xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 420bbdae89f29744847a166b831c675ad5affd2a)
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 <sys/mman.h>
14 #include <linux/rtnetlink.h>
15 #include <linux/sockios.h>
16 #include <linux/ethtool.h>
17 #include <fcntl.h>
18 
19 /* Verbs header. */
20 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
21 #ifdef PEDANTIC
22 #pragma GCC diagnostic ignored "-Wpedantic"
23 #endif
24 #include <infiniband/verbs.h>
25 #ifdef PEDANTIC
26 #pragma GCC diagnostic error "-Wpedantic"
27 #endif
28 
29 #include <rte_malloc.h>
30 #include <rte_ethdev_driver.h>
31 #include <rte_ethdev_pci.h>
32 #include <rte_pci.h>
33 #include <rte_bus_pci.h>
34 #include <rte_common.h>
35 #include <rte_kvargs.h>
36 #include <rte_rwlock.h>
37 #include <rte_spinlock.h>
38 #include <rte_string_fns.h>
39 #include <rte_alarm.h>
40 
41 #include <mlx5_glue.h>
42 #include <mlx5_devx_cmds.h>
43 #include <mlx5_common.h>
44 #include <mlx5_common_mp.h>
45 #include <mlx5_common_mr.h>
46 
47 #include "mlx5_defs.h"
48 #include "mlx5.h"
49 #include "mlx5_common_os.h"
50 #include "mlx5_utils.h"
51 #include "mlx5_rxtx.h"
52 #include "mlx5_autoconf.h"
53 #include "mlx5_mr.h"
54 #include "mlx5_flow.h"
55 #include "rte_pmd_mlx5.h"
56 #include "mlx5_verbs.h"
57 
58 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
59 
60 #ifndef HAVE_IBV_MLX5_MOD_MPW
61 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
62 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
63 #endif
64 
65 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
66 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
67 #endif
68 
69 /**
70  * Get mlx5 device attributes. The glue function query_device_ex() is called
71  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
72  * device attributes from the glue out parameter.
73  *
74  * @param dev
75  *   Pointer to ibv context.
76  *
77  * @param device_attr
78  *   Pointer to mlx5 device attributes.
79  *
80  * @return
81  *   0 on success, non zero error number otherwise
82  */
83 int
84 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
85 {
86 	int err;
87 	struct ibv_device_attr_ex attr_ex;
88 	memset(device_attr, 0, sizeof(*device_attr));
89 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
90 	if (err)
91 		return err;
92 
93 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
94 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
95 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
96 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
97 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
98 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
99 	device_attr->max_rwq_indirection_table_size =
100 		attr_ex.rss_caps.max_rwq_indirection_table_size;
101 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
102 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
103 
104 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
105 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
106 	if (err)
107 		return err;
108 
109 	device_attr->flags = dv_attr.flags;
110 	device_attr->comp_mask = dv_attr.comp_mask;
111 #ifdef HAVE_IBV_MLX5_MOD_SWP
112 	device_attr->sw_parsing_offloads =
113 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
114 #endif
115 	device_attr->min_single_stride_log_num_of_bytes =
116 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
117 	device_attr->max_single_stride_log_num_of_bytes =
118 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
119 	device_attr->min_single_wqe_log_num_of_strides =
120 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
121 	device_attr->max_single_wqe_log_num_of_strides =
122 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
123 	device_attr->stride_supported_qpts =
124 		dv_attr.striding_rq_caps.supported_qpts;
125 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
126 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
127 #endif
128 
129 	return err;
130 }
131 
132 /**
133  * Verbs callback to allocate a memory. This function should allocate the space
134  * according to the size provided residing inside a huge page.
135  * Please note that all allocation must respect the alignment from libmlx5
136  * (i.e. currently sysconf(_SC_PAGESIZE)).
137  *
138  * @param[in] size
139  *   The size in bytes of the memory to allocate.
140  * @param[in] data
141  *   A pointer to the callback data.
142  *
143  * @return
144  *   Allocated buffer, NULL otherwise and rte_errno is set.
145  */
146 static void *
147 mlx5_alloc_verbs_buf(size_t size, void *data)
148 {
149 	struct mlx5_priv *priv = data;
150 	void *ret;
151 	size_t alignment = sysconf(_SC_PAGESIZE);
152 	unsigned int socket = SOCKET_ID_ANY;
153 
154 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
155 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
156 
157 		socket = ctrl->socket;
158 	} else if (priv->verbs_alloc_ctx.type ==
159 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
160 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
161 
162 		socket = ctrl->socket;
163 	}
164 	MLX5_ASSERT(data != NULL);
165 	ret = rte_malloc_socket(__func__, size, alignment, socket);
166 	if (!ret && size)
167 		rte_errno = ENOMEM;
168 	return ret;
169 }
170 
171 /**
172  * Verbs callback to free a memory.
173  *
174  * @param[in] ptr
175  *   A pointer to the memory to free.
176  * @param[in] data
177  *   A pointer to the callback data.
178  */
179 static void
180 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
181 {
182 	MLX5_ASSERT(data != NULL);
183 	rte_free(ptr);
184 }
185 
186 /**
187  * Initialize DR related data within private structure.
188  * Routine checks the reference counter and does actual
189  * resources creation/initialization only if counter is zero.
190  *
191  * @param[in] priv
192  *   Pointer to the private device data structure.
193  *
194  * @return
195  *   Zero on success, positive error code otherwise.
196  */
197 static int
198 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
199 {
200 	struct mlx5_dev_ctx_shared *sh = priv->sh;
201 	char s[MLX5_HLIST_NAMESIZE];
202 	int err = 0;
203 
204 	if (!sh->flow_tbls)
205 		err = mlx5_alloc_table_hash_list(priv);
206 	else
207 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
208 			(void *)sh->flow_tbls);
209 	if (err)
210 		return err;
211 	/* Create tags hash list table. */
212 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
213 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
214 	if (!sh->tag_table) {
215 		DRV_LOG(ERR, "tags with hash creation failed.");
216 		err = ENOMEM;
217 		goto error;
218 	}
219 #ifdef HAVE_MLX5DV_DR
220 	void *domain;
221 
222 	if (sh->dv_refcnt) {
223 		/* Shared DV/DR structures is already initialized. */
224 		sh->dv_refcnt++;
225 		priv->dr_shared = 1;
226 		return 0;
227 	}
228 	/* Reference counter is zero, we should initialize structures. */
229 	domain = mlx5_glue->dr_create_domain(sh->ctx,
230 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
231 	if (!domain) {
232 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
233 		err = errno;
234 		goto error;
235 	}
236 	sh->rx_domain = domain;
237 	domain = mlx5_glue->dr_create_domain(sh->ctx,
238 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
239 	if (!domain) {
240 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
241 		err = errno;
242 		goto error;
243 	}
244 	pthread_mutex_init(&sh->dv_mutex, NULL);
245 	sh->tx_domain = domain;
246 #ifdef HAVE_MLX5DV_DR_ESWITCH
247 	if (priv->config.dv_esw_en) {
248 		domain  = mlx5_glue->dr_create_domain
249 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
250 		if (!domain) {
251 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
252 			err = errno;
253 			goto error;
254 		}
255 		sh->fdb_domain = domain;
256 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
257 	}
258 #endif
259 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
260 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
261 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
262 		if (sh->fdb_domain)
263 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
264 	}
265 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
266 #endif /* HAVE_MLX5DV_DR */
267 	sh->dv_refcnt++;
268 	priv->dr_shared = 1;
269 	return 0;
270 error:
271 	/* Rollback the created objects. */
272 	if (sh->rx_domain) {
273 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
274 		sh->rx_domain = NULL;
275 	}
276 	if (sh->tx_domain) {
277 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
278 		sh->tx_domain = NULL;
279 	}
280 	if (sh->fdb_domain) {
281 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
282 		sh->fdb_domain = NULL;
283 	}
284 	if (sh->esw_drop_action) {
285 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
286 		sh->esw_drop_action = NULL;
287 	}
288 	if (sh->pop_vlan_action) {
289 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
290 		sh->pop_vlan_action = NULL;
291 	}
292 	if (sh->tag_table) {
293 		/* tags should be destroyed with flow before. */
294 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
295 		sh->tag_table = NULL;
296 	}
297 	mlx5_free_table_hash_list(priv);
298 	return err;
299 }
300 
301 /**
302  * Destroy DR related data within private structure.
303  *
304  * @param[in] priv
305  *   Pointer to the private device data structure.
306  */
307 void
308 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
309 {
310 	struct mlx5_dev_ctx_shared *sh;
311 
312 	if (!priv->dr_shared)
313 		return;
314 	priv->dr_shared = 0;
315 	sh = priv->sh;
316 	MLX5_ASSERT(sh);
317 #ifdef HAVE_MLX5DV_DR
318 	MLX5_ASSERT(sh->dv_refcnt);
319 	if (sh->dv_refcnt && --sh->dv_refcnt)
320 		return;
321 	if (sh->rx_domain) {
322 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
323 		sh->rx_domain = NULL;
324 	}
325 	if (sh->tx_domain) {
326 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
327 		sh->tx_domain = NULL;
328 	}
329 #ifdef HAVE_MLX5DV_DR_ESWITCH
330 	if (sh->fdb_domain) {
331 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
332 		sh->fdb_domain = NULL;
333 	}
334 	if (sh->esw_drop_action) {
335 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
336 		sh->esw_drop_action = NULL;
337 	}
338 #endif
339 	if (sh->pop_vlan_action) {
340 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
341 		sh->pop_vlan_action = NULL;
342 	}
343 	pthread_mutex_destroy(&sh->dv_mutex);
344 #endif /* HAVE_MLX5DV_DR */
345 	if (sh->tag_table) {
346 		/* tags should be destroyed with flow before. */
347 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
348 		sh->tag_table = NULL;
349 	}
350 	mlx5_free_table_hash_list(priv);
351 }
352 
353 /**
354  * Spawn an Ethernet device from Verbs information.
355  *
356  * @param dpdk_dev
357  *   Backing DPDK device.
358  * @param spawn
359  *   Verbs device parameters (name, port, switch_info) to spawn.
360  * @param config
361  *   Device configuration parameters.
362  *
363  * @return
364  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
365  *   is set. The following errors are defined:
366  *
367  *   EBUSY: device is not supposed to be spawned.
368  *   EEXIST: device is already spawned
369  */
370 static struct rte_eth_dev *
371 mlx5_dev_spawn(struct rte_device *dpdk_dev,
372 	       struct mlx5_dev_spawn_data *spawn,
373 	       struct mlx5_dev_config config)
374 {
375 	const struct mlx5_switch_info *switch_info = &spawn->info;
376 	struct mlx5_dev_ctx_shared *sh = NULL;
377 	struct ibv_port_attr port_attr;
378 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
379 	struct rte_eth_dev *eth_dev = NULL;
380 	struct mlx5_priv *priv = NULL;
381 	int err = 0;
382 	unsigned int hw_padding = 0;
383 	unsigned int mps;
384 	unsigned int cqe_comp;
385 	unsigned int cqe_pad = 0;
386 	unsigned int tunnel_en = 0;
387 	unsigned int mpls_en = 0;
388 	unsigned int swp = 0;
389 	unsigned int mprq = 0;
390 	unsigned int mprq_min_stride_size_n = 0;
391 	unsigned int mprq_max_stride_size_n = 0;
392 	unsigned int mprq_min_stride_num_n = 0;
393 	unsigned int mprq_max_stride_num_n = 0;
394 	struct rte_ether_addr mac;
395 	char name[RTE_ETH_NAME_MAX_LEN];
396 	int own_domain_id = 0;
397 	uint16_t port_id;
398 	unsigned int i;
399 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
400 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
401 #endif
402 
403 	/* Determine if this port representor is supposed to be spawned. */
404 	if (switch_info->representor && dpdk_dev->devargs) {
405 		struct rte_eth_devargs eth_da;
406 
407 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
408 		if (err) {
409 			rte_errno = -err;
410 			DRV_LOG(ERR, "failed to process device arguments: %s",
411 				strerror(rte_errno));
412 			return NULL;
413 		}
414 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
415 			if (eth_da.representor_ports[i] ==
416 			    (uint16_t)switch_info->port_name)
417 				break;
418 		if (i == eth_da.nb_representor_ports) {
419 			rte_errno = EBUSY;
420 			return NULL;
421 		}
422 	}
423 	/* Build device name. */
424 	if (spawn->pf_bond <  0) {
425 		/* Single device. */
426 		if (!switch_info->representor)
427 			strlcpy(name, dpdk_dev->name, sizeof(name));
428 		else
429 			snprintf(name, sizeof(name), "%s_representor_%u",
430 				 dpdk_dev->name, switch_info->port_name);
431 	} else {
432 		/* Bonding device. */
433 		if (!switch_info->representor)
434 			snprintf(name, sizeof(name), "%s_%s",
435 				 dpdk_dev->name,
436 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
437 		else
438 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
439 				 dpdk_dev->name,
440 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
441 				 switch_info->port_name);
442 	}
443 	/* check if the device is already spawned */
444 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
445 		rte_errno = EEXIST;
446 		return NULL;
447 	}
448 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
449 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
450 		struct mlx5_mp_id mp_id;
451 
452 		eth_dev = rte_eth_dev_attach_secondary(name);
453 		if (eth_dev == NULL) {
454 			DRV_LOG(ERR, "can not attach rte ethdev");
455 			rte_errno = ENOMEM;
456 			return NULL;
457 		}
458 		eth_dev->device = dpdk_dev;
459 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
460 		err = mlx5_proc_priv_init(eth_dev);
461 		if (err)
462 			return NULL;
463 		mp_id.port_id = eth_dev->data->port_id;
464 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
465 		/* Receive command fd from primary process */
466 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
467 		if (err < 0)
468 			goto err_secondary;
469 		/* Remap UAR for Tx queues. */
470 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
471 		if (err)
472 			goto err_secondary;
473 		/*
474 		 * Ethdev pointer is still required as input since
475 		 * the primary device is not accessible from the
476 		 * secondary process.
477 		 */
478 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
479 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
480 		return eth_dev;
481 err_secondary:
482 		mlx5_dev_close(eth_dev);
483 		return NULL;
484 	}
485 	/*
486 	 * Some parameters ("tx_db_nc" in particularly) are needed in
487 	 * advance to create dv/verbs device context. We proceed the
488 	 * devargs here to get ones, and later proceed devargs again
489 	 * to override some hardware settings.
490 	 */
491 	err = mlx5_args(&config, dpdk_dev->devargs);
492 	if (err) {
493 		err = rte_errno;
494 		DRV_LOG(ERR, "failed to process device arguments: %s",
495 			strerror(rte_errno));
496 		goto error;
497 	}
498 	sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
499 	if (!sh)
500 		return NULL;
501 	config.devx = sh->devx;
502 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
503 	config.dest_tir = 1;
504 #endif
505 #ifdef HAVE_IBV_MLX5_MOD_SWP
506 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
507 #endif
508 	/*
509 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
510 	 * as all ConnectX-5 devices.
511 	 */
512 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
513 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
514 #endif
515 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
516 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
517 #endif
518 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
519 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
520 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
521 			DRV_LOG(DEBUG, "enhanced MPW is supported");
522 			mps = MLX5_MPW_ENHANCED;
523 		} else {
524 			DRV_LOG(DEBUG, "MPW is supported");
525 			mps = MLX5_MPW;
526 		}
527 	} else {
528 		DRV_LOG(DEBUG, "MPW isn't supported");
529 		mps = MLX5_MPW_DISABLED;
530 	}
531 #ifdef HAVE_IBV_MLX5_MOD_SWP
532 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
533 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
534 	DRV_LOG(DEBUG, "SWP support: %u", swp);
535 #endif
536 	config.swp = !!swp;
537 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
538 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
539 		struct mlx5dv_striding_rq_caps mprq_caps =
540 			dv_attr.striding_rq_caps;
541 
542 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
543 			mprq_caps.min_single_stride_log_num_of_bytes);
544 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
545 			mprq_caps.max_single_stride_log_num_of_bytes);
546 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
547 			mprq_caps.min_single_wqe_log_num_of_strides);
548 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
549 			mprq_caps.max_single_wqe_log_num_of_strides);
550 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
551 			mprq_caps.supported_qpts);
552 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
553 		mprq = 1;
554 		mprq_min_stride_size_n =
555 			mprq_caps.min_single_stride_log_num_of_bytes;
556 		mprq_max_stride_size_n =
557 			mprq_caps.max_single_stride_log_num_of_bytes;
558 		mprq_min_stride_num_n =
559 			mprq_caps.min_single_wqe_log_num_of_strides;
560 		mprq_max_stride_num_n =
561 			mprq_caps.max_single_wqe_log_num_of_strides;
562 	}
563 #endif
564 	if (RTE_CACHE_LINE_SIZE == 128 &&
565 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
566 		cqe_comp = 0;
567 	else
568 		cqe_comp = 1;
569 	config.cqe_comp = cqe_comp;
570 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
571 	/* Whether device supports 128B Rx CQE padding. */
572 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
573 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
574 #endif
575 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
576 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
577 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
578 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
579 			     (dv_attr.tunnel_offloads_caps &
580 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
581 			     (dv_attr.tunnel_offloads_caps &
582 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
583 	}
584 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
585 		tunnel_en ? "" : "not ");
586 #else
587 	DRV_LOG(WARNING,
588 		"tunnel offloading disabled due to old OFED/rdma-core version");
589 #endif
590 	config.tunnel_en = tunnel_en;
591 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
592 	mpls_en = ((dv_attr.tunnel_offloads_caps &
593 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
594 		   (dv_attr.tunnel_offloads_caps &
595 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
596 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
597 		mpls_en ? "" : "not ");
598 #else
599 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
600 		" old OFED/rdma-core version or firmware configuration");
601 #endif
602 	config.mpls_en = mpls_en;
603 	/* Check port status. */
604 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
605 	if (err) {
606 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
607 		goto error;
608 	}
609 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
610 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
611 		err = EINVAL;
612 		goto error;
613 	}
614 	if (port_attr.state != IBV_PORT_ACTIVE)
615 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
616 			mlx5_glue->port_state_str(port_attr.state),
617 			port_attr.state);
618 	/* Allocate private eth device data. */
619 	priv = rte_zmalloc("ethdev private structure",
620 			   sizeof(*priv),
621 			   RTE_CACHE_LINE_SIZE);
622 	if (priv == NULL) {
623 		DRV_LOG(ERR, "priv allocation failure");
624 		err = ENOMEM;
625 		goto error;
626 	}
627 	priv->sh = sh;
628 	priv->dev_port = spawn->phys_port;
629 	priv->pci_dev = spawn->pci_dev;
630 	priv->mtu = RTE_ETHER_MTU;
631 	priv->mp_id.port_id = port_id;
632 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
633 #ifndef RTE_ARCH_64
634 	/* Initialize UAR access locks for 32bit implementations. */
635 	rte_spinlock_init(&priv->uar_lock_cq);
636 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
637 		rte_spinlock_init(&priv->uar_lock[i]);
638 #endif
639 	/* Some internal functions rely on Netlink sockets, open them now. */
640 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
641 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
642 	priv->representor = !!switch_info->representor;
643 	priv->master = !!switch_info->master;
644 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
645 	priv->vport_meta_tag = 0;
646 	priv->vport_meta_mask = 0;
647 	priv->pf_bond = spawn->pf_bond;
648 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
649 	/*
650 	 * The DevX port query API is implemented. E-Switch may use
651 	 * either vport or reg_c[0] metadata register to match on
652 	 * vport index. The engaged part of metadata register is
653 	 * defined by mask.
654 	 */
655 	if (switch_info->representor || switch_info->master) {
656 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
657 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
658 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
659 						 &devx_port);
660 		if (err) {
661 			DRV_LOG(WARNING,
662 				"can't query devx port %d on device %s",
663 				spawn->phys_port,
664 				mlx5_os_get_dev_device_name(spawn->phys_dev));
665 			devx_port.comp_mask = 0;
666 		}
667 	}
668 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
669 		priv->vport_meta_tag = devx_port.reg_c_0.value;
670 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
671 		if (!priv->vport_meta_mask) {
672 			DRV_LOG(ERR, "vport zero mask for port %d"
673 				     " on bonding device %s",
674 				     spawn->phys_port,
675 				     mlx5_os_get_dev_device_name
676 							(spawn->phys_dev));
677 			err = ENOTSUP;
678 			goto error;
679 		}
680 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
681 			DRV_LOG(ERR, "invalid vport tag for port %d"
682 				     " on bonding device %s",
683 				     spawn->phys_port,
684 				     mlx5_os_get_dev_device_name
685 							(spawn->phys_dev));
686 			err = ENOTSUP;
687 			goto error;
688 		}
689 	}
690 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
691 		priv->vport_id = devx_port.vport_num;
692 	} else if (spawn->pf_bond >= 0) {
693 		DRV_LOG(ERR, "can't deduce vport index for port %d"
694 			     " on bonding device %s",
695 			     spawn->phys_port,
696 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
697 		err = ENOTSUP;
698 		goto error;
699 	} else {
700 		/* Suppose vport index in compatible way. */
701 		priv->vport_id = switch_info->representor ?
702 				 switch_info->port_name + 1 : -1;
703 	}
704 #else
705 	/*
706 	 * Kernel/rdma_core support single E-Switch per PF configurations
707 	 * only and vport_id field contains the vport index for
708 	 * associated VF, which is deduced from representor port name.
709 	 * For example, let's have the IB device port 10, it has
710 	 * attached network device eth0, which has port name attribute
711 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
712 	 * as 3 (2+1). This assigning schema should be changed if the
713 	 * multiple E-Switch instances per PF configurations or/and PCI
714 	 * subfunctions are added.
715 	 */
716 	priv->vport_id = switch_info->representor ?
717 			 switch_info->port_name + 1 : -1;
718 #endif
719 	/* representor_id field keeps the unmodified VF index. */
720 	priv->representor_id = switch_info->representor ?
721 			       switch_info->port_name : -1;
722 	/*
723 	 * Look for sibling devices in order to reuse their switch domain
724 	 * if any, otherwise allocate one.
725 	 */
726 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
727 		const struct mlx5_priv *opriv =
728 			rte_eth_devices[port_id].data->dev_private;
729 
730 		if (!opriv ||
731 		    opriv->sh != priv->sh ||
732 			opriv->domain_id ==
733 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
734 			continue;
735 		priv->domain_id = opriv->domain_id;
736 		break;
737 	}
738 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
739 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
740 		if (err) {
741 			err = rte_errno;
742 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
743 				strerror(rte_errno));
744 			goto error;
745 		}
746 		own_domain_id = 1;
747 	}
748 	/* Override some values set by hardware configuration. */
749 	mlx5_args(&config, dpdk_dev->devargs);
750 	err = mlx5_dev_check_sibling_config(priv, &config);
751 	if (err)
752 		goto error;
753 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
754 			    IBV_DEVICE_RAW_IP_CSUM);
755 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
756 		(config.hw_csum ? "" : "not "));
757 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
758 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
759 	DRV_LOG(DEBUG, "counters are not supported");
760 #endif
761 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
762 	if (config.dv_flow_en) {
763 		DRV_LOG(WARNING, "DV flow is not supported");
764 		config.dv_flow_en = 0;
765 	}
766 #endif
767 	config.ind_table_max_size =
768 		sh->device_attr.max_rwq_indirection_table_size;
769 	/*
770 	 * Remove this check once DPDK supports larger/variable
771 	 * indirection tables.
772 	 */
773 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
774 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
775 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
776 		config.ind_table_max_size);
777 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
778 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
779 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
780 		(config.hw_vlan_strip ? "" : "not "));
781 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
782 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
783 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
784 		(config.hw_fcs_strip ? "" : "not "));
785 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
786 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
787 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
788 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
789 			IBV_DEVICE_PCI_WRITE_END_PADDING);
790 #endif
791 	if (config.hw_padding && !hw_padding) {
792 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
793 		config.hw_padding = 0;
794 	} else if (config.hw_padding) {
795 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
796 	}
797 	config.tso = (sh->device_attr.max_tso > 0 &&
798 		      (sh->device_attr.tso_supported_qpts &
799 		       (1 << IBV_QPT_RAW_PACKET)));
800 	if (config.tso)
801 		config.tso_max_payload_sz = sh->device_attr.max_tso;
802 	/*
803 	 * MPW is disabled by default, while the Enhanced MPW is enabled
804 	 * by default.
805 	 */
806 	if (config.mps == MLX5_ARG_UNSET)
807 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
808 							  MLX5_MPW_DISABLED;
809 	else
810 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
811 	DRV_LOG(INFO, "%sMPS is %s",
812 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
813 		config.mps == MLX5_MPW ? "legacy " : "",
814 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
815 	if (config.cqe_comp && !cqe_comp) {
816 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
817 		config.cqe_comp = 0;
818 	}
819 	if (config.cqe_pad && !cqe_pad) {
820 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
821 		config.cqe_pad = 0;
822 	} else if (config.cqe_pad) {
823 		DRV_LOG(INFO, "Rx CQE padding is enabled");
824 	}
825 	if (config.devx) {
826 		priv->counter_fallback = 0;
827 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
828 		if (err) {
829 			err = -err;
830 			goto error;
831 		}
832 		if (!config.hca_attr.flow_counters_dump)
833 			priv->counter_fallback = 1;
834 #ifndef HAVE_IBV_DEVX_ASYNC
835 		priv->counter_fallback = 1;
836 #endif
837 		if (priv->counter_fallback)
838 			DRV_LOG(INFO, "Use fall-back DV counter management");
839 		/* Check for LRO support. */
840 		if (config.dest_tir && config.hca_attr.lro_cap &&
841 		    config.dv_flow_en) {
842 			/* TBD check tunnel lro caps. */
843 			config.lro.supported = config.hca_attr.lro_cap;
844 			DRV_LOG(DEBUG, "Device supports LRO");
845 			/*
846 			 * If LRO timeout is not configured by application,
847 			 * use the minimal supported value.
848 			 */
849 			if (!config.lro.timeout)
850 				config.lro.timeout =
851 				config.hca_attr.lro_timer_supported_periods[0];
852 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
853 				config.lro.timeout);
854 		}
855 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
856 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
857 		    config.dv_flow_en) {
858 			uint8_t reg_c_mask =
859 				config.hca_attr.qos.flow_meter_reg_c_ids;
860 			/*
861 			 * Meter needs two REG_C's for color match and pre-sfx
862 			 * flow match. Here get the REG_C for color match.
863 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
864 			 */
865 			reg_c_mask &= 0xfc;
866 			if (__builtin_popcount(reg_c_mask) < 1) {
867 				priv->mtr_en = 0;
868 				DRV_LOG(WARNING, "No available register for"
869 					" meter.");
870 			} else {
871 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
872 						      REG_C_0;
873 				priv->mtr_en = 1;
874 				priv->mtr_reg_share =
875 				      config.hca_attr.qos.flow_meter_reg_share;
876 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
877 					priv->mtr_color_reg);
878 			}
879 		}
880 #endif
881 	}
882 	if (config.mprq.enabled && mprq) {
883 		if (config.mprq.stride_num_n &&
884 		    (config.mprq.stride_num_n > mprq_max_stride_num_n ||
885 		     config.mprq.stride_num_n < mprq_min_stride_num_n)) {
886 			config.mprq.stride_num_n =
887 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
888 						mprq_min_stride_num_n),
889 					mprq_max_stride_num_n);
890 			DRV_LOG(WARNING,
891 				"the number of strides"
892 				" for Multi-Packet RQ is out of range,"
893 				" setting default value (%u)",
894 				1 << config.mprq.stride_num_n);
895 		}
896 		if (config.mprq.stride_size_n &&
897 		    (config.mprq.stride_size_n > mprq_max_stride_size_n ||
898 		     config.mprq.stride_size_n < mprq_min_stride_size_n)) {
899 			config.mprq.stride_size_n =
900 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
901 						mprq_min_stride_size_n),
902 					mprq_max_stride_size_n);
903 			DRV_LOG(WARNING,
904 				"the size of a stride"
905 				" for Multi-Packet RQ is out of range,"
906 				" setting default value (%u)",
907 				1 << config.mprq.stride_size_n);
908 		}
909 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
910 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
911 	} else if (config.mprq.enabled && !mprq) {
912 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
913 		config.mprq.enabled = 0;
914 	}
915 	if (config.max_dump_files_num == 0)
916 		config.max_dump_files_num = 128;
917 	eth_dev = rte_eth_dev_allocate(name);
918 	if (eth_dev == NULL) {
919 		DRV_LOG(ERR, "can not allocate rte ethdev");
920 		err = ENOMEM;
921 		goto error;
922 	}
923 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
924 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
925 	if (priv->representor) {
926 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
927 		eth_dev->data->representor_id = priv->representor_id;
928 	}
929 	/*
930 	 * Store associated network device interface index. This index
931 	 * is permanent throughout the lifetime of device. So, we may store
932 	 * the ifindex here and use the cached value further.
933 	 */
934 	MLX5_ASSERT(spawn->ifindex);
935 	priv->if_index = spawn->ifindex;
936 	eth_dev->data->dev_private = priv;
937 	priv->dev_data = eth_dev->data;
938 	eth_dev->data->mac_addrs = priv->mac;
939 	eth_dev->device = dpdk_dev;
940 	/* Configure the first MAC address by default. */
941 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
942 		DRV_LOG(ERR,
943 			"port %u cannot get MAC address, is mlx5_en"
944 			" loaded? (errno: %s)",
945 			eth_dev->data->port_id, strerror(rte_errno));
946 		err = ENODEV;
947 		goto error;
948 	}
949 	DRV_LOG(INFO,
950 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
951 		eth_dev->data->port_id,
952 		mac.addr_bytes[0], mac.addr_bytes[1],
953 		mac.addr_bytes[2], mac.addr_bytes[3],
954 		mac.addr_bytes[4], mac.addr_bytes[5]);
955 #ifdef RTE_LIBRTE_MLX5_DEBUG
956 	{
957 		char ifname[IF_NAMESIZE];
958 
959 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
960 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
961 				eth_dev->data->port_id, ifname);
962 		else
963 			DRV_LOG(DEBUG, "port %u ifname is unknown",
964 				eth_dev->data->port_id);
965 	}
966 #endif
967 	/* Get actual MTU if possible. */
968 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
969 	if (err) {
970 		err = rte_errno;
971 		goto error;
972 	}
973 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
974 		priv->mtu);
975 	/* Initialize burst functions to prevent crashes before link-up. */
976 	eth_dev->rx_pkt_burst = removed_rx_burst;
977 	eth_dev->tx_pkt_burst = removed_tx_burst;
978 	eth_dev->dev_ops = &mlx5_os_dev_ops;
979 	/* Register MAC address. */
980 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
981 	if (config.vf && config.vf_nl_en)
982 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
983 				      mlx5_ifindex(eth_dev),
984 				      eth_dev->data->mac_addrs,
985 				      MLX5_MAX_MAC_ADDRESSES);
986 	priv->flows = 0;
987 	priv->ctrl_flows = 0;
988 	TAILQ_INIT(&priv->flow_meters);
989 	TAILQ_INIT(&priv->flow_meter_profiles);
990 	/* Hint libmlx5 to use PMD allocator for data plane resources */
991 	struct mlx5dv_ctx_allocators alctr = {
992 		.alloc = &mlx5_alloc_verbs_buf,
993 		.free = &mlx5_free_verbs_buf,
994 		.data = priv,
995 	};
996 	mlx5_glue->dv_set_context_attr(sh->ctx,
997 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
998 				       (void *)((uintptr_t)&alctr));
999 	/* Bring Ethernet device up. */
1000 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1001 		eth_dev->data->port_id);
1002 	mlx5_set_link_up(eth_dev);
1003 	/*
1004 	 * Even though the interrupt handler is not installed yet,
1005 	 * interrupts will still trigger on the async_fd from
1006 	 * Verbs context returned by ibv_open_device().
1007 	 */
1008 	mlx5_link_update(eth_dev, 0);
1009 #ifdef HAVE_MLX5DV_DR_ESWITCH
1010 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1011 	      (switch_info->representor || switch_info->master)))
1012 		config.dv_esw_en = 0;
1013 #else
1014 	config.dv_esw_en = 0;
1015 #endif
1016 	/* Detect minimal data bytes to inline. */
1017 	mlx5_set_min_inline(spawn, &config);
1018 	/* Store device configuration on private structure. */
1019 	priv->config = config;
1020 	/* Create context for virtual machine VLAN workaround. */
1021 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1022 	if (config.dv_flow_en) {
1023 		err = mlx5_alloc_shared_dr(priv);
1024 		if (err)
1025 			goto error;
1026 		/*
1027 		 * RSS id is shared with meter flow id. Meter flow id can only
1028 		 * use the 24 MSB of the register.
1029 		 */
1030 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1031 				     MLX5_MTR_COLOR_BITS);
1032 		if (!priv->qrss_id_pool) {
1033 			DRV_LOG(ERR, "can't create flow id pool");
1034 			err = ENOMEM;
1035 			goto error;
1036 		}
1037 	}
1038 	/* Supported Verbs flow priority number detection. */
1039 	err = mlx5_flow_discover_priorities(eth_dev);
1040 	if (err < 0) {
1041 		err = -err;
1042 		goto error;
1043 	}
1044 	priv->config.flow_prio = err;
1045 	if (!priv->config.dv_esw_en &&
1046 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1047 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1048 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1049 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1050 	}
1051 	mlx5_set_metadata_mask(eth_dev);
1052 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1053 	    !priv->sh->dv_regc0_mask) {
1054 		DRV_LOG(ERR, "metadata mode %u is not supported "
1055 			     "(no metadata reg_c[0] is available)",
1056 			     priv->config.dv_xmeta_en);
1057 			err = ENOTSUP;
1058 			goto error;
1059 	}
1060 	/*
1061 	 * Allocate the buffer for flow creating, just once.
1062 	 * The allocation must be done before any flow creating.
1063 	 */
1064 	mlx5_flow_alloc_intermediate(eth_dev);
1065 	/* Query availability of metadata reg_c's. */
1066 	err = mlx5_flow_discover_mreg_c(eth_dev);
1067 	if (err < 0) {
1068 		err = -err;
1069 		goto error;
1070 	}
1071 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1072 		DRV_LOG(DEBUG,
1073 			"port %u extensive metadata register is not supported",
1074 			eth_dev->data->port_id);
1075 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1076 			DRV_LOG(ERR, "metadata mode %u is not supported "
1077 				     "(no metadata registers available)",
1078 				     priv->config.dv_xmeta_en);
1079 			err = ENOTSUP;
1080 			goto error;
1081 		}
1082 	}
1083 	if (priv->config.dv_flow_en &&
1084 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1085 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1086 	    priv->sh->dv_regc0_mask) {
1087 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1088 						      MLX5_FLOW_MREG_HTABLE_SZ);
1089 		if (!priv->mreg_cp_tbl) {
1090 			err = ENOMEM;
1091 			goto error;
1092 		}
1093 	}
1094 	return eth_dev;
1095 error:
1096 	if (priv) {
1097 		if (priv->mreg_cp_tbl)
1098 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1099 		if (priv->sh)
1100 			mlx5_os_free_shared_dr(priv);
1101 		if (priv->nl_socket_route >= 0)
1102 			close(priv->nl_socket_route);
1103 		if (priv->nl_socket_rdma >= 0)
1104 			close(priv->nl_socket_rdma);
1105 		if (priv->vmwa_context)
1106 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1107 		if (priv->qrss_id_pool)
1108 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1109 		if (own_domain_id)
1110 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1111 		rte_free(priv);
1112 		if (eth_dev != NULL)
1113 			eth_dev->data->dev_private = NULL;
1114 	}
1115 	if (eth_dev != NULL) {
1116 		/* mac_addrs must not be freed alone because part of
1117 		 * dev_private
1118 		 **/
1119 		eth_dev->data->mac_addrs = NULL;
1120 		rte_eth_dev_release_port(eth_dev);
1121 	}
1122 	if (sh)
1123 		mlx5_free_shared_dev_ctx(sh);
1124 	MLX5_ASSERT(err > 0);
1125 	rte_errno = err;
1126 	return NULL;
1127 }
1128 
1129 /**
1130  * Comparison callback to sort device data.
1131  *
1132  * This is meant to be used with qsort().
1133  *
1134  * @param a[in]
1135  *   Pointer to pointer to first data object.
1136  * @param b[in]
1137  *   Pointer to pointer to second data object.
1138  *
1139  * @return
1140  *   0 if both objects are equal, less than 0 if the first argument is less
1141  *   than the second, greater than 0 otherwise.
1142  */
1143 static int
1144 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1145 {
1146 	const struct mlx5_switch_info *si_a =
1147 		&((const struct mlx5_dev_spawn_data *)a)->info;
1148 	const struct mlx5_switch_info *si_b =
1149 		&((const struct mlx5_dev_spawn_data *)b)->info;
1150 	int ret;
1151 
1152 	/* Master device first. */
1153 	ret = si_b->master - si_a->master;
1154 	if (ret)
1155 		return ret;
1156 	/* Then representor devices. */
1157 	ret = si_b->representor - si_a->representor;
1158 	if (ret)
1159 		return ret;
1160 	/* Unidentified devices come last in no specific order. */
1161 	if (!si_a->representor)
1162 		return 0;
1163 	/* Order representors by name. */
1164 	return si_a->port_name - si_b->port_name;
1165 }
1166 
1167 /**
1168  * Match PCI information for possible slaves of bonding device.
1169  *
1170  * @param[in] ibv_dev
1171  *   Pointer to Infiniband device structure.
1172  * @param[in] pci_dev
1173  *   Pointer to PCI device structure to match PCI address.
1174  * @param[in] nl_rdma
1175  *   Netlink RDMA group socket handle.
1176  *
1177  * @return
1178  *   negative value if no bonding device found, otherwise
1179  *   positive index of slave PF in bonding.
1180  */
1181 static int
1182 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1183 			   const struct rte_pci_device *pci_dev,
1184 			   int nl_rdma)
1185 {
1186 	char ifname[IF_NAMESIZE + 1];
1187 	unsigned int ifindex;
1188 	unsigned int np, i;
1189 	FILE *file = NULL;
1190 	int pf = -1;
1191 
1192 	/*
1193 	 * Try to get master device name. If something goes
1194 	 * wrong suppose the lack of kernel support and no
1195 	 * bonding devices.
1196 	 */
1197 	if (nl_rdma < 0)
1198 		return -1;
1199 	if (!strstr(ibv_dev->name, "bond"))
1200 		return -1;
1201 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1202 	if (!np)
1203 		return -1;
1204 	/*
1205 	 * The Master device might not be on the predefined
1206 	 * port (not on port index 1, it is not garanted),
1207 	 * we have to scan all Infiniband device port and
1208 	 * find master.
1209 	 */
1210 	for (i = 1; i <= np; ++i) {
1211 		/* Check whether Infiniband port is populated. */
1212 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1213 		if (!ifindex)
1214 			continue;
1215 		if (!if_indextoname(ifindex, ifname))
1216 			continue;
1217 		/* Try to read bonding slave names from sysfs. */
1218 		MKSTR(slaves,
1219 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1220 		file = fopen(slaves, "r");
1221 		if (file)
1222 			break;
1223 	}
1224 	if (!file)
1225 		return -1;
1226 	/* Use safe format to check maximal buffer length. */
1227 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1228 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1229 		char tmp_str[IF_NAMESIZE + 32];
1230 		struct rte_pci_addr pci_addr;
1231 		struct mlx5_switch_info	info;
1232 
1233 		/* Process slave interface names in the loop. */
1234 		snprintf(tmp_str, sizeof(tmp_str),
1235 			 "/sys/class/net/%s", ifname);
1236 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1237 			DRV_LOG(WARNING, "can not get PCI address"
1238 					 " for netdev \"%s\"", ifname);
1239 			continue;
1240 		}
1241 		if (pci_dev->addr.domain != pci_addr.domain ||
1242 		    pci_dev->addr.bus != pci_addr.bus ||
1243 		    pci_dev->addr.devid != pci_addr.devid ||
1244 		    pci_dev->addr.function != pci_addr.function)
1245 			continue;
1246 		/* Slave interface PCI address match found. */
1247 		fclose(file);
1248 		snprintf(tmp_str, sizeof(tmp_str),
1249 			 "/sys/class/net/%s/phys_port_name", ifname);
1250 		file = fopen(tmp_str, "rb");
1251 		if (!file)
1252 			break;
1253 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1254 		if (fscanf(file, "%32s", tmp_str) == 1)
1255 			mlx5_translate_port_name(tmp_str, &info);
1256 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1257 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1258 			pf = info.port_name;
1259 		break;
1260 	}
1261 	if (file)
1262 		fclose(file);
1263 	return pf;
1264 }
1265 
1266 /**
1267  * DPDK callback to register a PCI device.
1268  *
1269  * This function spawns Ethernet devices out of a given PCI device.
1270  *
1271  * @param[in] pci_drv
1272  *   PCI driver structure (mlx5_driver).
1273  * @param[in] pci_dev
1274  *   PCI device information.
1275  *
1276  * @return
1277  *   0 on success, a negative errno value otherwise and rte_errno is set.
1278  */
1279 int
1280 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1281 		  struct rte_pci_device *pci_dev)
1282 {
1283 	struct ibv_device **ibv_list;
1284 	/*
1285 	 * Number of found IB Devices matching with requested PCI BDF.
1286 	 * nd != 1 means there are multiple IB devices over the same
1287 	 * PCI device and we have representors and master.
1288 	 */
1289 	unsigned int nd = 0;
1290 	/*
1291 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1292 	 * we have the single multiport IB device, and there may be
1293 	 * representors attached to some of found ports.
1294 	 */
1295 	unsigned int np = 0;
1296 	/*
1297 	 * Number of DPDK ethernet devices to Spawn - either over
1298 	 * multiple IB devices or multiple ports of single IB device.
1299 	 * Actually this is the number of iterations to spawn.
1300 	 */
1301 	unsigned int ns = 0;
1302 	/*
1303 	 * Bonding device
1304 	 *   < 0 - no bonding device (single one)
1305 	 *  >= 0 - bonding device (value is slave PF index)
1306 	 */
1307 	int bd = -1;
1308 	struct mlx5_dev_spawn_data *list = NULL;
1309 	struct mlx5_dev_config dev_config;
1310 	int ret;
1311 
1312 	if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1313 		DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1314 			" driver.");
1315 		return 1;
1316 	}
1317 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1318 		mlx5_pmd_socket_init();
1319 	ret = mlx5_init_once();
1320 	if (ret) {
1321 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1322 			strerror(rte_errno));
1323 		return -rte_errno;
1324 	}
1325 	MLX5_ASSERT(pci_drv == &mlx5_driver);
1326 	errno = 0;
1327 	ibv_list = mlx5_glue->get_device_list(&ret);
1328 	if (!ibv_list) {
1329 		rte_errno = errno ? errno : ENOSYS;
1330 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1331 		return -rte_errno;
1332 	}
1333 	/*
1334 	 * First scan the list of all Infiniband devices to find
1335 	 * matching ones, gathering into the list.
1336 	 */
1337 	struct ibv_device *ibv_match[ret + 1];
1338 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1339 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1340 	unsigned int i;
1341 
1342 	while (ret-- > 0) {
1343 		struct rte_pci_addr pci_addr;
1344 
1345 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1346 		bd = mlx5_device_bond_pci_match
1347 				(ibv_list[ret], pci_dev, nl_rdma);
1348 		if (bd >= 0) {
1349 			/*
1350 			 * Bonding device detected. Only one match is allowed,
1351 			 * the bonding is supported over multi-port IB device,
1352 			 * there should be no matches on representor PCI
1353 			 * functions or non VF LAG bonding devices with
1354 			 * specified address.
1355 			 */
1356 			if (nd) {
1357 				DRV_LOG(ERR,
1358 					"multiple PCI match on bonding device"
1359 					"\"%s\" found", ibv_list[ret]->name);
1360 				rte_errno = ENOENT;
1361 				ret = -rte_errno;
1362 				goto exit;
1363 			}
1364 			DRV_LOG(INFO, "PCI information matches for"
1365 				      " slave %d bonding device \"%s\"",
1366 				      bd, ibv_list[ret]->name);
1367 			ibv_match[nd++] = ibv_list[ret];
1368 			break;
1369 		}
1370 		if (mlx5_dev_to_pci_addr
1371 			(ibv_list[ret]->ibdev_path, &pci_addr))
1372 			continue;
1373 		if (pci_dev->addr.domain != pci_addr.domain ||
1374 		    pci_dev->addr.bus != pci_addr.bus ||
1375 		    pci_dev->addr.devid != pci_addr.devid ||
1376 		    pci_dev->addr.function != pci_addr.function)
1377 			continue;
1378 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1379 			ibv_list[ret]->name);
1380 		ibv_match[nd++] = ibv_list[ret];
1381 	}
1382 	ibv_match[nd] = NULL;
1383 	if (!nd) {
1384 		/* No device matches, just complain and bail out. */
1385 		DRV_LOG(WARNING,
1386 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1387 			" are kernel drivers loaded?",
1388 			pci_dev->addr.domain, pci_dev->addr.bus,
1389 			pci_dev->addr.devid, pci_dev->addr.function);
1390 		rte_errno = ENOENT;
1391 		ret = -rte_errno;
1392 		goto exit;
1393 	}
1394 	if (nd == 1) {
1395 		/*
1396 		 * Found single matching device may have multiple ports.
1397 		 * Each port may be representor, we have to check the port
1398 		 * number and check the representors existence.
1399 		 */
1400 		if (nl_rdma >= 0)
1401 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1402 		if (!np)
1403 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1404 					 " ports number", ibv_match[0]->name);
1405 		if (bd >= 0 && !np) {
1406 			DRV_LOG(ERR, "can not get ports"
1407 				     " for bonding device");
1408 			rte_errno = ENOENT;
1409 			ret = -rte_errno;
1410 			goto exit;
1411 		}
1412 	}
1413 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1414 	if (bd >= 0) {
1415 		/*
1416 		 * This may happen if there is VF LAG kernel support and
1417 		 * application is compiled with older rdma_core library.
1418 		 */
1419 		DRV_LOG(ERR,
1420 			"No kernel/verbs support for VF LAG bonding found.");
1421 		rte_errno = ENOTSUP;
1422 		ret = -rte_errno;
1423 		goto exit;
1424 	}
1425 #endif
1426 	/*
1427 	 * Now we can determine the maximal
1428 	 * amount of devices to be spawned.
1429 	 */
1430 	list = rte_zmalloc("device spawn data",
1431 			 sizeof(struct mlx5_dev_spawn_data) *
1432 			 (np ? np : nd),
1433 			 RTE_CACHE_LINE_SIZE);
1434 	if (!list) {
1435 		DRV_LOG(ERR, "spawn data array allocation failure");
1436 		rte_errno = ENOMEM;
1437 		ret = -rte_errno;
1438 		goto exit;
1439 	}
1440 	if (bd >= 0 || np > 1) {
1441 		/*
1442 		 * Single IB device with multiple ports found,
1443 		 * it may be E-Switch master device and representors.
1444 		 * We have to perform identification through the ports.
1445 		 */
1446 		MLX5_ASSERT(nl_rdma >= 0);
1447 		MLX5_ASSERT(ns == 0);
1448 		MLX5_ASSERT(nd == 1);
1449 		MLX5_ASSERT(np);
1450 		for (i = 1; i <= np; ++i) {
1451 			list[ns].max_port = np;
1452 			list[ns].phys_port = i;
1453 			list[ns].phys_dev = ibv_match[0];
1454 			list[ns].eth_dev = NULL;
1455 			list[ns].pci_dev = pci_dev;
1456 			list[ns].pf_bond = bd;
1457 			list[ns].ifindex = mlx5_nl_ifindex
1458 				(nl_rdma,
1459 				mlx5_os_get_dev_device_name
1460 						(list[ns].phys_dev), i);
1461 			if (!list[ns].ifindex) {
1462 				/*
1463 				 * No network interface index found for the
1464 				 * specified port, it means there is no
1465 				 * representor on this port. It's OK,
1466 				 * there can be disabled ports, for example
1467 				 * if sriov_numvfs < sriov_totalvfs.
1468 				 */
1469 				continue;
1470 			}
1471 			ret = -1;
1472 			if (nl_route >= 0)
1473 				ret = mlx5_nl_switch_info
1474 					       (nl_route,
1475 						list[ns].ifindex,
1476 						&list[ns].info);
1477 			if (ret || (!list[ns].info.representor &&
1478 				    !list[ns].info.master)) {
1479 				/*
1480 				 * We failed to recognize representors with
1481 				 * Netlink, let's try to perform the task
1482 				 * with sysfs.
1483 				 */
1484 				ret =  mlx5_sysfs_switch_info
1485 						(list[ns].ifindex,
1486 						 &list[ns].info);
1487 			}
1488 			if (!ret && bd >= 0) {
1489 				switch (list[ns].info.name_type) {
1490 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1491 					if (list[ns].info.port_name == bd)
1492 						ns++;
1493 					break;
1494 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1495 					/* Fallthrough */
1496 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1497 					if (list[ns].info.pf_num == bd)
1498 						ns++;
1499 					break;
1500 				default:
1501 					break;
1502 				}
1503 				continue;
1504 			}
1505 			if (!ret && (list[ns].info.representor ^
1506 				     list[ns].info.master))
1507 				ns++;
1508 		}
1509 		if (!ns) {
1510 			DRV_LOG(ERR,
1511 				"unable to recognize master/representors"
1512 				" on the IB device with multiple ports");
1513 			rte_errno = ENOENT;
1514 			ret = -rte_errno;
1515 			goto exit;
1516 		}
1517 	} else {
1518 		/*
1519 		 * The existence of several matching entries (nd > 1) means
1520 		 * port representors have been instantiated. No existing Verbs
1521 		 * call nor sysfs entries can tell them apart, this can only
1522 		 * be done through Netlink calls assuming kernel drivers are
1523 		 * recent enough to support them.
1524 		 *
1525 		 * In the event of identification failure through Netlink,
1526 		 * try again through sysfs, then:
1527 		 *
1528 		 * 1. A single IB device matches (nd == 1) with single
1529 		 *    port (np=0/1) and is not a representor, assume
1530 		 *    no switch support.
1531 		 *
1532 		 * 2. Otherwise no safe assumptions can be made;
1533 		 *    complain louder and bail out.
1534 		 */
1535 		for (i = 0; i != nd; ++i) {
1536 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1537 			list[ns].max_port = 1;
1538 			list[ns].phys_port = 1;
1539 			list[ns].phys_dev = ibv_match[i];
1540 			list[ns].eth_dev = NULL;
1541 			list[ns].pci_dev = pci_dev;
1542 			list[ns].pf_bond = -1;
1543 			list[ns].ifindex = 0;
1544 			if (nl_rdma >= 0)
1545 				list[ns].ifindex = mlx5_nl_ifindex
1546 				(nl_rdma,
1547 				mlx5_os_get_dev_device_name
1548 						(list[ns].phys_dev), 1);
1549 			if (!list[ns].ifindex) {
1550 				char ifname[IF_NAMESIZE];
1551 
1552 				/*
1553 				 * Netlink failed, it may happen with old
1554 				 * ib_core kernel driver (before 4.16).
1555 				 * We can assume there is old driver because
1556 				 * here we are processing single ports IB
1557 				 * devices. Let's try sysfs to retrieve
1558 				 * the ifindex. The method works for
1559 				 * master device only.
1560 				 */
1561 				if (nd > 1) {
1562 					/*
1563 					 * Multiple devices found, assume
1564 					 * representors, can not distinguish
1565 					 * master/representor and retrieve
1566 					 * ifindex via sysfs.
1567 					 */
1568 					continue;
1569 				}
1570 				ret = mlx5_get_ifname_sysfs
1571 					(ibv_match[i]->ibdev_path, ifname);
1572 				if (!ret)
1573 					list[ns].ifindex =
1574 						if_nametoindex(ifname);
1575 				if (!list[ns].ifindex) {
1576 					/*
1577 					 * No network interface index found
1578 					 * for the specified device, it means
1579 					 * there it is neither representor
1580 					 * nor master.
1581 					 */
1582 					continue;
1583 				}
1584 			}
1585 			ret = -1;
1586 			if (nl_route >= 0)
1587 				ret = mlx5_nl_switch_info
1588 					       (nl_route,
1589 						list[ns].ifindex,
1590 						&list[ns].info);
1591 			if (ret || (!list[ns].info.representor &&
1592 				    !list[ns].info.master)) {
1593 				/*
1594 				 * We failed to recognize representors with
1595 				 * Netlink, let's try to perform the task
1596 				 * with sysfs.
1597 				 */
1598 				ret =  mlx5_sysfs_switch_info
1599 						(list[ns].ifindex,
1600 						 &list[ns].info);
1601 			}
1602 			if (!ret && (list[ns].info.representor ^
1603 				     list[ns].info.master)) {
1604 				ns++;
1605 			} else if ((nd == 1) &&
1606 				   !list[ns].info.representor &&
1607 				   !list[ns].info.master) {
1608 				/*
1609 				 * Single IB device with
1610 				 * one physical port and
1611 				 * attached network device.
1612 				 * May be SRIOV is not enabled
1613 				 * or there is no representors.
1614 				 */
1615 				DRV_LOG(INFO, "no E-Switch support detected");
1616 				ns++;
1617 				break;
1618 			}
1619 		}
1620 		if (!ns) {
1621 			DRV_LOG(ERR,
1622 				"unable to recognize master/representors"
1623 				" on the multiple IB devices");
1624 			rte_errno = ENOENT;
1625 			ret = -rte_errno;
1626 			goto exit;
1627 		}
1628 	}
1629 	MLX5_ASSERT(ns);
1630 	/*
1631 	 * Sort list to probe devices in natural order for users convenience
1632 	 * (i.e. master first, then representors from lowest to highest ID).
1633 	 */
1634 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1635 	/* Default configuration. */
1636 	dev_config = (struct mlx5_dev_config){
1637 		.hw_padding = 0,
1638 		.mps = MLX5_ARG_UNSET,
1639 		.dbnc = MLX5_ARG_UNSET,
1640 		.rx_vec_en = 1,
1641 		.txq_inline_max = MLX5_ARG_UNSET,
1642 		.txq_inline_min = MLX5_ARG_UNSET,
1643 		.txq_inline_mpw = MLX5_ARG_UNSET,
1644 		.txqs_inline = MLX5_ARG_UNSET,
1645 		.vf_nl_en = 1,
1646 		.mr_ext_memseg_en = 1,
1647 		.mprq = {
1648 			.enabled = 0, /* Disabled by default. */
1649 			.stride_num_n = 0,
1650 			.stride_size_n = 0,
1651 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1652 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1653 		},
1654 		.dv_esw_en = 1,
1655 		.dv_flow_en = 1,
1656 		.log_hp_size = MLX5_ARG_UNSET,
1657 	};
1658 	/* Device specific configuration. */
1659 	switch (pci_dev->id.device_id) {
1660 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1661 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1662 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1663 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1664 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1665 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1666 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1667 		dev_config.vf = 1;
1668 		break;
1669 	default:
1670 		break;
1671 	}
1672 	for (i = 0; i != ns; ++i) {
1673 		uint32_t restore;
1674 
1675 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1676 						 &list[i],
1677 						 dev_config);
1678 		if (!list[i].eth_dev) {
1679 			if (rte_errno != EBUSY && rte_errno != EEXIST)
1680 				break;
1681 			/* Device is disabled or already spawned. Ignore it. */
1682 			continue;
1683 		}
1684 		restore = list[i].eth_dev->data->dev_flags;
1685 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1686 		/* Restore non-PCI flags cleared by the above call. */
1687 		list[i].eth_dev->data->dev_flags |= restore;
1688 		rte_eth_dev_probing_finish(list[i].eth_dev);
1689 	}
1690 	if (i != ns) {
1691 		DRV_LOG(ERR,
1692 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1693 			" encountering an error: %s",
1694 			pci_dev->addr.domain, pci_dev->addr.bus,
1695 			pci_dev->addr.devid, pci_dev->addr.function,
1696 			strerror(rte_errno));
1697 		ret = -rte_errno;
1698 		/* Roll back. */
1699 		while (i--) {
1700 			if (!list[i].eth_dev)
1701 				continue;
1702 			mlx5_dev_close(list[i].eth_dev);
1703 			/* mac_addrs must not be freed because in dev_private */
1704 			list[i].eth_dev->data->mac_addrs = NULL;
1705 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1706 		}
1707 		/* Restore original error. */
1708 		rte_errno = -ret;
1709 	} else {
1710 		ret = 0;
1711 	}
1712 exit:
1713 	/*
1714 	 * Do the routine cleanup:
1715 	 * - close opened Netlink sockets
1716 	 * - free allocated spawn data array
1717 	 * - free the Infiniband device list
1718 	 */
1719 	if (nl_rdma >= 0)
1720 		close(nl_rdma);
1721 	if (nl_route >= 0)
1722 		close(nl_route);
1723 	if (list)
1724 		rte_free(list);
1725 	MLX5_ASSERT(ibv_list);
1726 	mlx5_glue->free_device_list(ibv_list);
1727 	return ret;
1728 }
1729 
1730 static int
1731 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1732 {
1733 	char *env;
1734 	int value;
1735 
1736 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1737 	/* Get environment variable to store. */
1738 	env = getenv(MLX5_SHUT_UP_BF);
1739 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1740 	if (config->dbnc == MLX5_ARG_UNSET)
1741 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1742 	else
1743 		setenv(MLX5_SHUT_UP_BF,
1744 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1745 	return value;
1746 }
1747 
1748 static void
1749 mlx5_restore_doorbell_mapping_env(int value)
1750 {
1751 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1752 	/* Restore the original environment variable state. */
1753 	if (value == MLX5_ARG_UNSET)
1754 		unsetenv(MLX5_SHUT_UP_BF);
1755 	else
1756 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1757 }
1758 
1759 /**
1760  * Extract pdn of PD object using DV API.
1761  *
1762  * @param[in] pd
1763  *   Pointer to the verbs PD object.
1764  * @param[out] pdn
1765  *   Pointer to the PD object number variable.
1766  *
1767  * @return
1768  *   0 on success, error value otherwise.
1769  */
1770 int
1771 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1772 {
1773 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1774 	struct mlx5dv_obj obj;
1775 	struct mlx5dv_pd pd_info;
1776 	int ret = 0;
1777 
1778 	obj.pd.in = pd;
1779 	obj.pd.out = &pd_info;
1780 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1781 	if (ret) {
1782 		DRV_LOG(DEBUG, "Fail to get PD object info");
1783 		return ret;
1784 	}
1785 	*pdn = pd_info.pdn;
1786 	return 0;
1787 #else
1788 	(void)pd;
1789 	(void)pdn;
1790 	return -ENOTSUP;
1791 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1792 }
1793 
1794 /**
1795  * Function API to open IB device.
1796  *
1797  * This function calls the Linux glue APIs to open a device.
1798  *
1799  * @param[in] spawn
1800  *   Pointer to the IB device attributes (name, port, etc).
1801  * @param[out] config
1802  *   Pointer to device configuration structure.
1803  * @param[out] sh
1804  *   Pointer to shared context structure.
1805  *
1806  * @return
1807  *   0 on success, a positive error value otherwise.
1808  */
1809 int
1810 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1811 		     const struct mlx5_dev_config *config,
1812 		     struct mlx5_dev_ctx_shared *sh)
1813 {
1814 	int dbmap_env;
1815 	int err = 0;
1816 	/*
1817 	 * Configure environment variable "MLX5_BF_SHUT_UP"
1818 	 * before the device creation. The rdma_core library
1819 	 * checks the variable at device creation and
1820 	 * stores the result internally.
1821 	 */
1822 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
1823 	/* Try to open IB device with DV first, then usual Verbs. */
1824 	errno = 0;
1825 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1826 	if (sh->ctx) {
1827 		sh->devx = 1;
1828 		DRV_LOG(DEBUG, "DevX is supported");
1829 		/* The device is created, no need for environment. */
1830 		mlx5_restore_doorbell_mapping_env(dbmap_env);
1831 	} else {
1832 		/* The environment variable is still configured. */
1833 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1834 		err = errno ? errno : ENODEV;
1835 		/*
1836 		 * The environment variable is not needed anymore,
1837 		 * all device creation attempts are completed.
1838 		 */
1839 		mlx5_restore_doorbell_mapping_env(dbmap_env);
1840 		if (!sh->ctx)
1841 			return err;
1842 		DRV_LOG(DEBUG, "DevX is NOT supported");
1843 		err = 0;
1844 	}
1845 	return err;
1846 }
1847 
1848 /**
1849  * Install shared asynchronous device events handler.
1850  * This function is implemented to support event sharing
1851  * between multiple ports of single IB device.
1852  *
1853  * @param sh
1854  *   Pointer to mlx5_dev_ctx_shared object.
1855  */
1856 void
1857 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1858 {
1859 	int ret;
1860 	int flags;
1861 
1862 	sh->intr_handle.fd = -1;
1863 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1864 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1865 		    F_SETFL, flags | O_NONBLOCK);
1866 	if (ret) {
1867 		DRV_LOG(INFO, "failed to change file descriptor async event"
1868 			" queue");
1869 	} else {
1870 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1871 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1872 		if (rte_intr_callback_register(&sh->intr_handle,
1873 					mlx5_dev_interrupt_handler, sh)) {
1874 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
1875 			sh->intr_handle.fd = -1;
1876 		}
1877 	}
1878 	if (sh->devx) {
1879 #ifdef HAVE_IBV_DEVX_ASYNC
1880 		sh->intr_handle_devx.fd = -1;
1881 		sh->devx_comp =
1882 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1883 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1884 		if (!devx_comp) {
1885 			DRV_LOG(INFO, "failed to allocate devx_comp.");
1886 			return;
1887 		}
1888 		flags = fcntl(devx_comp->fd, F_GETFL);
1889 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1890 		if (ret) {
1891 			DRV_LOG(INFO, "failed to change file descriptor"
1892 				" devx comp");
1893 			return;
1894 		}
1895 		sh->intr_handle_devx.fd = devx_comp->fd;
1896 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1897 		if (rte_intr_callback_register(&sh->intr_handle_devx,
1898 					mlx5_dev_interrupt_handler_devx, sh)) {
1899 			DRV_LOG(INFO, "Fail to install the devx shared"
1900 				" interrupt.");
1901 			sh->intr_handle_devx.fd = -1;
1902 		}
1903 #endif /* HAVE_IBV_DEVX_ASYNC */
1904 	}
1905 }
1906 
1907 /**
1908  * Uninstall shared asynchronous device events handler.
1909  * This function is implemented to support event sharing
1910  * between multiple ports of single IB device.
1911  *
1912  * @param dev
1913  *   Pointer to mlx5_dev_ctx_shared object.
1914  */
1915 void
1916 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
1917 {
1918 	if (sh->intr_handle.fd >= 0)
1919 		mlx5_intr_callback_unregister(&sh->intr_handle,
1920 					      mlx5_dev_interrupt_handler, sh);
1921 #ifdef HAVE_IBV_DEVX_ASYNC
1922 	if (sh->intr_handle_devx.fd >= 0)
1923 		rte_intr_callback_unregister(&sh->intr_handle_devx,
1924 				  mlx5_dev_interrupt_handler_devx, sh);
1925 	if (sh->devx_comp)
1926 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1927 #endif
1928 }
1929 
1930 /**
1931  * Read statistics by a named counter.
1932  *
1933  * @param[in] priv
1934  *   Pointer to the private device data structure.
1935  * @param[in] ctr_name
1936  *   Pointer to the name of the statistic counter to read
1937  * @param[out] stat
1938  *   Pointer to read statistic value.
1939  * @return
1940  *   0 on success and stat is valud, 1 if failed to read the value
1941  *   rte_errno is set.
1942  *
1943  */
1944 int
1945 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
1946 		      uint64_t *stat)
1947 {
1948 	int fd;
1949 
1950 	if (priv->sh) {
1951 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
1952 			  priv->sh->ibdev_path,
1953 			  priv->dev_port,
1954 			  ctr_name);
1955 		fd = open(path, O_RDONLY);
1956 		if (fd != -1) {
1957 			char buf[21] = {'\0'};
1958 			ssize_t n = read(fd, buf, sizeof(buf));
1959 
1960 			close(fd);
1961 			if (n != -1) {
1962 				*stat = strtoull(buf, NULL, 10);
1963 				return 0;
1964 			}
1965 		}
1966 	}
1967 	*stat = 0;
1968 	return 1;
1969 }
1970 
1971 /**
1972  * Read device counters table.
1973  *
1974  * @param dev
1975  *   Pointer to Ethernet device.
1976  * @param[out] stats
1977  *   Counters table output buffer.
1978  *
1979  * @return
1980  *   0 on success and stats is filled, negative errno value otherwise and
1981  *   rte_errno is set.
1982  */
1983 int
1984 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
1985 {
1986 	struct mlx5_priv *priv = dev->data->dev_private;
1987 	struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
1988 	unsigned int i;
1989 	struct ifreq ifr;
1990 	unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
1991 	unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
1992 	struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
1993 	int ret;
1994 
1995 	et_stats->cmd = ETHTOOL_GSTATS;
1996 	et_stats->n_stats = xstats_ctrl->stats_n;
1997 	ifr.ifr_data = (caddr_t)et_stats;
1998 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1999 	if (ret) {
2000 		DRV_LOG(WARNING,
2001 			"port %u unable to read statistic values from device",
2002 			dev->data->port_id);
2003 		return ret;
2004 	}
2005 	for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2006 		if (xstats_ctrl->info[i].dev) {
2007 			ret = mlx5_os_read_dev_stat(priv,
2008 					    xstats_ctrl->info[i].ctr_name,
2009 					    &stats[i]);
2010 			/* return last xstats counter if fail to read. */
2011 			if (ret == 0)
2012 				xstats_ctrl->xstats[i] = stats[i];
2013 			else
2014 				stats[i] = xstats_ctrl->xstats[i];
2015 		} else {
2016 			stats[i] = (uint64_t)
2017 				et_stats->data[xstats_ctrl->dev_table_idx[i]];
2018 		}
2019 	}
2020 	return 0;
2021 }
2022 
2023 /**
2024  * Query the number of statistics provided by ETHTOOL.
2025  *
2026  * @param dev
2027  *   Pointer to Ethernet device.
2028  *
2029  * @return
2030  *   Number of statistics on success, negative errno value otherwise and
2031  *   rte_errno is set.
2032  */
2033 int
2034 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2035 {
2036 	struct ethtool_drvinfo drvinfo;
2037 	struct ifreq ifr;
2038 	int ret;
2039 
2040 	drvinfo.cmd = ETHTOOL_GDRVINFO;
2041 	ifr.ifr_data = (caddr_t)&drvinfo;
2042 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2043 	if (ret) {
2044 		DRV_LOG(WARNING, "port %u unable to query number of statistics",
2045 			dev->data->port_id);
2046 		return ret;
2047 	}
2048 	return drvinfo.n_stats;
2049 }
2050 
2051 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2052 	{
2053 		.dpdk_name = "rx_port_unicast_bytes",
2054 		.ctr_name = "rx_vport_unicast_bytes",
2055 	},
2056 	{
2057 		.dpdk_name = "rx_port_multicast_bytes",
2058 		.ctr_name = "rx_vport_multicast_bytes",
2059 	},
2060 	{
2061 		.dpdk_name = "rx_port_broadcast_bytes",
2062 		.ctr_name = "rx_vport_broadcast_bytes",
2063 	},
2064 	{
2065 		.dpdk_name = "rx_port_unicast_packets",
2066 		.ctr_name = "rx_vport_unicast_packets",
2067 	},
2068 	{
2069 		.dpdk_name = "rx_port_multicast_packets",
2070 		.ctr_name = "rx_vport_multicast_packets",
2071 	},
2072 	{
2073 		.dpdk_name = "rx_port_broadcast_packets",
2074 		.ctr_name = "rx_vport_broadcast_packets",
2075 	},
2076 	{
2077 		.dpdk_name = "tx_port_unicast_bytes",
2078 		.ctr_name = "tx_vport_unicast_bytes",
2079 	},
2080 	{
2081 		.dpdk_name = "tx_port_multicast_bytes",
2082 		.ctr_name = "tx_vport_multicast_bytes",
2083 	},
2084 	{
2085 		.dpdk_name = "tx_port_broadcast_bytes",
2086 		.ctr_name = "tx_vport_broadcast_bytes",
2087 	},
2088 	{
2089 		.dpdk_name = "tx_port_unicast_packets",
2090 		.ctr_name = "tx_vport_unicast_packets",
2091 	},
2092 	{
2093 		.dpdk_name = "tx_port_multicast_packets",
2094 		.ctr_name = "tx_vport_multicast_packets",
2095 	},
2096 	{
2097 		.dpdk_name = "tx_port_broadcast_packets",
2098 		.ctr_name = "tx_vport_broadcast_packets",
2099 	},
2100 	{
2101 		.dpdk_name = "rx_wqe_err",
2102 		.ctr_name = "rx_wqe_err",
2103 	},
2104 	{
2105 		.dpdk_name = "rx_crc_errors_phy",
2106 		.ctr_name = "rx_crc_errors_phy",
2107 	},
2108 	{
2109 		.dpdk_name = "rx_in_range_len_errors_phy",
2110 		.ctr_name = "rx_in_range_len_errors_phy",
2111 	},
2112 	{
2113 		.dpdk_name = "rx_symbol_err_phy",
2114 		.ctr_name = "rx_symbol_err_phy",
2115 	},
2116 	{
2117 		.dpdk_name = "tx_errors_phy",
2118 		.ctr_name = "tx_errors_phy",
2119 	},
2120 	{
2121 		.dpdk_name = "rx_out_of_buffer",
2122 		.ctr_name = "out_of_buffer",
2123 		.dev = 1,
2124 	},
2125 	{
2126 		.dpdk_name = "tx_packets_phy",
2127 		.ctr_name = "tx_packets_phy",
2128 	},
2129 	{
2130 		.dpdk_name = "rx_packets_phy",
2131 		.ctr_name = "rx_packets_phy",
2132 	},
2133 	{
2134 		.dpdk_name = "tx_discards_phy",
2135 		.ctr_name = "tx_discards_phy",
2136 	},
2137 	{
2138 		.dpdk_name = "rx_discards_phy",
2139 		.ctr_name = "rx_discards_phy",
2140 	},
2141 	{
2142 		.dpdk_name = "tx_bytes_phy",
2143 		.ctr_name = "tx_bytes_phy",
2144 	},
2145 	{
2146 		.dpdk_name = "rx_bytes_phy",
2147 		.ctr_name = "rx_bytes_phy",
2148 	},
2149 	/* Representor only */
2150 	{
2151 		.dpdk_name = "rx_packets",
2152 		.ctr_name = "vport_rx_packets",
2153 	},
2154 	{
2155 		.dpdk_name = "rx_bytes",
2156 		.ctr_name = "vport_rx_bytes",
2157 	},
2158 	{
2159 		.dpdk_name = "tx_packets",
2160 		.ctr_name = "vport_tx_packets",
2161 	},
2162 	{
2163 		.dpdk_name = "tx_bytes",
2164 		.ctr_name = "vport_tx_bytes",
2165 	},
2166 };
2167 
2168 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2169 
2170 /**
2171  * Init the structures to read device counters.
2172  *
2173  * @param dev
2174  *   Pointer to Ethernet device.
2175  */
2176 void
2177 mlx5_os_stats_init(struct rte_eth_dev *dev)
2178 {
2179 	struct mlx5_priv *priv = dev->data->dev_private;
2180 	struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2181 	struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2182 	unsigned int i;
2183 	unsigned int j;
2184 	struct ifreq ifr;
2185 	struct ethtool_gstrings *strings = NULL;
2186 	unsigned int dev_stats_n;
2187 	unsigned int str_sz;
2188 	int ret;
2189 
2190 	/* So that it won't aggregate for each init. */
2191 	xstats_ctrl->mlx5_stats_n = 0;
2192 	ret = mlx5_os_get_stats_n(dev);
2193 	if (ret < 0) {
2194 		DRV_LOG(WARNING, "port %u no extended statistics available",
2195 			dev->data->port_id);
2196 		return;
2197 	}
2198 	dev_stats_n = ret;
2199 	/* Allocate memory to grab stat names and values. */
2200 	str_sz = dev_stats_n * ETH_GSTRING_LEN;
2201 	strings = (struct ethtool_gstrings *)
2202 		  rte_malloc("xstats_strings",
2203 			     str_sz + sizeof(struct ethtool_gstrings), 0);
2204 	if (!strings) {
2205 		DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2206 		     dev->data->port_id);
2207 		return;
2208 	}
2209 	strings->cmd = ETHTOOL_GSTRINGS;
2210 	strings->string_set = ETH_SS_STATS;
2211 	strings->len = dev_stats_n;
2212 	ifr.ifr_data = (caddr_t)strings;
2213 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2214 	if (ret) {
2215 		DRV_LOG(WARNING, "port %u unable to get statistic names",
2216 			dev->data->port_id);
2217 		goto free;
2218 	}
2219 	for (i = 0; i != dev_stats_n; ++i) {
2220 		const char *curr_string = (const char *)
2221 			&strings->data[i * ETH_GSTRING_LEN];
2222 
2223 		for (j = 0; j != xstats_n; ++j) {
2224 			if (!strcmp(mlx5_counters_init[j].ctr_name,
2225 				    curr_string)) {
2226 				unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2227 
2228 				xstats_ctrl->dev_table_idx[idx] = i;
2229 				xstats_ctrl->info[idx] = mlx5_counters_init[j];
2230 				break;
2231 			}
2232 		}
2233 	}
2234 	/* Add dev counters. */
2235 	for (i = 0; i != xstats_n; ++i) {
2236 		if (mlx5_counters_init[i].dev) {
2237 			unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2238 
2239 			xstats_ctrl->info[idx] = mlx5_counters_init[i];
2240 			xstats_ctrl->hw_stats[idx] = 0;
2241 		}
2242 	}
2243 	MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2244 	xstats_ctrl->stats_n = dev_stats_n;
2245 	/* Copy to base at first time. */
2246 	ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2247 	if (ret)
2248 		DRV_LOG(ERR, "port %u cannot read device counters: %s",
2249 			dev->data->port_id, strerror(rte_errno));
2250 	mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2251 	stats_ctrl->imissed = 0;
2252 free:
2253 	rte_free(strings);
2254 }
2255 
2256 /**
2257  * Set the reg_mr and dereg_mr call backs
2258  *
2259  * @param reg_mr_cb[out]
2260  *   Pointer to reg_mr func
2261  * @param dereg_mr_cb[out]
2262  *   Pointer to dereg_mr func
2263  *
2264  */
2265 void
2266 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2267 		      mlx5_dereg_mr_t *dereg_mr_cb)
2268 {
2269 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2270 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2271 }
2272 
2273 const struct eth_dev_ops mlx5_os_dev_ops = {
2274 	.dev_configure = mlx5_dev_configure,
2275 	.dev_start = mlx5_dev_start,
2276 	.dev_stop = mlx5_dev_stop,
2277 	.dev_set_link_down = mlx5_set_link_down,
2278 	.dev_set_link_up = mlx5_set_link_up,
2279 	.dev_close = mlx5_dev_close,
2280 	.promiscuous_enable = mlx5_promiscuous_enable,
2281 	.promiscuous_disable = mlx5_promiscuous_disable,
2282 	.allmulticast_enable = mlx5_allmulticast_enable,
2283 	.allmulticast_disable = mlx5_allmulticast_disable,
2284 	.link_update = mlx5_link_update,
2285 	.stats_get = mlx5_stats_get,
2286 	.stats_reset = mlx5_stats_reset,
2287 	.xstats_get = mlx5_xstats_get,
2288 	.xstats_reset = mlx5_xstats_reset,
2289 	.xstats_get_names = mlx5_xstats_get_names,
2290 	.fw_version_get = mlx5_fw_version_get,
2291 	.dev_infos_get = mlx5_dev_infos_get,
2292 	.read_clock = mlx5_read_clock,
2293 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2294 	.vlan_filter_set = mlx5_vlan_filter_set,
2295 	.rx_queue_setup = mlx5_rx_queue_setup,
2296 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2297 	.tx_queue_setup = mlx5_tx_queue_setup,
2298 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2299 	.rx_queue_release = mlx5_rx_queue_release,
2300 	.tx_queue_release = mlx5_tx_queue_release,
2301 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2302 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2303 	.mac_addr_remove = mlx5_mac_addr_remove,
2304 	.mac_addr_add = mlx5_mac_addr_add,
2305 	.mac_addr_set = mlx5_mac_addr_set,
2306 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2307 	.mtu_set = mlx5_dev_set_mtu,
2308 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2309 	.vlan_offload_set = mlx5_vlan_offload_set,
2310 	.reta_update = mlx5_dev_rss_reta_update,
2311 	.reta_query = mlx5_dev_rss_reta_query,
2312 	.rss_hash_update = mlx5_rss_hash_update,
2313 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2314 	.filter_ctrl = mlx5_dev_filter_ctrl,
2315 	.rx_descriptor_status = mlx5_rx_descriptor_status,
2316 	.tx_descriptor_status = mlx5_tx_descriptor_status,
2317 	.rxq_info_get = mlx5_rxq_info_get,
2318 	.txq_info_get = mlx5_txq_info_get,
2319 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2320 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2321 	.rx_queue_count = mlx5_rx_queue_count,
2322 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2323 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2324 	.is_removed = mlx5_is_removed,
2325 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2326 	.get_module_info = mlx5_get_module_info,
2327 	.get_module_eeprom = mlx5_get_module_eeprom,
2328 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2329 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2330 };
2331 
2332 /* Available operations from secondary process. */
2333 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2334 	.stats_get = mlx5_stats_get,
2335 	.stats_reset = mlx5_stats_reset,
2336 	.xstats_get = mlx5_xstats_get,
2337 	.xstats_reset = mlx5_xstats_reset,
2338 	.xstats_get_names = mlx5_xstats_get_names,
2339 	.fw_version_get = mlx5_fw_version_get,
2340 	.dev_infos_get = mlx5_dev_infos_get,
2341 	.rx_descriptor_status = mlx5_rx_descriptor_status,
2342 	.tx_descriptor_status = mlx5_tx_descriptor_status,
2343 	.rxq_info_get = mlx5_rxq_info_get,
2344 	.txq_info_get = mlx5_txq_info_get,
2345 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2346 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2347 	.get_module_info = mlx5_get_module_info,
2348 	.get_module_eeprom = mlx5_get_module_eeprom,
2349 };
2350 
2351 /* Available operations in flow isolated mode. */
2352 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2353 	.dev_configure = mlx5_dev_configure,
2354 	.dev_start = mlx5_dev_start,
2355 	.dev_stop = mlx5_dev_stop,
2356 	.dev_set_link_down = mlx5_set_link_down,
2357 	.dev_set_link_up = mlx5_set_link_up,
2358 	.dev_close = mlx5_dev_close,
2359 	.promiscuous_enable = mlx5_promiscuous_enable,
2360 	.promiscuous_disable = mlx5_promiscuous_disable,
2361 	.allmulticast_enable = mlx5_allmulticast_enable,
2362 	.allmulticast_disable = mlx5_allmulticast_disable,
2363 	.link_update = mlx5_link_update,
2364 	.stats_get = mlx5_stats_get,
2365 	.stats_reset = mlx5_stats_reset,
2366 	.xstats_get = mlx5_xstats_get,
2367 	.xstats_reset = mlx5_xstats_reset,
2368 	.xstats_get_names = mlx5_xstats_get_names,
2369 	.fw_version_get = mlx5_fw_version_get,
2370 	.dev_infos_get = mlx5_dev_infos_get,
2371 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2372 	.vlan_filter_set = mlx5_vlan_filter_set,
2373 	.rx_queue_setup = mlx5_rx_queue_setup,
2374 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2375 	.tx_queue_setup = mlx5_tx_queue_setup,
2376 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2377 	.rx_queue_release = mlx5_rx_queue_release,
2378 	.tx_queue_release = mlx5_tx_queue_release,
2379 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2380 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2381 	.mac_addr_remove = mlx5_mac_addr_remove,
2382 	.mac_addr_add = mlx5_mac_addr_add,
2383 	.mac_addr_set = mlx5_mac_addr_set,
2384 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2385 	.mtu_set = mlx5_dev_set_mtu,
2386 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2387 	.vlan_offload_set = mlx5_vlan_offload_set,
2388 	.filter_ctrl = mlx5_dev_filter_ctrl,
2389 	.rx_descriptor_status = mlx5_rx_descriptor_status,
2390 	.tx_descriptor_status = mlx5_tx_descriptor_status,
2391 	.rxq_info_get = mlx5_rxq_info_get,
2392 	.txq_info_get = mlx5_txq_info_get,
2393 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2394 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2395 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2396 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2397 	.is_removed = mlx5_is_removed,
2398 	.get_module_info = mlx5_get_module_info,
2399 	.get_module_eeprom = mlx5_get_module_eeprom,
2400 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2401 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2402 };
2403