #endif
829 	config->tunnel_en = tunnel_en;
830 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
831 	mpls_en = ((dv_attr.tunnel_offloads_caps &
832 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
833 		   (dv_attr.tunnel_offloads_caps &
834 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
835 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
836 		mpls_en ? "" : "not ");
837 #else
838 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
839 		" old OFED/rdma-core version or firmware configuration");
840 #endif
841 	config->mpls_en = mpls_en;
842 	/* Check port status. */
843 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
844 	if (err) {
845 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
846 		goto error;
847 	}
848 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
849 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
850 		err = EINVAL;
851 		goto error;
852 	}
853 	if (port_attr.state != IBV_PORT_ACTIVE)
854 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
855 			mlx5_glue->port_state_str(port_attr.state),
856 			port_attr.state);
857 	/* Allocate private eth device data. */
858 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
859 			   sizeof(*priv),
860 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
861 	if (priv == NULL) {
862 		DRV_LOG(ERR, "priv allocation failure");
863 		err = ENOMEM;
864 		goto error;
865 	}
866 	priv->sh = sh;
867 	priv->dev_port = spawn->phys_port;
868 	priv->pci_dev = spawn->pci_dev;
869 	priv->mtu = RTE_ETHER_MTU;
870 	priv->mp_id.port_id = port_id;
871 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
872 	/* Some internal functions rely on Netlink sockets, open them now. */
873 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
874 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
875 	priv->representor = !!switch_info->representor;
876 	priv->master = !!switch_info->master;
877 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
878 	priv->vport_meta_tag = 0;
879 	priv->vport_meta_mask = 0;
880 	priv->pf_bond = spawn->pf_bond;
881 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
882 	/*
883 	 * The DevX port query API is implemented. E-Switch may use
884 	 * either vport or reg_c[0] metadata register to match on
885 	 * vport index. The engaged part of metadata register is
886 	 * defined by mask.
887 	 */
888 	if (switch_info->representor || switch_info->master) {
889 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
890 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
891 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
892 						 &devx_port);
893 		if (err) {
894 			DRV_LOG(WARNING,
895 				"can't query devx port %d on device %s",
896 				spawn->phys_port,
897 				mlx5_os_get_dev_device_name(spawn->phys_dev));
898 			devx_port.comp_mask = 0;
899 		}
900 	}
901 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
902 		priv->vport_meta_tag = devx_port.reg_c_0.value;
903 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
904 		if (!priv->vport_meta_mask) {
905 			DRV_LOG(ERR, "vport zero mask for port %d"
906 				     " on bonding device %s",
907 				     spawn->phys_port,
908 				     mlx5_os_get_dev_device_name
909 							(spawn->phys_dev));
910 			err = ENOTSUP;
911 			goto error;
912 		}
913 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
914 			DRV_LOG(ERR, "invalid vport tag for port %d"
915 				     " on bonding device %s",
916 				     spawn->phys_port,
917 				     mlx5_os_get_dev_device_name
918 							(spawn->phys_dev));
919 			err = ENOTSUP;
920 			goto error;
921 		}
922 	}
923 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
924 		priv->vport_id = devx_port.vport_num;
925 	} else if (spawn->pf_bond >= 0) {
926 		DRV_LOG(ERR, "can't deduce vport index for port %d"
927 			     " on bonding device %s",
928 			     spawn->phys_port,
929 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
930 		err = ENOTSUP;
931 		goto error;
932 	} else {
933 		/* Suppose vport index in compatible way. */
934 		priv->vport_id = switch_info->representor ?
935 				 switch_info->port_name + 1 : -1;
936 	}
937 #else
938 	/*
939 	 * Kernel/rdma_core support single E-Switch per PF configurations
940 	 * only and vport_id field contains the vport index for
941 	 * associated VF, which is deduced from representor port name.
942 	 * For example, let's have the IB device port 10, it has
943 	 * attached network device eth0, which has port name attribute
944 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
945 	 * as 3 (2+1). This assigning schema should be changed if the
946 	 * multiple E-Switch instances per PF configurations or/and PCI
947 	 * subfunctions are added.
948 	 */
949 	priv->vport_id = switch_info->representor ?
950 			 switch_info->port_name + 1 : -1;
951 #endif
952 	/* representor_id field keeps the unmodified VF index. */
953 	priv->representor_id = switch_info->representor ?
954 			       switch_info->port_name : -1;
955 	/*
956 	 * Look for sibling devices in order to reuse their switch domain
957 	 * if any, otherwise allocate one.
958 	 */
959 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
960 		const struct mlx5_priv *opriv =
961 			rte_eth_devices[port_id].data->dev_private;
962 
963 		if (!opriv ||
964 		    opriv->sh != priv->sh ||
965 			opriv->domain_id ==
966 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
967 			continue;
968 		priv->domain_id = opriv->domain_id;
969 		break;
970 	}
971 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
972 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
973 		if (err) {
974 			err = rte_errno;
975 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
976 				strerror(rte_errno));
977 			goto error;
978 		}
979 		own_domain_id = 1;
980 	}
981 	/* Override some values set by hardware configuration. */
982 	mlx5_args(config, dpdk_dev->devargs);
983 	err = mlx5_dev_check_sibling_config(priv, config);
984 	if (err)
985 		goto error;
986 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
987 			    IBV_DEVICE_RAW_IP_CSUM);
988 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
989 		(config->hw_csum ? "" : "not "));
990 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
991 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
992 	DRV_LOG(DEBUG, "counters are not supported");
993 #endif
994 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
995 	if (config->dv_flow_en) {
996 		DRV_LOG(WARNING, "DV flow is not supported");
997 		config->dv_flow_en = 0;
998 	}
999 #endif
1000 	config->ind_table_max_size =
1001 		sh->device_attr.max_rwq_indirection_table_size;
1002 	/*
1003 	 * Remove this check once DPDK supports larger/variable
1004 	 * indirection tables.
1005 	 */
1006 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1007 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1008 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1009 		config->ind_table_max_size);
1010 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1011 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1012 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1013 		(config->hw_vlan_strip ? "" : "not "));
1014 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1015 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1016 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1017 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1018 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1019 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1020 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1021 #endif
1022 	if (config->hw_padding && !hw_padding) {
1023 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1024 		config->hw_padding = 0;
1025 	} else if (config->hw_padding) {
1026 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1027 	}
1028 	config->tso = (sh->device_attr.max_tso > 0 &&
1029 		      (sh->device_attr.tso_supported_qpts &
1030 		       (1 << IBV_QPT_RAW_PACKET)));
1031 	if (config->tso)
1032 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1033 	/*
1034 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1035 	 * by default.
1036 	 */
1037 	if (config->mps == MLX5_ARG_UNSET)
1038 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1039 							  MLX5_MPW_DISABLED;
1040 	else
1041 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1042 	DRV_LOG(INFO, "%sMPS is %s",
1043 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1044 		config->mps == MLX5_MPW ? "legacy " : "",
1045 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1046 	if (config->cqe_comp && !cqe_comp) {
1047 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1048 		config->cqe_comp = 0;
1049 	}
1050 	if (config->cqe_pad && !cqe_pad) {
1051 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1052 		config->cqe_pad = 0;
1053 	} else if (config->cqe_pad) {
1054 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1055 	}
1056 	if (config->devx) {
1057 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1058 		if (err) {
1059 			err = -err;
1060 			goto error;
1061 		}
1062 		/* Check relax ordering support. */
1063 		if (config->hca_attr.relaxed_ordering_write &&
1064 		    config->hca_attr.relaxed_ordering_read  &&
1065 		    !haswell_broadwell_cpu)
1066 			sh->cmng.relaxed_ordering = 1;
1067 		/* Check for LRO support. */
1068 		if (config->dest_tir && config->hca_attr.lro_cap &&
1069 		    config->dv_flow_en) {
1070 			/* TBD check tunnel lro caps. */
1071 			config->lro.supported = config->hca_attr.lro_cap;
1072 			DRV_LOG(DEBUG, "Device supports LRO");
1073 			/*
1074 			 * If LRO timeout is not configured by application,
1075 			 * use the minimal supported value.
1076 			 */
1077 			if (!config->lro.timeout)
1078 				config->lro.timeout =
1079 				config->hca_attr.lro_timer_supported_periods[0];
1080 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1081 				config->lro.timeout);
1082 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1083 				"required for coalescing is %d bytes",
1084 				config->hca_attr.lro_min_mss_size);
1085 		}
1086 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1087 		if (config->hca_attr.qos.sup &&
1088 		    config->hca_attr.qos.srtcm_sup &&
1089 		    config->dv_flow_en) {
1090 			uint8_t reg_c_mask =
1091 				config->hca_attr.qos.flow_meter_reg_c_ids;
1092 			/*
1093 			 * Meter needs two REG_C's for color match and pre-sfx
1094 			 * flow match. Here get the REG_C for color match.
1095 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1096 			 */
1097 			reg_c_mask &= 0xfc;
1098 			if (__builtin_popcount(reg_c_mask) < 1) {
1099 				priv->mtr_en = 0;
1100 				DRV_LOG(WARNING, "No available register for"
1101 					" meter.");
1102 			} else {
1103 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1104 						      REG_C_0;
1105 				priv->mtr_en = 1;
1106 				priv->mtr_reg_share =
1107 				      config->hca_attr.qos.flow_meter_reg_share;
1108 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1109 					priv->mtr_color_reg);
1110 			}
1111 		}
1112 #endif
1113 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1114 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1115 		    config->dv_flow_en) {
1116 			priv->sampler_en = 1;
1117 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1118 		} else {
1119 			priv->sampler_en = 0;
1120 			if (!config->hca_attr.log_max_ft_sampler_num)
1121 				DRV_LOG(WARNING, "No available register for"
1122 						" Sampler.");
1123 			else
1124 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1125 		}
1126 #endif
1127 	}
1128 	if (config->tx_pp) {
1129 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1130 			config->hca_attr.dev_freq_khz);
1131 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1132 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1133 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1134 			config->hca_attr.cross_channel ? "" : "not ");
1135 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1136 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1137 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1138 			config->hca_attr.non_wire_sq ? "" : "not ");
1139 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1140 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1141 			config->hca_attr.log_max_static_sq_wq);
1142 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1143 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1144 		if (!config->devx) {
1145 			DRV_LOG(ERR, "DevX is required for packet pacing");
1146 			err = ENODEV;
1147 			goto error;
1148 		}
1149 		if (!config->hca_attr.qos.packet_pacing) {
1150 			DRV_LOG(ERR, "Packet pacing is not supported");
1151 			err = ENODEV;
1152 			goto error;
1153 		}
1154 		if (!config->hca_attr.cross_channel) {
1155 			DRV_LOG(ERR, "Cross channel operations are"
1156 				     " required for packet pacing");
1157 			err = ENODEV;
1158 			goto error;
1159 		}
1160 		if (!config->hca_attr.wqe_index_ignore) {
1161 			DRV_LOG(ERR, "WQE index ignore feature is"
1162 				     " required for packet pacing");
1163 			err = ENODEV;
1164 			goto error;
1165 		}
1166 		if (!config->hca_attr.non_wire_sq) {
1167 			DRV_LOG(ERR, "Non-wire SQ feature is"
1168 				     " required for packet pacing");
1169 			err = ENODEV;
1170 			goto error;
1171 		}
1172 		if (!config->hca_attr.log_max_static_sq_wq) {
1173 			DRV_LOG(ERR, "Static WQE SQ feature is"
1174 				     " required for packet pacing");
1175 			err = ENODEV;
1176 			goto error;
1177 		}
1178 		if (!config->hca_attr.qos.wqe_rate_pp) {
1179 			DRV_LOG(ERR, "WQE rate mode is required"
1180 				     " for packet pacing");
1181 			err = ENODEV;
1182 			goto error;
1183 		}
1184 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1185 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1186 			     " can't create queues for packet pacing");
1187 		err = ENODEV;
1188 		goto error;
1189 #endif
1190 	}
1191 	if (config->devx) {
1192 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1193 
1194 		err = config->hca_attr.access_register_user ?
1195 			mlx5_devx_cmd_register_read
1196 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1197 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1198 		if (!err) {
1199 			uint32_t ts_mode;
1200 
1201 			/* MTUTC register is read successfully. */
1202 			ts_mode = MLX5_GET(register_mtutc, reg,
1203 					   time_stamp_mode);
1204 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1205 				config->rt_timestamp = 1;
1206 		} else {
1207 			/* Kernel does not support register reading. */
1208 			if (config->hca_attr.dev_freq_khz ==
1209 						 (NS_PER_S / MS_PER_S))
1210 				config->rt_timestamp = 1;
1211 		}
1212 	}
1213 	/*
1214 	 * If HW has bug working with tunnel packet decapsulation and
1215 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1216 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1217 	 */
1218 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1219 		config->hw_fcs_strip = 0;
1220 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1221 		(config->hw_fcs_strip ? "" : "not "));
1222 	if (config->mprq.enabled && mprq) {
1223 		if (config->mprq.stride_num_n &&
1224 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1225 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1226 			config->mprq.stride_num_n =
1227 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1228 						mprq_min_stride_num_n),
1229 					mprq_max_stride_num_n);
1230 			DRV_LOG(WARNING,
1231 				"the number of strides"
1232 				" for Multi-Packet RQ is out of range,"
1233 				" setting default value (%u)",
1234 				1 << config->mprq.stride_num_n);
1235 		}
1236 		if (config->mprq.stride_size_n &&
1237 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1238 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1239 			config->mprq.stride_size_n =
1240 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1241 						mprq_min_stride_size_n),
1242 					mprq_max_stride_size_n);
1243 			DRV_LOG(WARNING,
1244 				"the size of a stride"
1245 				" for Multi-Packet RQ is out of range,"
1246 				" setting default value (%u)",
1247 				1 << config->mprq.stride_size_n);
1248 		}
1249 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1250 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1251 	} else if (config->mprq.enabled && !mprq) {
1252 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1253 		config->mprq.enabled = 0;
1254 	}
1255 	if (config->max_dump_files_num == 0)
1256 		config->max_dump_files_num = 128;
1257 	eth_dev = rte_eth_dev_allocate(name);
1258 	if (eth_dev == NULL) {
1259 		DRV_LOG(ERR, "can not allocate rte ethdev");
1260 		err = ENOMEM;
1261 		goto error;
1262 	}
1263 	if (priv->representor) {
1264 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1265 		eth_dev->data->representor_id = priv->representor_id;
1266 	}
1267 	/*
1268 	 * Store associated network device interface index. This index
1269 	 * is permanent throughout the lifetime of device. So, we may store
1270 	 * the ifindex here and use the cached value further.
1271 	 */
1272 	MLX5_ASSERT(spawn->ifindex);
1273 	priv->if_index = spawn->ifindex;
1274 	if (priv->pf_bond >= 0 && priv->master) {
1275 		/* Get bond interface info */
1276 		err = mlx5_sysfs_bond_info(priv->if_index,
1277 				     &priv->bond_ifindex,
1278 				     priv->bond_name);
1279 		if (err)
1280 			DRV_LOG(ERR, "unable to get bond info: %s",
1281 				strerror(rte_errno));
1282 		else
1283 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1284 				priv->if_index, priv->bond_ifindex,
1285 				priv->bond_name);
1286 	}
1287 	eth_dev->data->dev_private = priv;
1288 	priv->dev_data = eth_dev->data;
1289 	eth_dev->data->mac_addrs = priv->mac;
1290 	eth_dev->device = dpdk_dev;
1291 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1292 	/* Configure the first MAC address by default. */
1293 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1294 		DRV_LOG(ERR,
1295 			"port %u cannot get MAC address, is mlx5_en"
1296 			" loaded? (errno: %s)",
1297 			eth_dev->data->port_id, strerror(rte_errno));
1298 		err = ENODEV;
1299 		goto error;
1300 	}
1301 	DRV_LOG(INFO,
1302 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1303 		eth_dev->data->port_id,
1304 		mac.addr_bytes[0], mac.addr_bytes[1],
1305 		mac.addr_bytes[2], mac.addr_bytes[3],
1306 		mac.addr_bytes[4], mac.addr_bytes[5]);
1307 #ifdef RTE_LIBRTE_MLX5_DEBUG
1308 	{
1309 		char ifname[IF_NAMESIZE];
1310 
1311 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1312 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1313 				eth_dev->data->port_id, ifname);
1314 		else
1315 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1316 				eth_dev->data->port_id);
1317 	}
1318 #endif
1319 	/* Get actual MTU if possible. */
1320 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1321 	if (err) {
1322 		err = rte_errno;
1323 		goto error;
1324 	}
1325 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1326 		priv->mtu);
1327 	/* Initialize burst functions to prevent crashes before link-up. */
1328 	eth_dev->rx_pkt_burst = removed_rx_burst;
1329 	eth_dev->tx_pkt_burst = removed_tx_burst;
1330 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1331 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1332 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1333 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1334 	/* Register MAC address. */
1335 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1336 	if (config->vf && config->vf_nl_en)
1337 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1338 				      mlx5_ifindex(eth_dev),
1339 				      eth_dev->data->mac_addrs,
1340 				      MLX5_MAX_MAC_ADDRESSES);
1341 	priv->flows = 0;
1342 	priv->ctrl_flows = 0;
1343 	TAILQ_INIT(&priv->flow_meters);
1344 	TAILQ_INIT(&priv->flow_meter_profiles);
1345 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1346 	mlx5_glue->dv_set_context_attr(sh->ctx,
1347 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1348 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1349 				.alloc = &mlx5_alloc_verbs_buf,
1350 				.free = &mlx5_free_verbs_buf,
1351 				.data = priv,
1352 			}));
1353 	/* Bring Ethernet device up. */
1354 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1355 		eth_dev->data->port_id);
1356 	mlx5_set_link_up(eth_dev);
1357 	/*
1358 	 * Even though the interrupt handler is not installed yet,
1359 	 * interrupts will still trigger on the async_fd from
1360 	 * Verbs context returned by ibv_open_device().
1361 	 */
1362 	mlx5_link_update(eth_dev, 0);
1363 #ifdef HAVE_MLX5DV_DR_ESWITCH
1364 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1365 	      (switch_info->representor || switch_info->master)))
1366 		config->dv_esw_en = 0;
1367 #else
1368 	config->dv_esw_en = 0;
1369 #endif
1370 	/* Detect minimal data bytes to inline. */
1371 	mlx5_set_min_inline(spawn, config);
1372 	/* Store device configuration on private structure. */
1373 	priv->config = *config;
1374 	/* Create context for virtual machine VLAN workaround. */
1375 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1376 	if (config->dv_flow_en) {
1377 		err = mlx5_alloc_shared_dr(priv);
1378 		if (err)
1379 			goto error;
1380 		/*
1381 		 * RSS id is shared with meter flow id. Meter flow id can only
1382 		 * use the 24 MSB of the register.
1383 		 */
1384 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1385 				     MLX5_MTR_COLOR_BITS);
1386 		if (!priv->qrss_id_pool) {
1387 			DRV_LOG(ERR, "can't create flow id pool");
1388 			err = ENOMEM;
1389 			goto error;
1390 		}
1391 	}
1392 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1393 		priv->obj_ops = devx_obj_ops;
1394 		priv->obj_ops.drop_action_create =
1395 						ibv_obj_ops.drop_action_create;
1396 		priv->obj_ops.drop_action_destroy =
1397 						ibv_obj_ops.drop_action_destroy;
1398 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1399 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1400 #else
1401 		if (config->dv_esw_en)
1402 			priv->obj_ops.txq_obj_modify =
1403 						ibv_obj_ops.txq_obj_modify;
1404 #endif
1405 		/* Use specific wrappers for Tx object. */
1406 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1407 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1408 
1409 	} else {
1410 		priv->obj_ops = ibv_obj_ops;
1411 	}
1412 	/* Supported Verbs flow priority number detection. */
1413 	err = mlx5_flow_discover_priorities(eth_dev);
1414 	if (err < 0) {
1415 		err = -err;
1416 		goto error;
1417 	}
1418 	priv->config.flow_prio = err;
1419 	if (!priv->config.dv_esw_en &&
1420 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1421 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1422 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1423 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1424 	}
1425 	mlx5_set_metadata_mask(eth_dev);
1426 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1427 	    !priv->sh->dv_regc0_mask) {
1428 		DRV_LOG(ERR, "metadata mode %u is not supported "
1429 			     "(no metadata reg_c[0] is available)",
1430 			     priv->config.dv_xmeta_en);
1431 			err = ENOTSUP;
1432 			goto error;
1433 	}
1434 	/*
1435 	 * Allocate the buffer for flow creating, just once.
1436 	 * The allocation must be done before any flow creating.
1437 	 */
1438 	mlx5_flow_alloc_intermediate(eth_dev);
1439 	/* Query availability of metadata reg_c's. */
1440 	err = mlx5_flow_discover_mreg_c(eth_dev);
1441 	if (err < 0) {
1442 		err = -err;
1443 		goto error;
1444 	}
1445 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1446 		DRV_LOG(DEBUG,
1447 			"port %u extensive metadata register is not supported",
1448 			eth_dev->data->port_id);
1449 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1450 			DRV_LOG(ERR, "metadata mode %u is not supported "
1451 				     "(no metadata registers available)",
1452 				     priv->config.dv_xmeta_en);
1453 			err = ENOTSUP;
1454 			goto error;
1455 		}
1456 	}
1457 	if (priv->config.dv_flow_en &&
1458 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1459 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1460 	    priv->sh->dv_regc0_mask) {
1461 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1462 						      MLX5_FLOW_MREG_HTABLE_SZ);
1463 		if (!priv->mreg_cp_tbl) {
1464 			err = ENOMEM;
1465 			goto error;
1466 		}
1467 	}
1468 	mlx5_flow_counter_mode_config(eth_dev);
1469 	return eth_dev;
1470 error:
1471 	if (priv) {
1472 		if (priv->mreg_cp_tbl)
1473 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1474 		if (priv->sh)
1475 			mlx5_os_free_shared_dr(priv);
1476 		if (priv->nl_socket_route >= 0)
1477 			close(priv->nl_socket_route);
1478 		if (priv->nl_socket_rdma >= 0)
1479 			close(priv->nl_socket_rdma);
1480 		if (priv->vmwa_context)
1481 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1482 		if (priv->qrss_id_pool)
1483 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1484 		if (own_domain_id)
1485 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1486 		mlx5_free(priv);
1487 		if (eth_dev != NULL)
1488 			eth_dev->data->dev_private = NULL;
1489 	}
1490 	if (eth_dev != NULL) {
1491 		/* mac_addrs must not be freed alone because part of
1492 		 * dev_private
1493 		 **/
1494 		eth_dev->data->mac_addrs = NULL;
1495 		rte_eth_dev_release_port(eth_dev);
1496 	}
1497 	if (sh)
1498 		mlx5_free_shared_dev_ctx(sh);
1499 	MLX5_ASSERT(err > 0);
1500 	rte_errno = err;
1501 	return NULL;
1502 }
1503 
1504 /**
1505  * Comparison callback to sort device data.
1506  *
1507  * This is meant to be used with qsort().
1508  *
1509  * @param a[in]
1510  *   Pointer to pointer to first data object.
1511  * @param b[in]
1512  *   Pointer to pointer to second data object.
1513  *
1514  * @return
1515  *   0 if both objects are equal, less than 0 if the first argument is less
1516  *   than the second, greater than 0 otherwise.
1517  */
1518 static int
1519 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1520 {
1521 	const struct mlx5_switch_info *si_a =
1522 		&((const struct mlx5_dev_spawn_data *)a)->info;
1523 	const struct mlx5_switch_info *si_b =
1524 		&((const struct mlx5_dev_spawn_data *)b)->info;
1525 	int ret;
1526 
1527 	/* Master device first. */
1528 	ret = si_b->master - si_a->master;
1529 	if (ret)
1530 		return ret;
1531 	/* Then representor devices. */
1532 	ret = si_b->representor - si_a->representor;
1533 	if (ret)
1534 		return ret;
1535 	/* Unidentified devices come last in no specific order. */
1536 	if (!si_a->representor)
1537 		return 0;
1538 	/* Order representors by name. */
1539 	return si_a->port_name - si_b->port_name;
1540 }
1541 
1542 /**
1543  * Match PCI information for possible slaves of bonding device.
1544  *
1545  * @param[in] ibv_dev
1546  *   Pointer to Infiniband device structure.
1547  * @param[in] pci_dev
1548  *   Pointer to PCI device structure to match PCI address.
1549  * @param[in] nl_rdma
1550  *   Netlink RDMA group socket handle.
1551  *
1552  * @return
1553  *   negative value if no bonding device found, otherwise
1554  *   positive index of slave PF in bonding.
1555  */
1556 static int
1557 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1558 			   const struct rte_pci_device *pci_dev,
1559 			   int nl_rdma)
1560 {
1561 	char ifname[IF_NAMESIZE + 1];
1562 	unsigned int ifindex;
1563 	unsigned int np, i;
1564 	FILE *file = NULL;
1565 	int pf = -1;
1566 
1567 	/*
1568 	 * Try to get master device name. If something goes
1569 	 * wrong suppose the lack of kernel support and no
1570 	 * bonding devices.
1571 	 */
1572 	if (nl_rdma < 0)
1573 		return -1;
1574 	if (!strstr(ibv_dev->name, "bond"))
1575 		return -1;
1576 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1577 	if (!np)
1578 		return -1;
1579 	/*
1580 	 * The Master device might not be on the predefined
1581 	 * port (not on port index 1, it is not garanted),
1582 	 * we have to scan all Infiniband device port and
1583 	 * find master.
1584 	 */
1585 	for (i = 1; i <= np; ++i) {
1586 		/* Check whether Infiniband port is populated. */
1587 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1588 		if (!ifindex)
1589 			continue;
1590 		if (!if_indextoname(ifindex, ifname))
1591 			continue;
1592 		/* Try to read bonding slave names from sysfs. */
1593 		MKSTR(slaves,
1594 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1595 		file = fopen(slaves, "r");
1596 		if (file)
1597 			break;
1598 	}
1599 	if (!file)
1600 		return -1;
1601 	/* Use safe format to check maximal buffer length. */
1602 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1603 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1604 		char tmp_str[IF_NAMESIZE + 32];
1605 		struct rte_pci_addr pci_addr;
1606 		struct mlx5_switch_info	info;
1607 
1608 		/* Process slave interface names in the loop. */
1609 		snprintf(tmp_str, sizeof(tmp_str),
1610 			 "/sys/class/net/%s", ifname);
1611 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1612 			DRV_LOG(WARNING, "can not get PCI address"
1613 					 " for netdev \"%s\"", ifname);
1614 			continue;
1615 		}
1616 		if (pci_dev->addr.domain != pci_addr.domain ||
1617 		    pci_dev->addr.bus != pci_addr.bus ||
1618 		    pci_dev->addr.devid != pci_addr.devid ||
1619 		    pci_dev->addr.function != pci_addr.function)
1620 			continue;
1621 		/* Slave interface PCI address match found. */
1622 		fclose(file);
1623 		snprintf(tmp_str, sizeof(tmp_str),
1624 			 "/sys/class/net/%s/phys_port_name", ifname);
1625 		file = fopen(tmp_str, "rb");
1626 		if (!file)
1627 			break;
1628 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1629 		if (fscanf(file, "%32s", tmp_str) == 1)
1630 			mlx5_translate_port_name(tmp_str, &info);
1631 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1632 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1633 			pf = info.port_name;
1634 		break;
1635 	}
1636 	if (file)
1637 		fclose(file);
1638 	return pf;
1639 }
1640 
1641 /**
1642  * DPDK callback to register a PCI device.
1643  *
1644  * This function spawns Ethernet devices out of a given PCI device.
1645  *
1646  * @param[in] pci_drv
1647  *   PCI driver structure (mlx5_driver).
1648  * @param[in] pci_dev
1649  *   PCI device information.
1650  *
1651  * @return
1652  *   0 on success, a negative errno value otherwise and rte_errno is set.
1653  */
1654 int
1655 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1656 		  struct rte_pci_device *pci_dev)
1657 {
1658 	struct ibv_device **ibv_list;
1659 	/*
1660 	 * Number of found IB Devices matching with requested PCI BDF.
1661 	 * nd != 1 means there are multiple IB devices over the same
1662 	 * PCI device and we have representors and master.
1663 	 */
1664 	unsigned int nd = 0;
1665 	/*
1666 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1667 	 * we have the single multiport IB device, and there may be
1668 	 * representors attached to some of found ports.
1669 	 */
1670 	unsigned int np = 0;
1671 	/*
1672 	 * Number of DPDK ethernet devices to Spawn - either over
1673 	 * multiple IB devices or multiple ports of single IB device.
1674 	 * Actually this is the number of iterations to spawn.
1675 	 */
1676 	unsigned int ns = 0;
1677 	/*
1678 	 * Bonding device
1679 	 *   < 0 - no bonding device (single one)
1680 	 *  >= 0 - bonding device (value is slave PF index)
1681 	 */
1682 	int bd = -1;
1683 	struct mlx5_dev_spawn_data *list = NULL;
1684 	struct mlx5_dev_config dev_config;
1685 	unsigned int dev_config_vf;
1686 	int ret;
1687 
1688 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1689 		mlx5_pmd_socket_init();
1690 	ret = mlx5_init_once();
1691 	if (ret) {
1692 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1693 			strerror(rte_errno));
1694 		return -rte_errno;
1695 	}
1696 	errno = 0;
1697 	ibv_list = mlx5_glue->get_device_list(&ret);
1698 	if (!ibv_list) {
1699 		rte_errno = errno ? errno : ENOSYS;
1700 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1701 		return -rte_errno;
1702 	}
1703 	/*
1704 	 * First scan the list of all Infiniband devices to find
1705 	 * matching ones, gathering into the list.
1706 	 */
1707 	struct ibv_device *ibv_match[ret + 1];
1708 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1709 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1710 	unsigned int i;
1711 
1712 	while (ret-- > 0) {
1713 		struct rte_pci_addr pci_addr;
1714 
1715 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1716 		bd = mlx5_device_bond_pci_match
1717 				(ibv_list[ret], pci_dev, nl_rdma);
1718 		if (bd >= 0) {
1719 			/*
1720 			 * Bonding device detected. Only one match is allowed,
1721 			 * the bonding is supported over multi-port IB device,
1722 			 * there should be no matches on representor PCI
1723 			 * functions or non VF LAG bonding devices with
1724 			 * specified address.
1725 			 */
1726 			if (nd) {
1727 				DRV_LOG(ERR,
1728 					"multiple PCI match on bonding device"
1729 					"\"%s\" found", ibv_list[ret]->name);
1730 				rte_errno = ENOENT;
1731 				ret = -rte_errno;
1732 				goto exit;
1733 			}
1734 			DRV_LOG(INFO, "PCI information matches for"
1735 				      " slave %d bonding device \"%s\"",
1736 				      bd, ibv_list[ret]->name);
1737 			ibv_match[nd++] = ibv_list[ret];
1738 			break;
1739 		}
1740 		if (mlx5_dev_to_pci_addr
1741 			(ibv_list[ret]->ibdev_path, &pci_addr))
1742 			continue;
1743 		if (pci_dev->addr.domain != pci_addr.domain ||
1744 		    pci_dev->addr.bus != pci_addr.bus ||
1745 		    pci_dev->addr.devid != pci_addr.devid ||
1746 		    pci_dev->addr.function != pci_addr.function)
1747 			continue;
1748 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1749 			ibv_list[ret]->name);
1750 		ibv_match[nd++] = ibv_list[ret];
1751 	}
1752 	ibv_match[nd] = NULL;
1753 	if (!nd) {
1754 		/* No device matches, just complain and bail out. */
1755 		DRV_LOG(WARNING,
1756 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1757 			" are kernel drivers loaded?",
1758 			pci_dev->addr.domain, pci_dev->addr.bus,
1759 			pci_dev->addr.devid, pci_dev->addr.function);
1760 		rte_errno = ENOENT;
1761 		ret = -rte_errno;
1762 		goto exit;
1763 	}
1764 	if (nd == 1) {
1765 		/*
1766 		 * Found single matching device may have multiple ports.
1767 		 * Each port may be representor, we have to check the port
1768 		 * number and check the representors existence.
1769 		 */
1770 		if (nl_rdma >= 0)
1771 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1772 		if (!np)
1773 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1774 					 " ports number", ibv_match[0]->name);
1775 		if (bd >= 0 && !np) {
1776 			DRV_LOG(ERR, "can not get ports"
1777 				     " for bonding device");
1778 			rte_errno = ENOENT;
1779 			ret = -rte_errno;
1780 			goto exit;
1781 		}
1782 	}
1783 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1784 	if (bd >= 0) {
1785 		/*
1786 		 * This may happen if there is VF LAG kernel support and
1787 		 * application is compiled with older rdma_core library.
1788 		 */
1789 		DRV_LOG(ERR,
1790 			"No kernel/verbs support for VF LAG bonding found.");
1791 		rte_errno = ENOTSUP;
1792 		ret = -rte_errno;
1793 		goto exit;
1794 	}
1795 #endif
1796 	/*
1797 	 * Now we can determine the maximal
1798 	 * amount of devices to be spawned.
1799 	 */
1800 	list = mlx5_malloc(MLX5_MEM_ZERO,
1801 			   sizeof(struct mlx5_dev_spawn_data) *
1802 			   (np ? np : nd),
1803 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1804 	if (!list) {
1805 		DRV_LOG(ERR, "spawn data array allocation failure");
1806 		rte_errno = ENOMEM;
1807 		ret = -rte_errno;
1808 		goto exit;
1809 	}
1810 	if (bd >= 0 || np > 1) {
1811 		/*
1812 		 * Single IB device with multiple ports found,
1813 		 * it may be E-Switch master device and representors.
1814 		 * We have to perform identification through the ports.
1815 		 */
1816 		MLX5_ASSERT(nl_rdma >= 0);
1817 		MLX5_ASSERT(ns == 0);
1818 		MLX5_ASSERT(nd == 1);
1819 		MLX5_ASSERT(np);
1820 		for (i = 1; i <= np; ++i) {
1821 			list[ns].max_port = np;
1822 			list[ns].phys_port = i;
1823 			list[ns].phys_dev = ibv_match[0];
1824 			list[ns].eth_dev = NULL;
1825 			list[ns].pci_dev = pci_dev;
1826 			list[ns].pf_bond = bd;
1827 			list[ns].ifindex = mlx5_nl_ifindex
1828 				(nl_rdma,
1829 				mlx5_os_get_dev_device_name
1830 						(list[ns].phys_dev), i);
1831 			if (!list[ns].ifindex) {
1832 				/*
1833 				 * No network interface index found for the
1834 				 * specified port, it means there is no
1835 				 * representor on this port. It's OK,
1836 				 * there can be disabled ports, for example
1837 				 * if sriov_numvfs < sriov_totalvfs.
1838 				 */
1839 				continue;
1840 			}
1841 			ret = -1;
1842 			if (nl_route >= 0)
1843 				ret = mlx5_nl_switch_info
1844 					       (nl_route,
1845 						list[ns].ifindex,
1846 						&list[ns].info);
1847 			if (ret || (!list[ns].info.representor &&
1848 				    !list[ns].info.master)) {
1849 				/*
1850 				 * We failed to recognize representors with
1851 				 * Netlink, let's try to perform the task
1852 				 * with sysfs.
1853 				 */
1854 				ret =  mlx5_sysfs_switch_info
1855 						(list[ns].ifindex,
1856 						 &list[ns].info);
1857 			}
1858 			if (!ret && bd >= 0) {
1859 				switch (list[ns].info.name_type) {
1860 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1861 					if (list[ns].info.port_name == bd)
1862 						ns++;
1863 					break;
1864 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1865 					/* Fallthrough */
1866 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1867 					if (list[ns].info.pf_num == bd)
1868 						ns++;
1869 					break;
1870 				default:
1871 					break;
1872 				}
1873 				continue;
1874 			}
1875 			if (!ret && (list[ns].info.representor ^
1876 				     list[ns].info.master))
1877 				ns++;
1878 		}
1879 		if (!ns) {
1880 			DRV_LOG(ERR,
1881 				"unable to recognize master/representors"
1882 				" on the IB device with multiple ports");
1883 			rte_errno = ENOENT;
1884 			ret = -rte_errno;
1885 			goto exit;
1886 		}
1887 	} else {
1888 		/*
1889 		 * The existence of several matching entries (nd > 1) means
1890 		 * port representors have been instantiated. No existing Verbs
1891 		 * call nor sysfs entries can tell them apart, this can only
1892 		 * be done through Netlink calls assuming kernel drivers are
1893 		 * recent enough to support them.
1894 		 *
1895 		 * In the event of identification failure through Netlink,
1896 		 * try again through sysfs, then:
1897 		 *
1898 		 * 1. A single IB device matches (nd == 1) with single
1899 		 *    port (np=0/1) and is not a representor, assume
1900 		 *    no switch support.
1901 		 *
1902 		 * 2. Otherwise no safe assumptions can be made;
1903 		 *    complain louder and bail out.
1904 		 */
1905 		for (i = 0; i != nd; ++i) {
1906 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1907 			list[ns].max_port = 1;
1908 			list[ns].phys_port = 1;
1909 			list[ns].phys_dev = ibv_match[i];
1910 			list[ns].eth_dev = NULL;
1911 			list[ns].pci_dev = pci_dev;
1912 			list[ns].pf_bond = -1;
1913 			list[ns].ifindex = 0;
1914 			if (nl_rdma >= 0)
1915 				list[ns].ifindex = mlx5_nl_ifindex
1916 				(nl_rdma,
1917 				mlx5_os_get_dev_device_name
1918 						(list[ns].phys_dev), 1);
1919 			if (!list[ns].ifindex) {
1920 				char ifname[IF_NAMESIZE];
1921 
1922 				/*
1923 				 * Netlink failed, it may happen with old
1924 				 * ib_core kernel driver (before 4.16).
1925 				 * We can assume there is old driver because
1926 				 * here we are processing single ports IB
1927 				 * devices. Let's try sysfs to retrieve
1928 				 * the ifindex. The method works for
1929 				 * master device only.
1930 				 */
1931 				if (nd > 1) {
1932 					/*
1933 					 * Multiple devices found, assume
1934 					 * representors, can not distinguish
1935 					 * master/representor and retrieve
1936 					 * ifindex via sysfs.
1937 					 */
1938 					continue;
1939 				}
1940 				ret = mlx5_get_ifname_sysfs
1941 					(ibv_match[i]->ibdev_path, ifname);
1942 				if (!ret)
1943 					list[ns].ifindex =
1944 						if_nametoindex(ifname);
1945 				if (!list[ns].ifindex) {
1946 					/*
1947 					 * No network interface index found
1948 					 * for the specified device, it means
1949 					 * there it is neither representor
1950 					 * nor master.
1951 					 */
1952 					continue;
1953 				}
1954 			}
1955 			ret = -1;
1956 			if (nl_route >= 0)
1957 				ret = mlx5_nl_switch_info
1958 					       (nl_route,
1959 						list[ns].ifindex,
1960 						&list[ns].info);
1961 			if (ret || (!list[ns].info.representor &&
1962 				    !list[ns].info.master)) {
1963 				/*
1964 				 * We failed to recognize representors with
1965 				 * Netlink, let's try to perform the task
1966 				 * with sysfs.
1967 				 */
1968 				ret =  mlx5_sysfs_switch_info
1969 						(list[ns].ifindex,
1970 						 &list[ns].info);
1971 			}
1972 			if (!ret && (list[ns].info.representor ^
1973 				     list[ns].info.master)) {
1974 				ns++;
1975 			} else if ((nd == 1) &&
1976 				   !list[ns].info.representor &&
1977 				   !list[ns].info.master) {
1978 				/*
1979 				 * Single IB device with
1980 				 * one physical port and
1981 				 * attached network device.
1982 				 * May be SRIOV is not enabled
1983 				 * or there is no representors.
1984 				 */
1985 				DRV_LOG(INFO, "no E-Switch support detected");
1986 				ns++;
1987 				break;
1988 			}
1989 		}
1990 		if (!ns) {
1991 			DRV_LOG(ERR,
1992 				"unable to recognize master/representors"
1993 				" on the multiple IB devices");
1994 			rte_errno = ENOENT;
1995 			ret = -rte_errno;
1996 			goto exit;
1997 		}
1998 	}
1999 	MLX5_ASSERT(ns);
2000 	/*
2001 	 * Sort list to probe devices in natural order for users convenience
2002 	 * (i.e. master first, then representors from lowest to highest ID).
2003 	 */
2004 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2005 	/* Device specific configuration. */
2006 	switch (pci_dev->id.device_id) {
2007 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2008 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2009 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2010 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2011 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2012 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2013 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2014 		dev_config_vf = 1;
2015 		break;
2016 	default:
2017 		dev_config_vf = 0;
2018 		break;
2019 	}
2020 	for (i = 0; i != ns; ++i) {
2021 		uint32_t restore;
2022 
2023 		/* Default configuration. */
2024 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2025 		dev_config.vf = dev_config_vf;
2026 		dev_config.mps = MLX5_ARG_UNSET;
2027 		dev_config.dbnc = MLX5_ARG_UNSET;
2028 		dev_config.rx_vec_en = 1;
2029 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2030 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2031 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2032 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2033 		dev_config.vf_nl_en = 1;
2034 		dev_config.mr_ext_memseg_en = 1;
2035 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2036 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2037 		dev_config.dv_esw_en = 1;
2038 		dev_config.dv_flow_en = 1;
2039 		dev_config.decap_en = 1;
2040 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2041 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2042 						 &list[i],
2043 						 &dev_config);
2044 		if (!list[i].eth_dev) {
2045 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2046 				break;
2047 			/* Device is disabled or already spawned. Ignore it. */
2048 			continue;
2049 		}
2050 		restore = list[i].eth_dev->data->dev_flags;
2051 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2052 		/* Restore non-PCI flags cleared by the above call. */
2053 		list[i].eth_dev->data->dev_flags |= restore;
2054 		rte_eth_dev_probing_finish(list[i].eth_dev);
2055 	}
2056 	if (i != ns) {
2057 		DRV_LOG(ERR,
2058 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2059 			" encountering an error: %s",
2060 			pci_dev->addr.domain, pci_dev->addr.bus,
2061 			pci_dev->addr.devid, pci_dev->addr.function,
2062 			strerror(rte_errno));
2063 		ret = -rte_errno;
2064 		/* Roll back. */
2065 		while (i--) {
2066 			if (!list[i].eth_dev)
2067 				continue;
2068 			mlx5_dev_close(list[i].eth_dev);
2069 			/* mac_addrs must not be freed because in dev_private */
2070 			list[i].eth_dev->data->mac_addrs = NULL;
2071 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2072 		}
2073 		/* Restore original error. */
2074 		rte_errno = -ret;
2075 	} else {
2076 		ret = 0;
2077 	}
2078 exit:
2079 	/*
2080 	 * Do the routine cleanup:
2081 	 * - close opened Netlink sockets
2082 	 * - free allocated spawn data array
2083 	 * - free the Infiniband device list
2084 	 */
2085 	if (nl_rdma >= 0)
2086 		close(nl_rdma);
2087 	if (nl_route >= 0)
2088 		close(nl_route);
2089 	if (list)
2090 		mlx5_free(list);
2091 	MLX5_ASSERT(ibv_list);
2092 	mlx5_glue->free_device_list(ibv_list);
2093 	return ret;
2094 }
2095 
2096 static int
2097 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2098 {
2099 	char *env;
2100 	int value;
2101 
2102 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2103 	/* Get environment variable to store. */
2104 	env = getenv(MLX5_SHUT_UP_BF);
2105 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2106 	if (config->dbnc == MLX5_ARG_UNSET)
2107 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2108 	else
2109 		setenv(MLX5_SHUT_UP_BF,
2110 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2111 	return value;
2112 }
2113 
2114 static void
2115 mlx5_restore_doorbell_mapping_env(int value)
2116 {
2117 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2118 	/* Restore the original environment variable state. */
2119 	if (value == MLX5_ARG_UNSET)
2120 		unsetenv(MLX5_SHUT_UP_BF);
2121 	else
2122 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2123 }
2124 
2125 /**
2126  * Extract pdn of PD object using DV API.
2127  *
2128  * @param[in] pd
2129  *   Pointer to the verbs PD object.
2130  * @param[out] pdn
2131  *   Pointer to the PD object number variable.
2132  *
2133  * @return
2134  *   0 on success, error value otherwise.
2135  */
2136 int
2137 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2138 {
2139 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2140 	struct mlx5dv_obj obj;
2141 	struct mlx5dv_pd pd_info;
2142 	int ret = 0;
2143 
2144 	obj.pd.in = pd;
2145 	obj.pd.out = &pd_info;
2146 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2147 	if (ret) {
2148 		DRV_LOG(DEBUG, "Fail to get PD object info");
2149 		return ret;
2150 	}
2151 	*pdn = pd_info.pdn;
2152 	return 0;
2153 #else
2154 	(void)pd;
2155 	(void)pdn;
2156 	return -ENOTSUP;
2157 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2158 }
2159 
2160 /**
2161  * Function API to open IB device.
2162  *
2163  * This function calls the Linux glue APIs to open a device.
2164  *
2165  * @param[in] spawn
2166  *   Pointer to the IB device attributes (name, port, etc).
2167  * @param[out] config
2168  *   Pointer to device configuration structure.
2169  * @param[out] sh
2170  *   Pointer to shared context structure.
2171  *
2172  * @return
2173  *   0 on success, a positive error value otherwise.
2174  */
2175 int
2176 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2177 		     const struct mlx5_dev_config *config,
2178 		     struct mlx5_dev_ctx_shared *sh)
2179 {
2180 	int dbmap_env;
2181 	int err = 0;
2182 
2183 	sh->numa_node = spawn->pci_dev->device.numa_node;
2184 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2185 	/*
2186 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2187 	 * before the device creation. The rdma_core library
2188 	 * checks the variable at device creation and
2189 	 * stores the result internally.
2190 	 */
2191 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2192 	/* Try to open IB device with DV first, then usual Verbs. */
2193 	errno = 0;
2194 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2195 	if (sh->ctx) {
2196 		sh->devx = 1;
2197 		DRV_LOG(DEBUG, "DevX is supported");
2198 		/* The device is created, no need for environment. */
2199 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2200 	} else {
2201 		/* The environment variable is still configured. */
2202 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2203 		err = errno ? errno : ENODEV;
2204 		/*
2205 		 * The environment variable is not needed anymore,
2206 		 * all device creation attempts are completed.
2207 		 */
2208 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2209 		if (!sh->ctx)
2210 			return err;
2211 		DRV_LOG(DEBUG, "DevX is NOT supported");
2212 		err = 0;
2213 	}
2214 	return err;
2215 }
2216 
2217 /**
2218  * Install shared asynchronous device events handler.
2219  * This function is implemented to support event sharing
2220  * between multiple ports of single IB device.
2221  *
2222  * @param sh
2223  *   Pointer to mlx5_dev_ctx_shared object.
2224  */
2225 void
2226 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2227 {
2228 	int ret;
2229 	int flags;
2230 
2231 	sh->intr_handle.fd = -1;
2232 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2233 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2234 		    F_SETFL, flags | O_NONBLOCK);
2235 	if (ret) {
2236 		DRV_LOG(INFO, "failed to change file descriptor async event"
2237 			" queue");
2238 	} else {
2239 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2240 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2241 		if (rte_intr_callback_register(&sh->intr_handle,
2242 					mlx5_dev_interrupt_handler, sh)) {
2243 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2244 			sh->intr_handle.fd = -1;
2245 		}
2246 	}
2247 	if (sh->devx) {
2248 #ifdef HAVE_IBV_DEVX_ASYNC
2249 		sh->intr_handle_devx.fd = -1;
2250 		sh->devx_comp =
2251 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2252 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2253 		if (!devx_comp) {
2254 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2255 			return;
2256 		}
2257 		flags = fcntl(devx_comp->fd, F_GETFL);
2258 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2259 		if (ret) {
2260 			DRV_LOG(INFO, "failed to change file descriptor"
2261 				" devx comp");
2262 			return;
2263 		}
2264 		sh->intr_handle_devx.fd = devx_comp->fd;
2265 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2266 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2267 					mlx5_dev_interrupt_handler_devx, sh)) {
2268 			DRV_LOG(INFO, "Fail to install the devx shared"
2269 				" interrupt.");
2270 			sh->intr_handle_devx.fd = -1;
2271 		}
2272 #endif /* HAVE_IBV_DEVX_ASYNC */
2273 	}
2274 }
2275 
2276 /**
2277  * Uninstall shared asynchronous device events handler.
2278  * This function is implemented to support event sharing
2279  * between multiple ports of single IB device.
2280  *
2281  * @param dev
2282  *   Pointer to mlx5_dev_ctx_shared object.
2283  */
2284 void
2285 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2286 {
2287 	if (sh->intr_handle.fd >= 0)
2288 		mlx5_intr_callback_unregister(&sh->intr_handle,
2289 					      mlx5_dev_interrupt_handler, sh);
2290 #ifdef HAVE_IBV_DEVX_ASYNC
2291 	if (sh->intr_handle_devx.fd >= 0)
2292 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2293 				  mlx5_dev_interrupt_handler_devx, sh);
2294 	if (sh->devx_comp)
2295 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2296 #endif
2297 }
2298 
2299 /**
2300  * Read statistics by a named counter.
2301  *
2302  * @param[in] priv
2303  *   Pointer to the private device data structure.
2304  * @param[in] ctr_name
2305  *   Pointer to the name of the statistic counter to read
2306  * @param[out] stat
2307  *   Pointer to read statistic value.
2308  * @return
2309  *   0 on success and stat is valud, 1 if failed to read the value
2310  *   rte_errno is set.
2311  *
2312  */
2313 int
2314 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2315 		      uint64_t *stat)
2316 {
2317 	int fd;
2318 
2319 	if (priv->sh) {
2320 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2321 		      priv->sh->ibdev_path,
2322 		      priv->dev_port,
2323 		      ctr_name);
2324 		fd = open(path, O_RDONLY);
2325 		/*
2326 		 * in switchdev the file location is not per port
2327 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2328 		 */
2329 		if (fd == -1) {
2330 			MKSTR(path1, "%s/hw_counters/%s",
2331 			      priv->sh->ibdev_path,
2332 			      ctr_name);
2333 			fd = open(path1, O_RDONLY);
2334 		}
2335 		if (fd != -1) {
2336 			char buf[21] = {'\0'};
2337 			ssize_t n = read(fd, buf, sizeof(buf));
2338 
2339 			close(fd);
2340 			if (n != -1) {
2341 				*stat = strtoull(buf, NULL, 10);
2342 				return 0;
2343 			}
2344 		}
2345 	}
2346 	*stat = 0;
2347 	return 1;
2348 }
2349 
2350 /**
2351  * Set the reg_mr and dereg_mr call backs
2352  *
2353  * @param reg_mr_cb[out]
2354  *   Pointer to reg_mr func
2355  * @param dereg_mr_cb[out]
2356  *   Pointer to dereg_mr func
2357  *
2358  */
2359 void
2360 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2361 		      mlx5_dereg_mr_t *dereg_mr_cb)
2362 {
2363 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2364 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2365 }
2366 
2367 /**
2368  * Remove a MAC address from device
2369  *
2370  * @param dev
2371  *   Pointer to Ethernet device structure.
2372  * @param index
2373  *   MAC address index.
2374  */
2375 void
2376 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2377 {
2378 	struct mlx5_priv *priv = dev->data->dev_private;
2379 	const int vf = priv->config.vf;
2380 
2381 	if (vf)
2382 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2383 					mlx5_ifindex(dev), priv->mac_own,
2384 					&dev->data->mac_addrs[index], index);
2385 }
2386 
2387 /**
2388  * Adds a MAC address to the device
2389  *
2390  * @param dev
2391  *   Pointer to Ethernet device structure.
2392  * @param mac_addr
2393  *   MAC address to register.
2394  * @param index
2395  *   MAC address index.
2396  *
2397  * @return
2398  *   0 on success, a negative errno value otherwise
2399  */
2400 int
2401 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2402 		     uint32_t index)
2403 {
2404 	struct mlx5_priv *priv = dev->data->dev_private;
2405 	const int vf = priv->config.vf;
2406 	int ret = 0;
2407 
2408 	if (vf)
2409 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2410 					   mlx5_ifindex(dev), priv->mac_own,
2411 					   mac, index);
2412 	return ret;
2413 }
2414 
2415 /**
2416  * Modify a VF MAC address
2417  *
2418  * @param priv
2419  *   Pointer to device private data.
2420  * @param mac_addr
2421  *   MAC address to modify into.
2422  * @param iface_idx
2423  *   Net device interface index
2424  * @param vf_index
2425  *   VF index
2426  *
2427  * @return
2428  *   0 on success, a negative errno value otherwise
2429  */
2430 int
2431 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2432 			   unsigned int iface_idx,
2433 			   struct rte_ether_addr *mac_addr,
2434 			   int vf_index)
2435 {
2436 	return mlx5_nl_vf_mac_addr_modify
2437 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2438 }
2439 
2440 /**
2441  * Set device promiscuous mode
2442  *
2443  * @param dev
2444  *   Pointer to Ethernet device structure.
2445  * @param enable
2446  *   0 - promiscuous is disabled, otherwise - enabled
2447  *
2448  * @return
2449  *   0 on success, a negative error value otherwise
2450  */
2451 int
2452 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2453 {
2454 	struct mlx5_priv *priv = dev->data->dev_private;
2455 
2456 	return mlx5_nl_promisc(priv->nl_socket_route,
2457 			       mlx5_ifindex(dev), !!enable);
2458 }
2459 
2460 /**
2461  * Set device promiscuous mode
2462  *
2463  * @param dev
2464  *   Pointer to Ethernet device structure.
2465  * @param enable
2466  *   0 - all multicase is disabled, otherwise - enabled
2467  *
2468  * @return
2469  *   0 on success, a negative error value otherwise
2470  */
2471 int
2472 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2473 {
2474 	struct mlx5_priv *priv = dev->data->dev_private;
2475 
2476 	return mlx5_nl_allmulti(priv->nl_socket_route,
2477 				mlx5_ifindex(dev), !!enable);
2478 }
2479 
2480 /**
2481  * Flush device MAC addresses
2482  *
2483  * @param dev
2484  *   Pointer to Ethernet device structure.
2485  *
2486  */
2487 void
2488 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2489 {
2490 	struct mlx5_priv *priv = dev->data->dev_private;
2491 
2492 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2493 			       dev->data->mac_addrs,
2494 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2495 }
2496 
2497 const struct eth_dev_ops mlx5_os_dev_ops = {
2498 	.dev_configure = mlx5_dev_configure,
2499 	.dev_start = mlx5_dev_start,
2500 	.dev_stop = mlx5_dev_stop,
2501 	.dev_set_link_down = mlx5_set_link_down,
2502 	.dev_set_link_up = mlx5_set_link_up,
2503 	.dev_close = mlx5_dev_close,
2504 	.promiscuous_enable = mlx5_promiscuous_enable,
2505 	.promiscuous_disable = mlx5_promiscuous_disable,
2506 	.allmulticast_enable = mlx5_allmulticast_enable,
2507 	.allmulticast_disable = mlx5_allmulticast_disable,
2508 	.link_update = mlx5_link_update,
2509 	.stats_get = mlx5_stats_get,
2510 	.stats_reset = mlx5_stats_reset,
2511 	.xstats_get = mlx5_xstats_get,
2512 	.xstats_reset = mlx5_xstats_reset,
2513 	.xstats_get_names = mlx5_xstats_get_names,
2514 	.fw_version_get = mlx5_fw_version_get,
2515 	.dev_infos_get = mlx5_dev_infos_get,
2516 	.read_clock = mlx5_txpp_read_clock,
2517 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2518 	.vlan_filter_set = mlx5_vlan_filter_set,
2519 	.rx_queue_setup = mlx5_rx_queue_setup,
2520 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2521 	.tx_queue_setup = mlx5_tx_queue_setup,
2522 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2523 	.rx_queue_release = mlx5_rx_queue_release,
2524 	.tx_queue_release = mlx5_tx_queue_release,
2525 	.rx_queue_start = mlx5_rx_queue_start,
2526 	.rx_queue_stop = mlx5_rx_queue_stop,
2527 	.tx_queue_start = mlx5_tx_queue_start,
2528 	.tx_queue_stop = mlx5_tx_queue_stop,
2529 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2530 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2531 	.mac_addr_remove = mlx5_mac_addr_remove,
2532 	.mac_addr_add = mlx5_mac_addr_add,
2533 	.mac_addr_set = mlx5_mac_addr_set,
2534 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2535 	.mtu_set = mlx5_dev_set_mtu,
2536 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2537 	.vlan_offload_set = mlx5_vlan_offload_set,
2538 	.reta_update = mlx5_dev_rss_reta_update,
2539 	.reta_query = mlx5_dev_rss_reta_query,
2540 	.rss_hash_update = mlx5_rss_hash_update,
2541 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2542 	.filter_ctrl = mlx5_dev_filter_ctrl,
2543 	.rxq_info_get = mlx5_rxq_info_get,
2544 	.txq_info_get = mlx5_txq_info_get,
2545 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2546 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2547 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2548 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2549 	.is_removed = mlx5_is_removed,
2550 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2551 	.get_module_info = mlx5_get_module_info,
2552 	.get_module_eeprom = mlx5_get_module_eeprom,
2553 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2554 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2555 };
2556 
2557 /* Available operations from secondary process. */
2558 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2559 	.stats_get = mlx5_stats_get,
2560 	.stats_reset = mlx5_stats_reset,
2561 	.xstats_get = mlx5_xstats_get,
2562 	.xstats_reset = mlx5_xstats_reset,
2563 	.xstats_get_names = mlx5_xstats_get_names,
2564 	.fw_version_get = mlx5_fw_version_get,
2565 	.dev_infos_get = mlx5_dev_infos_get,
2566 	.read_clock = mlx5_txpp_read_clock,
2567 	.rx_queue_start = mlx5_rx_queue_start,
2568 	.rx_queue_stop = mlx5_rx_queue_stop,
2569 	.tx_queue_start = mlx5_tx_queue_start,
2570 	.tx_queue_stop = mlx5_tx_queue_stop,
2571 	.rxq_info_get = mlx5_rxq_info_get,
2572 	.txq_info_get = mlx5_txq_info_get,
2573 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2574 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2575 	.get_module_info = mlx5_get_module_info,
2576 	.get_module_eeprom = mlx5_get_module_eeprom,
2577 };
2578 
2579 /* Available operations in flow isolated mode. */
2580 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2581 	.dev_configure = mlx5_dev_configure,
2582 	.dev_start = mlx5_dev_start,
2583 	.dev_stop = mlx5_dev_stop,
2584 	.dev_set_link_down = mlx5_set_link_down,
2585 	.dev_set_link_up = mlx5_set_link_up,
2586 	.dev_close = mlx5_dev_close,
2587 	.promiscuous_enable = mlx5_promiscuous_enable,
2588 	.promiscuous_disable = mlx5_promiscuous_disable,
2589 	.allmulticast_enable = mlx5_allmulticast_enable,
2590 	.allmulticast_disable = mlx5_allmulticast_disable,
2591 	.link_update = mlx5_link_update,
2592 	.stats_get = mlx5_stats_get,
2593 	.stats_reset = mlx5_stats_reset,
2594 	.xstats_get = mlx5_xstats_get,
2595 	.xstats_reset = mlx5_xstats_reset,
2596 	.xstats_get_names = mlx5_xstats_get_names,
2597 	.fw_version_get = mlx5_fw_version_get,
2598 	.dev_infos_get = mlx5_dev_infos_get,
2599 	.read_clock = mlx5_txpp_read_clock,
2600 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2601 	.vlan_filter_set = mlx5_vlan_filter_set,
2602 	.rx_queue_setup = mlx5_rx_queue_setup,
2603 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2604 	.tx_queue_setup = mlx5_tx_queue_setup,
2605 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2606 	.rx_queue_release = mlx5_rx_queue_release,
2607 	.tx_queue_release = mlx5_tx_queue_release,
2608 	.rx_queue_start = mlx5_rx_queue_start,
2609 	.rx_queue_stop = mlx5_rx_queue_stop,
2610 	.tx_queue_start = mlx5_tx_queue_start,
2611 	.tx_queue_stop = mlx5_tx_queue_stop,
2612 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2613 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2614 	.mac_addr_remove = mlx5_mac_addr_remove,
2615 	.mac_addr_add = mlx5_mac_addr_add,
2616 	.mac_addr_set = mlx5_mac_addr_set,
2617 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2618 	.mtu_set = mlx5_dev_set_mtu,
2619 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2620 	.vlan_offload_set = mlx5_vlan_offload_set,
2621 	.filter_ctrl = mlx5_dev_filter_ctrl,
2622 	.rxq_info_get = mlx5_rxq_info_get,
2623 	.txq_info_get = mlx5_txq_info_get,
2624 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2625 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2626 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2627 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2628 	.is_removed = mlx5_is_removed,
2629 	.get_module_info = mlx5_get_module_info,
2630 	.get_module_eeprom = mlx5_get_module_eeprom,
2631 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2632 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2633 };
2634