xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision ab27cdd93a1cc605171e162b3f5bca6699ea42cc)
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 /* Verbs header. */
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27 
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
31 #include <rte_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_kvargs.h>
35 #include <rte_rwlock.h>
36 #include <rte_spinlock.h>
37 #include <rte_string_fns.h>
38 #include <rte_alarm.h>
39 #include <rte_eal_paging.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 #include <mlx5_malloc.h>
47 
48 #include "mlx5_defs.h"
49 #include "mlx5.h"
50 #include "mlx5_common_os.h"
51 #include "mlx5_utils.h"
52 #include "mlx5_rxtx.h"
53 #include "mlx5_autoconf.h"
54 #include "mlx5_mr.h"
55 #include "mlx5_flow.h"
56 #include "rte_pmd_mlx5.h"
57 #include "mlx5_verbs.h"
58 
59 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
60 
61 #ifndef HAVE_IBV_MLX5_MOD_MPW
62 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
63 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
64 #endif
65 
66 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
67 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
68 #endif
69 
70 /**
71  * Get MAC address by querying netdevice.
72  *
73  * @param[in] dev
74  *   Pointer to Ethernet device.
75  * @param[out] mac
76  *   MAC address output buffer.
77  *
78  * @return
79  *   0 on success, a negative errno value otherwise and rte_errno is set.
80  */
81 static int
82 mlx5_get_mac(struct rte_eth_dev *dev, uint8_t (*mac)[RTE_ETHER_ADDR_LEN])
83 {
84 	struct ifreq request;
85 	int ret;
86 
87 	ret = mlx5_ifreq(dev, SIOCGIFHWADDR, &request);
88 	if (ret)
89 		return ret;
90 	memcpy(mac, request.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
91 	return 0;
92 }
93 
94 /**
95  * Get mlx5 device attributes. The glue function query_device_ex() is called
96  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
97  * device attributes from the glue out parameter.
98  *
99  * @param dev
100  *   Pointer to ibv context.
101  *
102  * @param device_attr
103  *   Pointer to mlx5 device attributes.
104  *
105  * @return
106  *   0 on success, non zero error number otherwise
107  */
108 int
109 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
110 {
111 	int err;
112 	struct ibv_device_attr_ex attr_ex;
113 	memset(device_attr, 0, sizeof(*device_attr));
114 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
115 	if (err)
116 		return err;
117 
118 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
119 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
120 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
121 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
122 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
123 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
124 	device_attr->max_rwq_indirection_table_size =
125 		attr_ex.rss_caps.max_rwq_indirection_table_size;
126 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
127 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
128 
129 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
130 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
131 	if (err)
132 		return err;
133 
134 	device_attr->flags = dv_attr.flags;
135 	device_attr->comp_mask = dv_attr.comp_mask;
136 #ifdef HAVE_IBV_MLX5_MOD_SWP
137 	device_attr->sw_parsing_offloads =
138 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
139 #endif
140 	device_attr->min_single_stride_log_num_of_bytes =
141 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
142 	device_attr->max_single_stride_log_num_of_bytes =
143 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
144 	device_attr->min_single_wqe_log_num_of_strides =
145 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
146 	device_attr->max_single_wqe_log_num_of_strides =
147 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
148 	device_attr->stride_supported_qpts =
149 		dv_attr.striding_rq_caps.supported_qpts;
150 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
151 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
152 #endif
153 
154 	return err;
155 }
156 
157 /**
158  * Verbs callback to allocate a memory. This function should allocate the space
159  * according to the size provided residing inside a huge page.
160  * Please note that all allocation must respect the alignment from libmlx5
161  * (i.e. currently rte_mem_page_size()).
162  *
163  * @param[in] size
164  *   The size in bytes of the memory to allocate.
165  * @param[in] data
166  *   A pointer to the callback data.
167  *
168  * @return
169  *   Allocated buffer, NULL otherwise and rte_errno is set.
170  */
171 static void *
172 mlx5_alloc_verbs_buf(size_t size, void *data)
173 {
174 	struct mlx5_priv *priv = data;
175 	void *ret;
176 	unsigned int socket = SOCKET_ID_ANY;
177 	size_t alignment = rte_mem_page_size();
178 	if (alignment == (size_t)-1) {
179 		DRV_LOG(ERR, "Failed to get mem page size");
180 		rte_errno = ENOMEM;
181 		return NULL;
182 	}
183 
184 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
185 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
186 
187 		socket = ctrl->socket;
188 	} else if (priv->verbs_alloc_ctx.type ==
189 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
190 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
191 
192 		socket = ctrl->socket;
193 	}
194 	MLX5_ASSERT(data != NULL);
195 	ret = mlx5_malloc(0, size, alignment, socket);
196 	if (!ret && size)
197 		rte_errno = ENOMEM;
198 	return ret;
199 }
200 
201 /**
202  * Verbs callback to free a memory.
203  *
204  * @param[in] ptr
205  *   A pointer to the memory to free.
206  * @param[in] data
207  *   A pointer to the callback data.
208  */
209 static void
210 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
211 {
212 	MLX5_ASSERT(data != NULL);
213 	mlx5_free(ptr);
214 }
215 
216 /**
217  * Initialize DR related data within private structure.
218  * Routine checks the reference counter and does actual
219  * resources creation/initialization only if counter is zero.
220  *
221  * @param[in] priv
222  *   Pointer to the private device data structure.
223  *
224  * @return
225  *   Zero on success, positive error code otherwise.
226  */
227 static int
228 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
229 {
230 	struct mlx5_dev_ctx_shared *sh = priv->sh;
231 	char s[MLX5_HLIST_NAMESIZE];
232 	int err = 0;
233 
234 	if (!sh->flow_tbls)
235 		err = mlx5_alloc_table_hash_list(priv);
236 	else
237 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
238 			(void *)sh->flow_tbls);
239 	if (err)
240 		return err;
241 	/* Create tags hash list table. */
242 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
243 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
244 	if (!sh->tag_table) {
245 		DRV_LOG(ERR, "tags with hash creation failed.");
246 		err = ENOMEM;
247 		goto error;
248 	}
249 #ifdef HAVE_MLX5DV_DR
250 	void *domain;
251 
252 	if (sh->dv_refcnt) {
253 		/* Shared DV/DR structures is already initialized. */
254 		sh->dv_refcnt++;
255 		priv->dr_shared = 1;
256 		return 0;
257 	}
258 	/* Reference counter is zero, we should initialize structures. */
259 	domain = mlx5_glue->dr_create_domain(sh->ctx,
260 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
261 	if (!domain) {
262 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
263 		err = errno;
264 		goto error;
265 	}
266 	sh->rx_domain = domain;
267 	domain = mlx5_glue->dr_create_domain(sh->ctx,
268 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
269 	if (!domain) {
270 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
271 		err = errno;
272 		goto error;
273 	}
274 	pthread_mutex_init(&sh->dv_mutex, NULL);
275 	sh->tx_domain = domain;
276 #ifdef HAVE_MLX5DV_DR_ESWITCH
277 	if (priv->config.dv_esw_en) {
278 		domain  = mlx5_glue->dr_create_domain
279 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
280 		if (!domain) {
281 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
282 			err = errno;
283 			goto error;
284 		}
285 		sh->fdb_domain = domain;
286 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
287 	}
288 #endif
289 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
290 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
291 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
292 		if (sh->fdb_domain)
293 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
294 	}
295 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
296 #endif /* HAVE_MLX5DV_DR */
297 	sh->dv_refcnt++;
298 	priv->dr_shared = 1;
299 	return 0;
300 error:
301 	/* Rollback the created objects. */
302 	if (sh->rx_domain) {
303 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
304 		sh->rx_domain = NULL;
305 	}
306 	if (sh->tx_domain) {
307 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
308 		sh->tx_domain = NULL;
309 	}
310 	if (sh->fdb_domain) {
311 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
312 		sh->fdb_domain = NULL;
313 	}
314 	if (sh->esw_drop_action) {
315 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
316 		sh->esw_drop_action = NULL;
317 	}
318 	if (sh->pop_vlan_action) {
319 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
320 		sh->pop_vlan_action = NULL;
321 	}
322 	if (sh->tag_table) {
323 		/* tags should be destroyed with flow before. */
324 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
325 		sh->tag_table = NULL;
326 	}
327 	mlx5_free_table_hash_list(priv);
328 	return err;
329 }
330 
331 /**
332  * Destroy DR related data within private structure.
333  *
334  * @param[in] priv
335  *   Pointer to the private device data structure.
336  */
337 void
338 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
339 {
340 	struct mlx5_dev_ctx_shared *sh;
341 
342 	if (!priv->dr_shared)
343 		return;
344 	priv->dr_shared = 0;
345 	sh = priv->sh;
346 	MLX5_ASSERT(sh);
347 #ifdef HAVE_MLX5DV_DR
348 	MLX5_ASSERT(sh->dv_refcnt);
349 	if (sh->dv_refcnt && --sh->dv_refcnt)
350 		return;
351 	if (sh->rx_domain) {
352 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
353 		sh->rx_domain = NULL;
354 	}
355 	if (sh->tx_domain) {
356 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
357 		sh->tx_domain = NULL;
358 	}
359 #ifdef HAVE_MLX5DV_DR_ESWITCH
360 	if (sh->fdb_domain) {
361 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
362 		sh->fdb_domain = NULL;
363 	}
364 	if (sh->esw_drop_action) {
365 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
366 		sh->esw_drop_action = NULL;
367 	}
368 #endif
369 	if (sh->pop_vlan_action) {
370 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
371 		sh->pop_vlan_action = NULL;
372 	}
373 	pthread_mutex_destroy(&sh->dv_mutex);
374 #endif /* HAVE_MLX5DV_DR */
375 	if (sh->tag_table) {
376 		/* tags should be destroyed with flow before. */
377 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
378 		sh->tag_table = NULL;
379 	}
380 	mlx5_free_table_hash_list(priv);
381 }
382 
383 /**
384  * Spawn an Ethernet device from Verbs information.
385  *
386  * @param dpdk_dev
387  *   Backing DPDK device.
388  * @param spawn
389  *   Verbs device parameters (name, port, switch_info) to spawn.
390  * @param config
391  *   Device configuration parameters.
392  *
393  * @return
394  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
395  *   is set. The following errors are defined:
396  *
397  *   EBUSY: device is not supposed to be spawned.
398  *   EEXIST: device is already spawned
399  */
400 static struct rte_eth_dev *
401 mlx5_dev_spawn(struct rte_device *dpdk_dev,
402 	       struct mlx5_dev_spawn_data *spawn,
403 	       struct mlx5_dev_config config)
404 {
405 	const struct mlx5_switch_info *switch_info = &spawn->info;
406 	struct mlx5_dev_ctx_shared *sh = NULL;
407 	struct ibv_port_attr port_attr;
408 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
409 	struct rte_eth_dev *eth_dev = NULL;
410 	struct mlx5_priv *priv = NULL;
411 	int err = 0;
412 	unsigned int hw_padding = 0;
413 	unsigned int mps;
414 	unsigned int cqe_comp;
415 	unsigned int cqe_pad = 0;
416 	unsigned int tunnel_en = 0;
417 	unsigned int mpls_en = 0;
418 	unsigned int swp = 0;
419 	unsigned int mprq = 0;
420 	unsigned int mprq_min_stride_size_n = 0;
421 	unsigned int mprq_max_stride_size_n = 0;
422 	unsigned int mprq_min_stride_num_n = 0;
423 	unsigned int mprq_max_stride_num_n = 0;
424 	struct rte_ether_addr mac;
425 	char name[RTE_ETH_NAME_MAX_LEN];
426 	int own_domain_id = 0;
427 	uint16_t port_id;
428 	unsigned int i;
429 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
430 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
431 #endif
432 
433 	/* Determine if this port representor is supposed to be spawned. */
434 	if (switch_info->representor && dpdk_dev->devargs) {
435 		struct rte_eth_devargs eth_da;
436 
437 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
438 		if (err) {
439 			rte_errno = -err;
440 			DRV_LOG(ERR, "failed to process device arguments: %s",
441 				strerror(rte_errno));
442 			return NULL;
443 		}
444 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
445 			if (eth_da.representor_ports[i] ==
446 			    (uint16_t)switch_info->port_name)
447 				break;
448 		if (i == eth_da.nb_representor_ports) {
449 			rte_errno = EBUSY;
450 			return NULL;
451 		}
452 	}
453 	/* Build device name. */
454 	if (spawn->pf_bond <  0) {
455 		/* Single device. */
456 		if (!switch_info->representor)
457 			strlcpy(name, dpdk_dev->name, sizeof(name));
458 		else
459 			snprintf(name, sizeof(name), "%s_representor_%u",
460 				 dpdk_dev->name, switch_info->port_name);
461 	} else {
462 		/* Bonding device. */
463 		if (!switch_info->representor)
464 			snprintf(name, sizeof(name), "%s_%s",
465 				 dpdk_dev->name,
466 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
467 		else
468 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
469 				 dpdk_dev->name,
470 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
471 				 switch_info->port_name);
472 	}
473 	/* check if the device is already spawned */
474 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
475 		rte_errno = EEXIST;
476 		return NULL;
477 	}
478 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
479 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
480 		struct mlx5_mp_id mp_id;
481 
482 		eth_dev = rte_eth_dev_attach_secondary(name);
483 		if (eth_dev == NULL) {
484 			DRV_LOG(ERR, "can not attach rte ethdev");
485 			rte_errno = ENOMEM;
486 			return NULL;
487 		}
488 		eth_dev->device = dpdk_dev;
489 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
490 		err = mlx5_proc_priv_init(eth_dev);
491 		if (err)
492 			return NULL;
493 		mp_id.port_id = eth_dev->data->port_id;
494 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
495 		/* Receive command fd from primary process */
496 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
497 		if (err < 0)
498 			goto err_secondary;
499 		/* Remap UAR for Tx queues. */
500 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
501 		if (err)
502 			goto err_secondary;
503 		/*
504 		 * Ethdev pointer is still required as input since
505 		 * the primary device is not accessible from the
506 		 * secondary process.
507 		 */
508 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
509 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
510 		return eth_dev;
511 err_secondary:
512 		mlx5_dev_close(eth_dev);
513 		return NULL;
514 	}
515 	/*
516 	 * Some parameters ("tx_db_nc" in particularly) are needed in
517 	 * advance to create dv/verbs device context. We proceed the
518 	 * devargs here to get ones, and later proceed devargs again
519 	 * to override some hardware settings.
520 	 */
521 	err = mlx5_args(&config, dpdk_dev->devargs);
522 	if (err) {
523 		err = rte_errno;
524 		DRV_LOG(ERR, "failed to process device arguments: %s",
525 			strerror(rte_errno));
526 		goto error;
527 	}
528 	mlx5_malloc_mem_select(config.sys_mem_en);
529 	sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
530 	if (!sh)
531 		return NULL;
532 	config.devx = sh->devx;
533 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
534 	config.dest_tir = 1;
535 #endif
536 #ifdef HAVE_IBV_MLX5_MOD_SWP
537 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
538 #endif
539 	/*
540 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
541 	 * as all ConnectX-5 devices.
542 	 */
543 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
544 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
545 #endif
546 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
547 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
548 #endif
549 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
550 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
551 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
552 			DRV_LOG(DEBUG, "enhanced MPW is supported");
553 			mps = MLX5_MPW_ENHANCED;
554 		} else {
555 			DRV_LOG(DEBUG, "MPW is supported");
556 			mps = MLX5_MPW;
557 		}
558 	} else {
559 		DRV_LOG(DEBUG, "MPW isn't supported");
560 		mps = MLX5_MPW_DISABLED;
561 	}
562 #ifdef HAVE_IBV_MLX5_MOD_SWP
563 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
564 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
565 	DRV_LOG(DEBUG, "SWP support: %u", swp);
566 #endif
567 	config.swp = !!swp;
568 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
569 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
570 		struct mlx5dv_striding_rq_caps mprq_caps =
571 			dv_attr.striding_rq_caps;
572 
573 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
574 			mprq_caps.min_single_stride_log_num_of_bytes);
575 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
576 			mprq_caps.max_single_stride_log_num_of_bytes);
577 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
578 			mprq_caps.min_single_wqe_log_num_of_strides);
579 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
580 			mprq_caps.max_single_wqe_log_num_of_strides);
581 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
582 			mprq_caps.supported_qpts);
583 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
584 		mprq = 1;
585 		mprq_min_stride_size_n =
586 			mprq_caps.min_single_stride_log_num_of_bytes;
587 		mprq_max_stride_size_n =
588 			mprq_caps.max_single_stride_log_num_of_bytes;
589 		mprq_min_stride_num_n =
590 			mprq_caps.min_single_wqe_log_num_of_strides;
591 		mprq_max_stride_num_n =
592 			mprq_caps.max_single_wqe_log_num_of_strides;
593 	}
594 #endif
595 	if (RTE_CACHE_LINE_SIZE == 128 &&
596 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
597 		cqe_comp = 0;
598 	else
599 		cqe_comp = 1;
600 	config.cqe_comp = cqe_comp;
601 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
602 	/* Whether device supports 128B Rx CQE padding. */
603 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
604 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
605 #endif
606 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
607 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
608 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
609 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
610 			     (dv_attr.tunnel_offloads_caps &
611 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
612 			     (dv_attr.tunnel_offloads_caps &
613 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
614 	}
615 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
616 		tunnel_en ? "" : "not ");
617 #else
618 	DRV_LOG(WARNING,
619 		"tunnel offloading disabled due to old OFED/rdma-core version");
620 #endif
621 	config.tunnel_en = tunnel_en;
622 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
623 	mpls_en = ((dv_attr.tunnel_offloads_caps &
624 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
625 		   (dv_attr.tunnel_offloads_caps &
626 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
627 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
628 		mpls_en ? "" : "not ");
629 #else
630 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
631 		" old OFED/rdma-core version or firmware configuration");
632 #endif
633 	config.mpls_en = mpls_en;
634 	/* Check port status. */
635 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
636 	if (err) {
637 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
638 		goto error;
639 	}
640 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
641 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
642 		err = EINVAL;
643 		goto error;
644 	}
645 	if (port_attr.state != IBV_PORT_ACTIVE)
646 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
647 			mlx5_glue->port_state_str(port_attr.state),
648 			port_attr.state);
649 	/* Allocate private eth device data. */
650 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
651 			   sizeof(*priv),
652 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
653 	if (priv == NULL) {
654 		DRV_LOG(ERR, "priv allocation failure");
655 		err = ENOMEM;
656 		goto error;
657 	}
658 	priv->sh = sh;
659 	priv->dev_port = spawn->phys_port;
660 	priv->pci_dev = spawn->pci_dev;
661 	priv->mtu = RTE_ETHER_MTU;
662 	priv->mp_id.port_id = port_id;
663 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
664 	/* Some internal functions rely on Netlink sockets, open them now. */
665 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
666 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
667 	priv->representor = !!switch_info->representor;
668 	priv->master = !!switch_info->master;
669 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
670 	priv->vport_meta_tag = 0;
671 	priv->vport_meta_mask = 0;
672 	priv->pf_bond = spawn->pf_bond;
673 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
674 	/*
675 	 * The DevX port query API is implemented. E-Switch may use
676 	 * either vport or reg_c[0] metadata register to match on
677 	 * vport index. The engaged part of metadata register is
678 	 * defined by mask.
679 	 */
680 	if (switch_info->representor || switch_info->master) {
681 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
682 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
683 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
684 						 &devx_port);
685 		if (err) {
686 			DRV_LOG(WARNING,
687 				"can't query devx port %d on device %s",
688 				spawn->phys_port,
689 				mlx5_os_get_dev_device_name(spawn->phys_dev));
690 			devx_port.comp_mask = 0;
691 		}
692 	}
693 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
694 		priv->vport_meta_tag = devx_port.reg_c_0.value;
695 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
696 		if (!priv->vport_meta_mask) {
697 			DRV_LOG(ERR, "vport zero mask for port %d"
698 				     " on bonding device %s",
699 				     spawn->phys_port,
700 				     mlx5_os_get_dev_device_name
701 							(spawn->phys_dev));
702 			err = ENOTSUP;
703 			goto error;
704 		}
705 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
706 			DRV_LOG(ERR, "invalid vport tag for port %d"
707 				     " on bonding device %s",
708 				     spawn->phys_port,
709 				     mlx5_os_get_dev_device_name
710 							(spawn->phys_dev));
711 			err = ENOTSUP;
712 			goto error;
713 		}
714 	}
715 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
716 		priv->vport_id = devx_port.vport_num;
717 	} else if (spawn->pf_bond >= 0) {
718 		DRV_LOG(ERR, "can't deduce vport index for port %d"
719 			     " on bonding device %s",
720 			     spawn->phys_port,
721 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
722 		err = ENOTSUP;
723 		goto error;
724 	} else {
725 		/* Suppose vport index in compatible way. */
726 		priv->vport_id = switch_info->representor ?
727 				 switch_info->port_name + 1 : -1;
728 	}
729 #else
730 	/*
731 	 * Kernel/rdma_core support single E-Switch per PF configurations
732 	 * only and vport_id field contains the vport index for
733 	 * associated VF, which is deduced from representor port name.
734 	 * For example, let's have the IB device port 10, it has
735 	 * attached network device eth0, which has port name attribute
736 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
737 	 * as 3 (2+1). This assigning schema should be changed if the
738 	 * multiple E-Switch instances per PF configurations or/and PCI
739 	 * subfunctions are added.
740 	 */
741 	priv->vport_id = switch_info->representor ?
742 			 switch_info->port_name + 1 : -1;
743 #endif
744 	/* representor_id field keeps the unmodified VF index. */
745 	priv->representor_id = switch_info->representor ?
746 			       switch_info->port_name : -1;
747 	/*
748 	 * Look for sibling devices in order to reuse their switch domain
749 	 * if any, otherwise allocate one.
750 	 */
751 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
752 		const struct mlx5_priv *opriv =
753 			rte_eth_devices[port_id].data->dev_private;
754 
755 		if (!opriv ||
756 		    opriv->sh != priv->sh ||
757 			opriv->domain_id ==
758 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
759 			continue;
760 		priv->domain_id = opriv->domain_id;
761 		break;
762 	}
763 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
764 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
765 		if (err) {
766 			err = rte_errno;
767 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
768 				strerror(rte_errno));
769 			goto error;
770 		}
771 		own_domain_id = 1;
772 	}
773 	/* Override some values set by hardware configuration. */
774 	mlx5_args(&config, dpdk_dev->devargs);
775 	err = mlx5_dev_check_sibling_config(priv, &config);
776 	if (err)
777 		goto error;
778 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
779 			    IBV_DEVICE_RAW_IP_CSUM);
780 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
781 		(config.hw_csum ? "" : "not "));
782 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
783 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
784 	DRV_LOG(DEBUG, "counters are not supported");
785 #endif
786 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
787 	if (config.dv_flow_en) {
788 		DRV_LOG(WARNING, "DV flow is not supported");
789 		config.dv_flow_en = 0;
790 	}
791 #endif
792 	config.ind_table_max_size =
793 		sh->device_attr.max_rwq_indirection_table_size;
794 	/*
795 	 * Remove this check once DPDK supports larger/variable
796 	 * indirection tables.
797 	 */
798 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
799 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
800 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
801 		config.ind_table_max_size);
802 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
803 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
804 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
805 		(config.hw_vlan_strip ? "" : "not "));
806 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
807 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
808 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
809 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
810 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
811 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
812 			IBV_DEVICE_PCI_WRITE_END_PADDING);
813 #endif
814 	if (config.hw_padding && !hw_padding) {
815 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
816 		config.hw_padding = 0;
817 	} else if (config.hw_padding) {
818 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
819 	}
820 	config.tso = (sh->device_attr.max_tso > 0 &&
821 		      (sh->device_attr.tso_supported_qpts &
822 		       (1 << IBV_QPT_RAW_PACKET)));
823 	if (config.tso)
824 		config.tso_max_payload_sz = sh->device_attr.max_tso;
825 	/*
826 	 * MPW is disabled by default, while the Enhanced MPW is enabled
827 	 * by default.
828 	 */
829 	if (config.mps == MLX5_ARG_UNSET)
830 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
831 							  MLX5_MPW_DISABLED;
832 	else
833 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
834 	DRV_LOG(INFO, "%sMPS is %s",
835 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
836 		config.mps == MLX5_MPW ? "legacy " : "",
837 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
838 	if (config.cqe_comp && !cqe_comp) {
839 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
840 		config.cqe_comp = 0;
841 	}
842 	if (config.cqe_pad && !cqe_pad) {
843 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
844 		config.cqe_pad = 0;
845 	} else if (config.cqe_pad) {
846 		DRV_LOG(INFO, "Rx CQE padding is enabled");
847 	}
848 	if (config.devx) {
849 		priv->counter_fallback = 0;
850 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
851 		if (err) {
852 			err = -err;
853 			goto error;
854 		}
855 		if (!config.hca_attr.flow_counters_dump)
856 			priv->counter_fallback = 1;
857 #ifndef HAVE_IBV_DEVX_ASYNC
858 		priv->counter_fallback = 1;
859 #endif
860 		if (priv->counter_fallback)
861 			DRV_LOG(INFO, "Use fall-back DV counter management");
862 		/* Check for LRO support. */
863 		if (config.dest_tir && config.hca_attr.lro_cap &&
864 		    config.dv_flow_en) {
865 			/* TBD check tunnel lro caps. */
866 			config.lro.supported = config.hca_attr.lro_cap;
867 			DRV_LOG(DEBUG, "Device supports LRO");
868 			/*
869 			 * If LRO timeout is not configured by application,
870 			 * use the minimal supported value.
871 			 */
872 			if (!config.lro.timeout)
873 				config.lro.timeout =
874 				config.hca_attr.lro_timer_supported_periods[0];
875 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
876 				config.lro.timeout);
877 		}
878 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
879 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
880 		    config.dv_flow_en) {
881 			uint8_t reg_c_mask =
882 				config.hca_attr.qos.flow_meter_reg_c_ids;
883 			/*
884 			 * Meter needs two REG_C's for color match and pre-sfx
885 			 * flow match. Here get the REG_C for color match.
886 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
887 			 */
888 			reg_c_mask &= 0xfc;
889 			if (__builtin_popcount(reg_c_mask) < 1) {
890 				priv->mtr_en = 0;
891 				DRV_LOG(WARNING, "No available register for"
892 					" meter.");
893 			} else {
894 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
895 						      REG_C_0;
896 				priv->mtr_en = 1;
897 				priv->mtr_reg_share =
898 				      config.hca_attr.qos.flow_meter_reg_share;
899 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
900 					priv->mtr_color_reg);
901 			}
902 		}
903 #endif
904 	}
905 	if (config.tx_pp) {
906 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
907 			config.hca_attr.dev_freq_khz);
908 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
909 			config.hca_attr.qos.packet_pacing ? "" : "not ");
910 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
911 			config.hca_attr.cross_channel ? "" : "not ");
912 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
913 			config.hca_attr.wqe_index_ignore ? "" : "not ");
914 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
915 			config.hca_attr.non_wire_sq ? "" : "not ");
916 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
917 			config.hca_attr.log_max_static_sq_wq ? "" : "not ",
918 			config.hca_attr.log_max_static_sq_wq);
919 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
920 			config.hca_attr.qos.wqe_rate_pp ? "" : "not ");
921 		if (!config.devx) {
922 			DRV_LOG(ERR, "DevX is required for packet pacing");
923 			err = ENODEV;
924 			goto error;
925 		}
926 		if (!config.hca_attr.qos.packet_pacing) {
927 			DRV_LOG(ERR, "Packet pacing is not supported");
928 			err = ENODEV;
929 			goto error;
930 		}
931 		if (!config.hca_attr.cross_channel) {
932 			DRV_LOG(ERR, "Cross channel operations are"
933 				     " required for packet pacing");
934 			err = ENODEV;
935 			goto error;
936 		}
937 		if (!config.hca_attr.wqe_index_ignore) {
938 			DRV_LOG(ERR, "WQE index ignore feature is"
939 				     " required for packet pacing");
940 			err = ENODEV;
941 			goto error;
942 		}
943 		if (!config.hca_attr.non_wire_sq) {
944 			DRV_LOG(ERR, "Non-wire SQ feature is"
945 				     " required for packet pacing");
946 			err = ENODEV;
947 			goto error;
948 		}
949 		if (!config.hca_attr.log_max_static_sq_wq) {
950 			DRV_LOG(ERR, "Static WQE SQ feature is"
951 				     " required for packet pacing");
952 			err = ENODEV;
953 			goto error;
954 		}
955 		if (!config.hca_attr.qos.wqe_rate_pp) {
956 			DRV_LOG(ERR, "WQE rate mode is required"
957 				     " for packet pacing");
958 			err = ENODEV;
959 			goto error;
960 		}
961 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
962 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
963 			     " can't create queues for packet pacing");
964 		err = ENODEV;
965 		goto error;
966 #endif
967 	}
968 	if (config.devx) {
969 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
970 
971 		err = mlx5_devx_cmd_register_read
972 			(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
973 			reg, MLX5_ST_SZ_DW(register_mtutc));
974 		if (!err) {
975 			uint32_t ts_mode;
976 
977 			/* MTUTC register is read successfully. */
978 			ts_mode = MLX5_GET(register_mtutc, reg,
979 					   time_stamp_mode);
980 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
981 				config.rt_timestamp = 1;
982 		} else {
983 			/* Kernel does not support register reading. */
984 			if (config.hca_attr.dev_freq_khz ==
985 						 (NS_PER_S / MS_PER_S))
986 				config.rt_timestamp = 1;
987 		}
988 	}
989 	/*
990 	 * If HW has bug working with tunnel packet decapsulation and
991 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
992 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
993 	 */
994 	if (config.hca_attr.scatter_fcs_w_decap_disable && config.decap_en)
995 		config.hw_fcs_strip = 0;
996 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
997 		(config.hw_fcs_strip ? "" : "not "));
998 	if (config.mprq.enabled && mprq) {
999 		if (config.mprq.stride_num_n &&
1000 		    (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1001 		     config.mprq.stride_num_n < mprq_min_stride_num_n)) {
1002 			config.mprq.stride_num_n =
1003 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1004 						mprq_min_stride_num_n),
1005 					mprq_max_stride_num_n);
1006 			DRV_LOG(WARNING,
1007 				"the number of strides"
1008 				" for Multi-Packet RQ is out of range,"
1009 				" setting default value (%u)",
1010 				1 << config.mprq.stride_num_n);
1011 		}
1012 		if (config.mprq.stride_size_n &&
1013 		    (config.mprq.stride_size_n > mprq_max_stride_size_n ||
1014 		     config.mprq.stride_size_n < mprq_min_stride_size_n)) {
1015 			config.mprq.stride_size_n =
1016 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1017 						mprq_min_stride_size_n),
1018 					mprq_max_stride_size_n);
1019 			DRV_LOG(WARNING,
1020 				"the size of a stride"
1021 				" for Multi-Packet RQ is out of range,"
1022 				" setting default value (%u)",
1023 				1 << config.mprq.stride_size_n);
1024 		}
1025 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1026 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1027 	} else if (config.mprq.enabled && !mprq) {
1028 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1029 		config.mprq.enabled = 0;
1030 	}
1031 	if (config.max_dump_files_num == 0)
1032 		config.max_dump_files_num = 128;
1033 	eth_dev = rte_eth_dev_allocate(name);
1034 	if (eth_dev == NULL) {
1035 		DRV_LOG(ERR, "can not allocate rte ethdev");
1036 		err = ENOMEM;
1037 		goto error;
1038 	}
1039 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1040 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1041 	if (priv->representor) {
1042 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1043 		eth_dev->data->representor_id = priv->representor_id;
1044 	}
1045 	/*
1046 	 * Store associated network device interface index. This index
1047 	 * is permanent throughout the lifetime of device. So, we may store
1048 	 * the ifindex here and use the cached value further.
1049 	 */
1050 	MLX5_ASSERT(spawn->ifindex);
1051 	priv->if_index = spawn->ifindex;
1052 	eth_dev->data->dev_private = priv;
1053 	priv->dev_data = eth_dev->data;
1054 	eth_dev->data->mac_addrs = priv->mac;
1055 	eth_dev->device = dpdk_dev;
1056 	/* Configure the first MAC address by default. */
1057 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1058 		DRV_LOG(ERR,
1059 			"port %u cannot get MAC address, is mlx5_en"
1060 			" loaded? (errno: %s)",
1061 			eth_dev->data->port_id, strerror(rte_errno));
1062 		err = ENODEV;
1063 		goto error;
1064 	}
1065 	DRV_LOG(INFO,
1066 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1067 		eth_dev->data->port_id,
1068 		mac.addr_bytes[0], mac.addr_bytes[1],
1069 		mac.addr_bytes[2], mac.addr_bytes[3],
1070 		mac.addr_bytes[4], mac.addr_bytes[5]);
1071 #ifdef RTE_LIBRTE_MLX5_DEBUG
1072 	{
1073 		char ifname[IF_NAMESIZE];
1074 
1075 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1076 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1077 				eth_dev->data->port_id, ifname);
1078 		else
1079 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1080 				eth_dev->data->port_id);
1081 	}
1082 #endif
1083 	/* Get actual MTU if possible. */
1084 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1085 	if (err) {
1086 		err = rte_errno;
1087 		goto error;
1088 	}
1089 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1090 		priv->mtu);
1091 	/* Initialize burst functions to prevent crashes before link-up. */
1092 	eth_dev->rx_pkt_burst = removed_rx_burst;
1093 	eth_dev->tx_pkt_burst = removed_tx_burst;
1094 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1095 	/* Register MAC address. */
1096 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1097 	if (config.vf && config.vf_nl_en)
1098 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1099 				      mlx5_ifindex(eth_dev),
1100 				      eth_dev->data->mac_addrs,
1101 				      MLX5_MAX_MAC_ADDRESSES);
1102 	priv->flows = 0;
1103 	priv->ctrl_flows = 0;
1104 	TAILQ_INIT(&priv->flow_meters);
1105 	TAILQ_INIT(&priv->flow_meter_profiles);
1106 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1107 	mlx5_glue->dv_set_context_attr(sh->ctx,
1108 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1109 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1110 				.alloc = &mlx5_alloc_verbs_buf,
1111 				.free = &mlx5_free_verbs_buf,
1112 				.data = priv,
1113 			}));
1114 	/* Bring Ethernet device up. */
1115 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1116 		eth_dev->data->port_id);
1117 	mlx5_set_link_up(eth_dev);
1118 	/*
1119 	 * Even though the interrupt handler is not installed yet,
1120 	 * interrupts will still trigger on the async_fd from
1121 	 * Verbs context returned by ibv_open_device().
1122 	 */
1123 	mlx5_link_update(eth_dev, 0);
1124 #ifdef HAVE_MLX5DV_DR_ESWITCH
1125 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1126 	      (switch_info->representor || switch_info->master)))
1127 		config.dv_esw_en = 0;
1128 #else
1129 	config.dv_esw_en = 0;
1130 #endif
1131 	/* Detect minimal data bytes to inline. */
1132 	mlx5_set_min_inline(spawn, &config);
1133 	/* Store device configuration on private structure. */
1134 	priv->config = config;
1135 	/* Create context for virtual machine VLAN workaround. */
1136 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1137 	if (config.dv_flow_en) {
1138 		err = mlx5_alloc_shared_dr(priv);
1139 		if (err)
1140 			goto error;
1141 		/*
1142 		 * RSS id is shared with meter flow id. Meter flow id can only
1143 		 * use the 24 MSB of the register.
1144 		 */
1145 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1146 				     MLX5_MTR_COLOR_BITS);
1147 		if (!priv->qrss_id_pool) {
1148 			DRV_LOG(ERR, "can't create flow id pool");
1149 			err = ENOMEM;
1150 			goto error;
1151 		}
1152 	}
1153 	/* Supported Verbs flow priority number detection. */
1154 	err = mlx5_flow_discover_priorities(eth_dev);
1155 	if (err < 0) {
1156 		err = -err;
1157 		goto error;
1158 	}
1159 	priv->config.flow_prio = err;
1160 	if (!priv->config.dv_esw_en &&
1161 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1162 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1163 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1164 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1165 	}
1166 	mlx5_set_metadata_mask(eth_dev);
1167 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1168 	    !priv->sh->dv_regc0_mask) {
1169 		DRV_LOG(ERR, "metadata mode %u is not supported "
1170 			     "(no metadata reg_c[0] is available)",
1171 			     priv->config.dv_xmeta_en);
1172 			err = ENOTSUP;
1173 			goto error;
1174 	}
1175 	/*
1176 	 * Allocate the buffer for flow creating, just once.
1177 	 * The allocation must be done before any flow creating.
1178 	 */
1179 	mlx5_flow_alloc_intermediate(eth_dev);
1180 	/* Query availability of metadata reg_c's. */
1181 	err = mlx5_flow_discover_mreg_c(eth_dev);
1182 	if (err < 0) {
1183 		err = -err;
1184 		goto error;
1185 	}
1186 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1187 		DRV_LOG(DEBUG,
1188 			"port %u extensive metadata register is not supported",
1189 			eth_dev->data->port_id);
1190 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1191 			DRV_LOG(ERR, "metadata mode %u is not supported "
1192 				     "(no metadata registers available)",
1193 				     priv->config.dv_xmeta_en);
1194 			err = ENOTSUP;
1195 			goto error;
1196 		}
1197 	}
1198 	if (priv->config.dv_flow_en &&
1199 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1200 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1201 	    priv->sh->dv_regc0_mask) {
1202 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1203 						      MLX5_FLOW_MREG_HTABLE_SZ);
1204 		if (!priv->mreg_cp_tbl) {
1205 			err = ENOMEM;
1206 			goto error;
1207 		}
1208 	}
1209 	return eth_dev;
1210 error:
1211 	if (priv) {
1212 		if (priv->mreg_cp_tbl)
1213 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1214 		if (priv->sh)
1215 			mlx5_os_free_shared_dr(priv);
1216 		if (priv->nl_socket_route >= 0)
1217 			close(priv->nl_socket_route);
1218 		if (priv->nl_socket_rdma >= 0)
1219 			close(priv->nl_socket_rdma);
1220 		if (priv->vmwa_context)
1221 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1222 		if (priv->qrss_id_pool)
1223 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1224 		if (own_domain_id)
1225 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1226 		mlx5_free(priv);
1227 		if (eth_dev != NULL)
1228 			eth_dev->data->dev_private = NULL;
1229 	}
1230 	if (eth_dev != NULL) {
1231 		/* mac_addrs must not be freed alone because part of
1232 		 * dev_private
1233 		 **/
1234 		eth_dev->data->mac_addrs = NULL;
1235 		rte_eth_dev_release_port(eth_dev);
1236 	}
1237 	if (sh)
1238 		mlx5_free_shared_dev_ctx(sh);
1239 	MLX5_ASSERT(err > 0);
1240 	rte_errno = err;
1241 	return NULL;
1242 }
1243 
1244 /**
1245  * Comparison callback to sort device data.
1246  *
1247  * This is meant to be used with qsort().
1248  *
1249  * @param a[in]
1250  *   Pointer to pointer to first data object.
1251  * @param b[in]
1252  *   Pointer to pointer to second data object.
1253  *
1254  * @return
1255  *   0 if both objects are equal, less than 0 if the first argument is less
1256  *   than the second, greater than 0 otherwise.
1257  */
1258 static int
1259 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1260 {
1261 	const struct mlx5_switch_info *si_a =
1262 		&((const struct mlx5_dev_spawn_data *)a)->info;
1263 	const struct mlx5_switch_info *si_b =
1264 		&((const struct mlx5_dev_spawn_data *)b)->info;
1265 	int ret;
1266 
1267 	/* Master device first. */
1268 	ret = si_b->master - si_a->master;
1269 	if (ret)
1270 		return ret;
1271 	/* Then representor devices. */
1272 	ret = si_b->representor - si_a->representor;
1273 	if (ret)
1274 		return ret;
1275 	/* Unidentified devices come last in no specific order. */
1276 	if (!si_a->representor)
1277 		return 0;
1278 	/* Order representors by name. */
1279 	return si_a->port_name - si_b->port_name;
1280 }
1281 
1282 /**
1283  * Match PCI information for possible slaves of bonding device.
1284  *
1285  * @param[in] ibv_dev
1286  *   Pointer to Infiniband device structure.
1287  * @param[in] pci_dev
1288  *   Pointer to PCI device structure to match PCI address.
1289  * @param[in] nl_rdma
1290  *   Netlink RDMA group socket handle.
1291  *
1292  * @return
1293  *   negative value if no bonding device found, otherwise
1294  *   positive index of slave PF in bonding.
1295  */
1296 static int
1297 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1298 			   const struct rte_pci_device *pci_dev,
1299 			   int nl_rdma)
1300 {
1301 	char ifname[IF_NAMESIZE + 1];
1302 	unsigned int ifindex;
1303 	unsigned int np, i;
1304 	FILE *file = NULL;
1305 	int pf = -1;
1306 
1307 	/*
1308 	 * Try to get master device name. If something goes
1309 	 * wrong suppose the lack of kernel support and no
1310 	 * bonding devices.
1311 	 */
1312 	if (nl_rdma < 0)
1313 		return -1;
1314 	if (!strstr(ibv_dev->name, "bond"))
1315 		return -1;
1316 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1317 	if (!np)
1318 		return -1;
1319 	/*
1320 	 * The Master device might not be on the predefined
1321 	 * port (not on port index 1, it is not garanted),
1322 	 * we have to scan all Infiniband device port and
1323 	 * find master.
1324 	 */
1325 	for (i = 1; i <= np; ++i) {
1326 		/* Check whether Infiniband port is populated. */
1327 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1328 		if (!ifindex)
1329 			continue;
1330 		if (!if_indextoname(ifindex, ifname))
1331 			continue;
1332 		/* Try to read bonding slave names from sysfs. */
1333 		MKSTR(slaves,
1334 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1335 		file = fopen(slaves, "r");
1336 		if (file)
1337 			break;
1338 	}
1339 	if (!file)
1340 		return -1;
1341 	/* Use safe format to check maximal buffer length. */
1342 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1343 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1344 		char tmp_str[IF_NAMESIZE + 32];
1345 		struct rte_pci_addr pci_addr;
1346 		struct mlx5_switch_info	info;
1347 
1348 		/* Process slave interface names in the loop. */
1349 		snprintf(tmp_str, sizeof(tmp_str),
1350 			 "/sys/class/net/%s", ifname);
1351 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1352 			DRV_LOG(WARNING, "can not get PCI address"
1353 					 " for netdev \"%s\"", ifname);
1354 			continue;
1355 		}
1356 		if (pci_dev->addr.domain != pci_addr.domain ||
1357 		    pci_dev->addr.bus != pci_addr.bus ||
1358 		    pci_dev->addr.devid != pci_addr.devid ||
1359 		    pci_dev->addr.function != pci_addr.function)
1360 			continue;
1361 		/* Slave interface PCI address match found. */
1362 		fclose(file);
1363 		snprintf(tmp_str, sizeof(tmp_str),
1364 			 "/sys/class/net/%s/phys_port_name", ifname);
1365 		file = fopen(tmp_str, "rb");
1366 		if (!file)
1367 			break;
1368 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1369 		if (fscanf(file, "%32s", tmp_str) == 1)
1370 			mlx5_translate_port_name(tmp_str, &info);
1371 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1372 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1373 			pf = info.port_name;
1374 		break;
1375 	}
1376 	if (file)
1377 		fclose(file);
1378 	return pf;
1379 }
1380 
1381 /**
1382  * DPDK callback to register a PCI device.
1383  *
1384  * This function spawns Ethernet devices out of a given PCI device.
1385  *
1386  * @param[in] pci_drv
1387  *   PCI driver structure (mlx5_driver).
1388  * @param[in] pci_dev
1389  *   PCI device information.
1390  *
1391  * @return
1392  *   0 on success, a negative errno value otherwise and rte_errno is set.
1393  */
1394 int
1395 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1396 		  struct rte_pci_device *pci_dev)
1397 {
1398 	struct ibv_device **ibv_list;
1399 	/*
1400 	 * Number of found IB Devices matching with requested PCI BDF.
1401 	 * nd != 1 means there are multiple IB devices over the same
1402 	 * PCI device and we have representors and master.
1403 	 */
1404 	unsigned int nd = 0;
1405 	/*
1406 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1407 	 * we have the single multiport IB device, and there may be
1408 	 * representors attached to some of found ports.
1409 	 */
1410 	unsigned int np = 0;
1411 	/*
1412 	 * Number of DPDK ethernet devices to Spawn - either over
1413 	 * multiple IB devices or multiple ports of single IB device.
1414 	 * Actually this is the number of iterations to spawn.
1415 	 */
1416 	unsigned int ns = 0;
1417 	/*
1418 	 * Bonding device
1419 	 *   < 0 - no bonding device (single one)
1420 	 *  >= 0 - bonding device (value is slave PF index)
1421 	 */
1422 	int bd = -1;
1423 	struct mlx5_dev_spawn_data *list = NULL;
1424 	struct mlx5_dev_config dev_config;
1425 	int ret;
1426 
1427 	if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1428 		DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1429 			" driver.");
1430 		return 1;
1431 	}
1432 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1433 		mlx5_pmd_socket_init();
1434 	ret = mlx5_init_once();
1435 	if (ret) {
1436 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1437 			strerror(rte_errno));
1438 		return -rte_errno;
1439 	}
1440 	MLX5_ASSERT(pci_drv == &mlx5_driver);
1441 	errno = 0;
1442 	ibv_list = mlx5_glue->get_device_list(&ret);
1443 	if (!ibv_list) {
1444 		rte_errno = errno ? errno : ENOSYS;
1445 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1446 		return -rte_errno;
1447 	}
1448 	/*
1449 	 * First scan the list of all Infiniband devices to find
1450 	 * matching ones, gathering into the list.
1451 	 */
1452 	struct ibv_device *ibv_match[ret + 1];
1453 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1454 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1455 	unsigned int i;
1456 
1457 	while (ret-- > 0) {
1458 		struct rte_pci_addr pci_addr;
1459 
1460 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1461 		bd = mlx5_device_bond_pci_match
1462 				(ibv_list[ret], pci_dev, nl_rdma);
1463 		if (bd >= 0) {
1464 			/*
1465 			 * Bonding device detected. Only one match is allowed,
1466 			 * the bonding is supported over multi-port IB device,
1467 			 * there should be no matches on representor PCI
1468 			 * functions or non VF LAG bonding devices with
1469 			 * specified address.
1470 			 */
1471 			if (nd) {
1472 				DRV_LOG(ERR,
1473 					"multiple PCI match on bonding device"
1474 					"\"%s\" found", ibv_list[ret]->name);
1475 				rte_errno = ENOENT;
1476 				ret = -rte_errno;
1477 				goto exit;
1478 			}
1479 			DRV_LOG(INFO, "PCI information matches for"
1480 				      " slave %d bonding device \"%s\"",
1481 				      bd, ibv_list[ret]->name);
1482 			ibv_match[nd++] = ibv_list[ret];
1483 			break;
1484 		}
1485 		if (mlx5_dev_to_pci_addr
1486 			(ibv_list[ret]->ibdev_path, &pci_addr))
1487 			continue;
1488 		if (pci_dev->addr.domain != pci_addr.domain ||
1489 		    pci_dev->addr.bus != pci_addr.bus ||
1490 		    pci_dev->addr.devid != pci_addr.devid ||
1491 		    pci_dev->addr.function != pci_addr.function)
1492 			continue;
1493 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1494 			ibv_list[ret]->name);
1495 		ibv_match[nd++] = ibv_list[ret];
1496 	}
1497 	ibv_match[nd] = NULL;
1498 	if (!nd) {
1499 		/* No device matches, just complain and bail out. */
1500 		DRV_LOG(WARNING,
1501 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1502 			" are kernel drivers loaded?",
1503 			pci_dev->addr.domain, pci_dev->addr.bus,
1504 			pci_dev->addr.devid, pci_dev->addr.function);
1505 		rte_errno = ENOENT;
1506 		ret = -rte_errno;
1507 		goto exit;
1508 	}
1509 	if (nd == 1) {
1510 		/*
1511 		 * Found single matching device may have multiple ports.
1512 		 * Each port may be representor, we have to check the port
1513 		 * number and check the representors existence.
1514 		 */
1515 		if (nl_rdma >= 0)
1516 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1517 		if (!np)
1518 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1519 					 " ports number", ibv_match[0]->name);
1520 		if (bd >= 0 && !np) {
1521 			DRV_LOG(ERR, "can not get ports"
1522 				     " for bonding device");
1523 			rte_errno = ENOENT;
1524 			ret = -rte_errno;
1525 			goto exit;
1526 		}
1527 	}
1528 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1529 	if (bd >= 0) {
1530 		/*
1531 		 * This may happen if there is VF LAG kernel support and
1532 		 * application is compiled with older rdma_core library.
1533 		 */
1534 		DRV_LOG(ERR,
1535 			"No kernel/verbs support for VF LAG bonding found.");
1536 		rte_errno = ENOTSUP;
1537 		ret = -rte_errno;
1538 		goto exit;
1539 	}
1540 #endif
1541 	/*
1542 	 * Now we can determine the maximal
1543 	 * amount of devices to be spawned.
1544 	 */
1545 	list = mlx5_malloc(MLX5_MEM_ZERO,
1546 			   sizeof(struct mlx5_dev_spawn_data) *
1547 			   (np ? np : nd),
1548 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1549 	if (!list) {
1550 		DRV_LOG(ERR, "spawn data array allocation failure");
1551 		rte_errno = ENOMEM;
1552 		ret = -rte_errno;
1553 		goto exit;
1554 	}
1555 	if (bd >= 0 || np > 1) {
1556 		/*
1557 		 * Single IB device with multiple ports found,
1558 		 * it may be E-Switch master device and representors.
1559 		 * We have to perform identification through the ports.
1560 		 */
1561 		MLX5_ASSERT(nl_rdma >= 0);
1562 		MLX5_ASSERT(ns == 0);
1563 		MLX5_ASSERT(nd == 1);
1564 		MLX5_ASSERT(np);
1565 		for (i = 1; i <= np; ++i) {
1566 			list[ns].max_port = np;
1567 			list[ns].phys_port = i;
1568 			list[ns].phys_dev = ibv_match[0];
1569 			list[ns].eth_dev = NULL;
1570 			list[ns].pci_dev = pci_dev;
1571 			list[ns].pf_bond = bd;
1572 			list[ns].ifindex = mlx5_nl_ifindex
1573 				(nl_rdma,
1574 				mlx5_os_get_dev_device_name
1575 						(list[ns].phys_dev), i);
1576 			if (!list[ns].ifindex) {
1577 				/*
1578 				 * No network interface index found for the
1579 				 * specified port, it means there is no
1580 				 * representor on this port. It's OK,
1581 				 * there can be disabled ports, for example
1582 				 * if sriov_numvfs < sriov_totalvfs.
1583 				 */
1584 				continue;
1585 			}
1586 			ret = -1;
1587 			if (nl_route >= 0)
1588 				ret = mlx5_nl_switch_info
1589 					       (nl_route,
1590 						list[ns].ifindex,
1591 						&list[ns].info);
1592 			if (ret || (!list[ns].info.representor &&
1593 				    !list[ns].info.master)) {
1594 				/*
1595 				 * We failed to recognize representors with
1596 				 * Netlink, let's try to perform the task
1597 				 * with sysfs.
1598 				 */
1599 				ret =  mlx5_sysfs_switch_info
1600 						(list[ns].ifindex,
1601 						 &list[ns].info);
1602 			}
1603 			if (!ret && bd >= 0) {
1604 				switch (list[ns].info.name_type) {
1605 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1606 					if (list[ns].info.port_name == bd)
1607 						ns++;
1608 					break;
1609 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1610 					/* Fallthrough */
1611 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1612 					if (list[ns].info.pf_num == bd)
1613 						ns++;
1614 					break;
1615 				default:
1616 					break;
1617 				}
1618 				continue;
1619 			}
1620 			if (!ret && (list[ns].info.representor ^
1621 				     list[ns].info.master))
1622 				ns++;
1623 		}
1624 		if (!ns) {
1625 			DRV_LOG(ERR,
1626 				"unable to recognize master/representors"
1627 				" on the IB device with multiple ports");
1628 			rte_errno = ENOENT;
1629 			ret = -rte_errno;
1630 			goto exit;
1631 		}
1632 	} else {
1633 		/*
1634 		 * The existence of several matching entries (nd > 1) means
1635 		 * port representors have been instantiated. No existing Verbs
1636 		 * call nor sysfs entries can tell them apart, this can only
1637 		 * be done through Netlink calls assuming kernel drivers are
1638 		 * recent enough to support them.
1639 		 *
1640 		 * In the event of identification failure through Netlink,
1641 		 * try again through sysfs, then:
1642 		 *
1643 		 * 1. A single IB device matches (nd == 1) with single
1644 		 *    port (np=0/1) and is not a representor, assume
1645 		 *    no switch support.
1646 		 *
1647 		 * 2. Otherwise no safe assumptions can be made;
1648 		 *    complain louder and bail out.
1649 		 */
1650 		for (i = 0; i != nd; ++i) {
1651 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1652 			list[ns].max_port = 1;
1653 			list[ns].phys_port = 1;
1654 			list[ns].phys_dev = ibv_match[i];
1655 			list[ns].eth_dev = NULL;
1656 			list[ns].pci_dev = pci_dev;
1657 			list[ns].pf_bond = -1;
1658 			list[ns].ifindex = 0;
1659 			if (nl_rdma >= 0)
1660 				list[ns].ifindex = mlx5_nl_ifindex
1661 				(nl_rdma,
1662 				mlx5_os_get_dev_device_name
1663 						(list[ns].phys_dev), 1);
1664 			if (!list[ns].ifindex) {
1665 				char ifname[IF_NAMESIZE];
1666 
1667 				/*
1668 				 * Netlink failed, it may happen with old
1669 				 * ib_core kernel driver (before 4.16).
1670 				 * We can assume there is old driver because
1671 				 * here we are processing single ports IB
1672 				 * devices. Let's try sysfs to retrieve
1673 				 * the ifindex. The method works for
1674 				 * master device only.
1675 				 */
1676 				if (nd > 1) {
1677 					/*
1678 					 * Multiple devices found, assume
1679 					 * representors, can not distinguish
1680 					 * master/representor and retrieve
1681 					 * ifindex via sysfs.
1682 					 */
1683 					continue;
1684 				}
1685 				ret = mlx5_get_ifname_sysfs
1686 					(ibv_match[i]->ibdev_path, ifname);
1687 				if (!ret)
1688 					list[ns].ifindex =
1689 						if_nametoindex(ifname);
1690 				if (!list[ns].ifindex) {
1691 					/*
1692 					 * No network interface index found
1693 					 * for the specified device, it means
1694 					 * there it is neither representor
1695 					 * nor master.
1696 					 */
1697 					continue;
1698 				}
1699 			}
1700 			ret = -1;
1701 			if (nl_route >= 0)
1702 				ret = mlx5_nl_switch_info
1703 					       (nl_route,
1704 						list[ns].ifindex,
1705 						&list[ns].info);
1706 			if (ret || (!list[ns].info.representor &&
1707 				    !list[ns].info.master)) {
1708 				/*
1709 				 * We failed to recognize representors with
1710 				 * Netlink, let's try to perform the task
1711 				 * with sysfs.
1712 				 */
1713 				ret =  mlx5_sysfs_switch_info
1714 						(list[ns].ifindex,
1715 						 &list[ns].info);
1716 			}
1717 			if (!ret && (list[ns].info.representor ^
1718 				     list[ns].info.master)) {
1719 				ns++;
1720 			} else if ((nd == 1) &&
1721 				   !list[ns].info.representor &&
1722 				   !list[ns].info.master) {
1723 				/*
1724 				 * Single IB device with
1725 				 * one physical port and
1726 				 * attached network device.
1727 				 * May be SRIOV is not enabled
1728 				 * or there is no representors.
1729 				 */
1730 				DRV_LOG(INFO, "no E-Switch support detected");
1731 				ns++;
1732 				break;
1733 			}
1734 		}
1735 		if (!ns) {
1736 			DRV_LOG(ERR,
1737 				"unable to recognize master/representors"
1738 				" on the multiple IB devices");
1739 			rte_errno = ENOENT;
1740 			ret = -rte_errno;
1741 			goto exit;
1742 		}
1743 	}
1744 	MLX5_ASSERT(ns);
1745 	/*
1746 	 * Sort list to probe devices in natural order for users convenience
1747 	 * (i.e. master first, then representors from lowest to highest ID).
1748 	 */
1749 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1750 	/* Default configuration. */
1751 	dev_config = (struct mlx5_dev_config){
1752 		.hw_padding = 0,
1753 		.mps = MLX5_ARG_UNSET,
1754 		.dbnc = MLX5_ARG_UNSET,
1755 		.rx_vec_en = 1,
1756 		.txq_inline_max = MLX5_ARG_UNSET,
1757 		.txq_inline_min = MLX5_ARG_UNSET,
1758 		.txq_inline_mpw = MLX5_ARG_UNSET,
1759 		.txqs_inline = MLX5_ARG_UNSET,
1760 		.vf_nl_en = 1,
1761 		.mr_ext_memseg_en = 1,
1762 		.mprq = {
1763 			.enabled = 0, /* Disabled by default. */
1764 			.stride_num_n = 0,
1765 			.stride_size_n = 0,
1766 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1767 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1768 		},
1769 		.dv_esw_en = 1,
1770 		.dv_flow_en = 1,
1771 		.decap_en = 1,
1772 		.log_hp_size = MLX5_ARG_UNSET,
1773 	};
1774 	/* Device specific configuration. */
1775 	switch (pci_dev->id.device_id) {
1776 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1777 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1778 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1779 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1780 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1781 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1782 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1783 		dev_config.vf = 1;
1784 		break;
1785 	default:
1786 		break;
1787 	}
1788 	for (i = 0; i != ns; ++i) {
1789 		uint32_t restore;
1790 
1791 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1792 						 &list[i],
1793 						 dev_config);
1794 		if (!list[i].eth_dev) {
1795 			if (rte_errno != EBUSY && rte_errno != EEXIST)
1796 				break;
1797 			/* Device is disabled or already spawned. Ignore it. */
1798 			continue;
1799 		}
1800 		restore = list[i].eth_dev->data->dev_flags;
1801 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1802 		/* Restore non-PCI flags cleared by the above call. */
1803 		list[i].eth_dev->data->dev_flags |= restore;
1804 		rte_eth_dev_probing_finish(list[i].eth_dev);
1805 	}
1806 	if (i != ns) {
1807 		DRV_LOG(ERR,
1808 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1809 			" encountering an error: %s",
1810 			pci_dev->addr.domain, pci_dev->addr.bus,
1811 			pci_dev->addr.devid, pci_dev->addr.function,
1812 			strerror(rte_errno));
1813 		ret = -rte_errno;
1814 		/* Roll back. */
1815 		while (i--) {
1816 			if (!list[i].eth_dev)
1817 				continue;
1818 			mlx5_dev_close(list[i].eth_dev);
1819 			/* mac_addrs must not be freed because in dev_private */
1820 			list[i].eth_dev->data->mac_addrs = NULL;
1821 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1822 		}
1823 		/* Restore original error. */
1824 		rte_errno = -ret;
1825 	} else {
1826 		ret = 0;
1827 	}
1828 exit:
1829 	/*
1830 	 * Do the routine cleanup:
1831 	 * - close opened Netlink sockets
1832 	 * - free allocated spawn data array
1833 	 * - free the Infiniband device list
1834 	 */
1835 	if (nl_rdma >= 0)
1836 		close(nl_rdma);
1837 	if (nl_route >= 0)
1838 		close(nl_route);
1839 	if (list)
1840 		mlx5_free(list);
1841 	MLX5_ASSERT(ibv_list);
1842 	mlx5_glue->free_device_list(ibv_list);
1843 	return ret;
1844 }
1845 
1846 static int
1847 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1848 {
1849 	char *env;
1850 	int value;
1851 
1852 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1853 	/* Get environment variable to store. */
1854 	env = getenv(MLX5_SHUT_UP_BF);
1855 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1856 	if (config->dbnc == MLX5_ARG_UNSET)
1857 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1858 	else
1859 		setenv(MLX5_SHUT_UP_BF,
1860 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1861 	return value;
1862 }
1863 
1864 static void
1865 mlx5_restore_doorbell_mapping_env(int value)
1866 {
1867 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1868 	/* Restore the original environment variable state. */
1869 	if (value == MLX5_ARG_UNSET)
1870 		unsetenv(MLX5_SHUT_UP_BF);
1871 	else
1872 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1873 }
1874 
1875 /**
1876  * Extract pdn of PD object using DV API.
1877  *
1878  * @param[in] pd
1879  *   Pointer to the verbs PD object.
1880  * @param[out] pdn
1881  *   Pointer to the PD object number variable.
1882  *
1883  * @return
1884  *   0 on success, error value otherwise.
1885  */
1886 int
1887 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1888 {
1889 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1890 	struct mlx5dv_obj obj;
1891 	struct mlx5dv_pd pd_info;
1892 	int ret = 0;
1893 
1894 	obj.pd.in = pd;
1895 	obj.pd.out = &pd_info;
1896 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1897 	if (ret) {
1898 		DRV_LOG(DEBUG, "Fail to get PD object info");
1899 		return ret;
1900 	}
1901 	*pdn = pd_info.pdn;
1902 	return 0;
1903 #else
1904 	(void)pd;
1905 	(void)pdn;
1906 	return -ENOTSUP;
1907 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1908 }
1909 
1910 /**
1911  * Function API to open IB device.
1912  *
1913  * This function calls the Linux glue APIs to open a device.
1914  *
1915  * @param[in] spawn
1916  *   Pointer to the IB device attributes (name, port, etc).
1917  * @param[out] config
1918  *   Pointer to device configuration structure.
1919  * @param[out] sh
1920  *   Pointer to shared context structure.
1921  *
1922  * @return
1923  *   0 on success, a positive error value otherwise.
1924  */
1925 int
1926 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1927 		     const struct mlx5_dev_config *config,
1928 		     struct mlx5_dev_ctx_shared *sh)
1929 {
1930 	int dbmap_env;
1931 	int err = 0;
1932 
1933 	sh->numa_node = spawn->pci_dev->device.numa_node;
1934 	pthread_mutex_init(&sh->txpp.mutex, NULL);
1935 	/*
1936 	 * Configure environment variable "MLX5_BF_SHUT_UP"
1937 	 * before the device creation. The rdma_core library
1938 	 * checks the variable at device creation and
1939 	 * stores the result internally.
1940 	 */
1941 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
1942 	/* Try to open IB device with DV first, then usual Verbs. */
1943 	errno = 0;
1944 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1945 	if (sh->ctx) {
1946 		sh->devx = 1;
1947 		DRV_LOG(DEBUG, "DevX is supported");
1948 		/* The device is created, no need for environment. */
1949 		mlx5_restore_doorbell_mapping_env(dbmap_env);
1950 	} else {
1951 		/* The environment variable is still configured. */
1952 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1953 		err = errno ? errno : ENODEV;
1954 		/*
1955 		 * The environment variable is not needed anymore,
1956 		 * all device creation attempts are completed.
1957 		 */
1958 		mlx5_restore_doorbell_mapping_env(dbmap_env);
1959 		if (!sh->ctx)
1960 			return err;
1961 		DRV_LOG(DEBUG, "DevX is NOT supported");
1962 		err = 0;
1963 	}
1964 	return err;
1965 }
1966 
1967 /**
1968  * Install shared asynchronous device events handler.
1969  * This function is implemented to support event sharing
1970  * between multiple ports of single IB device.
1971  *
1972  * @param sh
1973  *   Pointer to mlx5_dev_ctx_shared object.
1974  */
1975 void
1976 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1977 {
1978 	int ret;
1979 	int flags;
1980 
1981 	sh->intr_handle.fd = -1;
1982 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1983 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1984 		    F_SETFL, flags | O_NONBLOCK);
1985 	if (ret) {
1986 		DRV_LOG(INFO, "failed to change file descriptor async event"
1987 			" queue");
1988 	} else {
1989 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1990 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1991 		if (rte_intr_callback_register(&sh->intr_handle,
1992 					mlx5_dev_interrupt_handler, sh)) {
1993 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
1994 			sh->intr_handle.fd = -1;
1995 		}
1996 	}
1997 	if (sh->devx) {
1998 #ifdef HAVE_IBV_DEVX_ASYNC
1999 		sh->intr_handle_devx.fd = -1;
2000 		sh->devx_comp =
2001 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2002 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2003 		if (!devx_comp) {
2004 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2005 			return;
2006 		}
2007 		flags = fcntl(devx_comp->fd, F_GETFL);
2008 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2009 		if (ret) {
2010 			DRV_LOG(INFO, "failed to change file descriptor"
2011 				" devx comp");
2012 			return;
2013 		}
2014 		sh->intr_handle_devx.fd = devx_comp->fd;
2015 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2016 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2017 					mlx5_dev_interrupt_handler_devx, sh)) {
2018 			DRV_LOG(INFO, "Fail to install the devx shared"
2019 				" interrupt.");
2020 			sh->intr_handle_devx.fd = -1;
2021 		}
2022 #endif /* HAVE_IBV_DEVX_ASYNC */
2023 	}
2024 }
2025 
2026 /**
2027  * Uninstall shared asynchronous device events handler.
2028  * This function is implemented to support event sharing
2029  * between multiple ports of single IB device.
2030  *
2031  * @param dev
2032  *   Pointer to mlx5_dev_ctx_shared object.
2033  */
2034 void
2035 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2036 {
2037 	if (sh->intr_handle.fd >= 0)
2038 		mlx5_intr_callback_unregister(&sh->intr_handle,
2039 					      mlx5_dev_interrupt_handler, sh);
2040 #ifdef HAVE_IBV_DEVX_ASYNC
2041 	if (sh->intr_handle_devx.fd >= 0)
2042 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2043 				  mlx5_dev_interrupt_handler_devx, sh);
2044 	if (sh->devx_comp)
2045 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2046 #endif
2047 }
2048 
2049 /**
2050  * Read statistics by a named counter.
2051  *
2052  * @param[in] priv
2053  *   Pointer to the private device data structure.
2054  * @param[in] ctr_name
2055  *   Pointer to the name of the statistic counter to read
2056  * @param[out] stat
2057  *   Pointer to read statistic value.
2058  * @return
2059  *   0 on success and stat is valud, 1 if failed to read the value
2060  *   rte_errno is set.
2061  *
2062  */
2063 int
2064 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2065 		      uint64_t *stat)
2066 {
2067 	int fd;
2068 
2069 	if (priv->sh) {
2070 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2071 			  priv->sh->ibdev_path,
2072 			  priv->dev_port,
2073 			  ctr_name);
2074 		fd = open(path, O_RDONLY);
2075 		if (fd != -1) {
2076 			char buf[21] = {'\0'};
2077 			ssize_t n = read(fd, buf, sizeof(buf));
2078 
2079 			close(fd);
2080 			if (n != -1) {
2081 				*stat = strtoull(buf, NULL, 10);
2082 				return 0;
2083 			}
2084 		}
2085 	}
2086 	*stat = 0;
2087 	return 1;
2088 }
2089 
2090 /**
2091  * Read device counters table.
2092  *
2093  * @param dev
2094  *   Pointer to Ethernet device.
2095  * @param[out] stats
2096  *   Counters table output buffer.
2097  *
2098  * @return
2099  *   0 on success and stats is filled, negative errno value otherwise and
2100  *   rte_errno is set.
2101  */
2102 int
2103 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
2104 {
2105 	struct mlx5_priv *priv = dev->data->dev_private;
2106 	struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2107 	unsigned int i;
2108 	struct ifreq ifr;
2109 	unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
2110 	unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
2111 	struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
2112 	int ret;
2113 
2114 	et_stats->cmd = ETHTOOL_GSTATS;
2115 	et_stats->n_stats = xstats_ctrl->stats_n;
2116 	ifr.ifr_data = (caddr_t)et_stats;
2117 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2118 	if (ret) {
2119 		DRV_LOG(WARNING,
2120 			"port %u unable to read statistic values from device",
2121 			dev->data->port_id);
2122 		return ret;
2123 	}
2124 	for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2125 		if (xstats_ctrl->info[i].dev) {
2126 			ret = mlx5_os_read_dev_stat(priv,
2127 					    xstats_ctrl->info[i].ctr_name,
2128 					    &stats[i]);
2129 			/* return last xstats counter if fail to read. */
2130 			if (ret == 0)
2131 				xstats_ctrl->xstats[i] = stats[i];
2132 			else
2133 				stats[i] = xstats_ctrl->xstats[i];
2134 		} else {
2135 			stats[i] = (uint64_t)
2136 				et_stats->data[xstats_ctrl->dev_table_idx[i]];
2137 		}
2138 	}
2139 	return 0;
2140 }
2141 
2142 /**
2143  * Query the number of statistics provided by ETHTOOL.
2144  *
2145  * @param dev
2146  *   Pointer to Ethernet device.
2147  *
2148  * @return
2149  *   Number of statistics on success, negative errno value otherwise and
2150  *   rte_errno is set.
2151  */
2152 int
2153 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2154 {
2155 	struct ethtool_drvinfo drvinfo;
2156 	struct ifreq ifr;
2157 	int ret;
2158 
2159 	drvinfo.cmd = ETHTOOL_GDRVINFO;
2160 	ifr.ifr_data = (caddr_t)&drvinfo;
2161 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2162 	if (ret) {
2163 		DRV_LOG(WARNING, "port %u unable to query number of statistics",
2164 			dev->data->port_id);
2165 		return ret;
2166 	}
2167 	return drvinfo.n_stats;
2168 }
2169 
2170 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2171 	{
2172 		.dpdk_name = "rx_port_unicast_bytes",
2173 		.ctr_name = "rx_vport_unicast_bytes",
2174 	},
2175 	{
2176 		.dpdk_name = "rx_port_multicast_bytes",
2177 		.ctr_name = "rx_vport_multicast_bytes",
2178 	},
2179 	{
2180 		.dpdk_name = "rx_port_broadcast_bytes",
2181 		.ctr_name = "rx_vport_broadcast_bytes",
2182 	},
2183 	{
2184 		.dpdk_name = "rx_port_unicast_packets",
2185 		.ctr_name = "rx_vport_unicast_packets",
2186 	},
2187 	{
2188 		.dpdk_name = "rx_port_multicast_packets",
2189 		.ctr_name = "rx_vport_multicast_packets",
2190 	},
2191 	{
2192 		.dpdk_name = "rx_port_broadcast_packets",
2193 		.ctr_name = "rx_vport_broadcast_packets",
2194 	},
2195 	{
2196 		.dpdk_name = "tx_port_unicast_bytes",
2197 		.ctr_name = "tx_vport_unicast_bytes",
2198 	},
2199 	{
2200 		.dpdk_name = "tx_port_multicast_bytes",
2201 		.ctr_name = "tx_vport_multicast_bytes",
2202 	},
2203 	{
2204 		.dpdk_name = "tx_port_broadcast_bytes",
2205 		.ctr_name = "tx_vport_broadcast_bytes",
2206 	},
2207 	{
2208 		.dpdk_name = "tx_port_unicast_packets",
2209 		.ctr_name = "tx_vport_unicast_packets",
2210 	},
2211 	{
2212 		.dpdk_name = "tx_port_multicast_packets",
2213 		.ctr_name = "tx_vport_multicast_packets",
2214 	},
2215 	{
2216 		.dpdk_name = "tx_port_broadcast_packets",
2217 		.ctr_name = "tx_vport_broadcast_packets",
2218 	},
2219 	{
2220 		.dpdk_name = "rx_wqe_err",
2221 		.ctr_name = "rx_wqe_err",
2222 	},
2223 	{
2224 		.dpdk_name = "rx_crc_errors_phy",
2225 		.ctr_name = "rx_crc_errors_phy",
2226 	},
2227 	{
2228 		.dpdk_name = "rx_in_range_len_errors_phy",
2229 		.ctr_name = "rx_in_range_len_errors_phy",
2230 	},
2231 	{
2232 		.dpdk_name = "rx_symbol_err_phy",
2233 		.ctr_name = "rx_symbol_err_phy",
2234 	},
2235 	{
2236 		.dpdk_name = "tx_errors_phy",
2237 		.ctr_name = "tx_errors_phy",
2238 	},
2239 	{
2240 		.dpdk_name = "rx_out_of_buffer",
2241 		.ctr_name = "out_of_buffer",
2242 		.dev = 1,
2243 	},
2244 	{
2245 		.dpdk_name = "tx_packets_phy",
2246 		.ctr_name = "tx_packets_phy",
2247 	},
2248 	{
2249 		.dpdk_name = "rx_packets_phy",
2250 		.ctr_name = "rx_packets_phy",
2251 	},
2252 	{
2253 		.dpdk_name = "tx_discards_phy",
2254 		.ctr_name = "tx_discards_phy",
2255 	},
2256 	{
2257 		.dpdk_name = "rx_discards_phy",
2258 		.ctr_name = "rx_discards_phy",
2259 	},
2260 	{
2261 		.dpdk_name = "tx_bytes_phy",
2262 		.ctr_name = "tx_bytes_phy",
2263 	},
2264 	{
2265 		.dpdk_name = "rx_bytes_phy",
2266 		.ctr_name = "rx_bytes_phy",
2267 	},
2268 	/* Representor only */
2269 	{
2270 		.dpdk_name = "rx_packets",
2271 		.ctr_name = "vport_rx_packets",
2272 	},
2273 	{
2274 		.dpdk_name = "rx_bytes",
2275 		.ctr_name = "vport_rx_bytes",
2276 	},
2277 	{
2278 		.dpdk_name = "tx_packets",
2279 		.ctr_name = "vport_tx_packets",
2280 	},
2281 	{
2282 		.dpdk_name = "tx_bytes",
2283 		.ctr_name = "vport_tx_bytes",
2284 	},
2285 };
2286 
2287 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2288 
2289 /**
2290  * Init the structures to read device counters.
2291  *
2292  * @param dev
2293  *   Pointer to Ethernet device.
2294  */
2295 void
2296 mlx5_os_stats_init(struct rte_eth_dev *dev)
2297 {
2298 	struct mlx5_priv *priv = dev->data->dev_private;
2299 	struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2300 	struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2301 	unsigned int i;
2302 	unsigned int j;
2303 	struct ifreq ifr;
2304 	struct ethtool_gstrings *strings = NULL;
2305 	unsigned int dev_stats_n;
2306 	unsigned int str_sz;
2307 	int ret;
2308 
2309 	/* So that it won't aggregate for each init. */
2310 	xstats_ctrl->mlx5_stats_n = 0;
2311 	ret = mlx5_os_get_stats_n(dev);
2312 	if (ret < 0) {
2313 		DRV_LOG(WARNING, "port %u no extended statistics available",
2314 			dev->data->port_id);
2315 		return;
2316 	}
2317 	dev_stats_n = ret;
2318 	/* Allocate memory to grab stat names and values. */
2319 	str_sz = dev_stats_n * ETH_GSTRING_LEN;
2320 	strings = (struct ethtool_gstrings *)
2321 		  mlx5_malloc(0, str_sz + sizeof(struct ethtool_gstrings), 0,
2322 			      SOCKET_ID_ANY);
2323 	if (!strings) {
2324 		DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2325 		     dev->data->port_id);
2326 		return;
2327 	}
2328 	strings->cmd = ETHTOOL_GSTRINGS;
2329 	strings->string_set = ETH_SS_STATS;
2330 	strings->len = dev_stats_n;
2331 	ifr.ifr_data = (caddr_t)strings;
2332 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2333 	if (ret) {
2334 		DRV_LOG(WARNING, "port %u unable to get statistic names",
2335 			dev->data->port_id);
2336 		goto free;
2337 	}
2338 	for (i = 0; i != dev_stats_n; ++i) {
2339 		const char *curr_string = (const char *)
2340 			&strings->data[i * ETH_GSTRING_LEN];
2341 
2342 		for (j = 0; j != xstats_n; ++j) {
2343 			if (!strcmp(mlx5_counters_init[j].ctr_name,
2344 				    curr_string)) {
2345 				unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2346 
2347 				xstats_ctrl->dev_table_idx[idx] = i;
2348 				xstats_ctrl->info[idx] = mlx5_counters_init[j];
2349 				break;
2350 			}
2351 		}
2352 	}
2353 	/* Add dev counters. */
2354 	for (i = 0; i != xstats_n; ++i) {
2355 		if (mlx5_counters_init[i].dev) {
2356 			unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2357 
2358 			xstats_ctrl->info[idx] = mlx5_counters_init[i];
2359 			xstats_ctrl->hw_stats[idx] = 0;
2360 		}
2361 	}
2362 	MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2363 	xstats_ctrl->stats_n = dev_stats_n;
2364 	/* Copy to base at first time. */
2365 	ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2366 	if (ret)
2367 		DRV_LOG(ERR, "port %u cannot read device counters: %s",
2368 			dev->data->port_id, strerror(rte_errno));
2369 	mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2370 	stats_ctrl->imissed = 0;
2371 free:
2372 	mlx5_free(strings);
2373 }
2374 
2375 /**
2376  * Set the reg_mr and dereg_mr call backs
2377  *
2378  * @param reg_mr_cb[out]
2379  *   Pointer to reg_mr func
2380  * @param dereg_mr_cb[out]
2381  *   Pointer to dereg_mr func
2382  *
2383  */
2384 void
2385 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2386 		      mlx5_dereg_mr_t *dereg_mr_cb)
2387 {
2388 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2389 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2390 }
2391 
2392 /**
2393  * Remove a MAC address from device
2394  *
2395  * @param dev
2396  *   Pointer to Ethernet device structure.
2397  * @param index
2398  *   MAC address index.
2399  */
2400 void
2401 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2402 {
2403 	struct mlx5_priv *priv = dev->data->dev_private;
2404 	const int vf = priv->config.vf;
2405 
2406 	if (vf)
2407 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2408 					mlx5_ifindex(dev), priv->mac_own,
2409 					&dev->data->mac_addrs[index], index);
2410 }
2411 
2412 /**
2413  * Adds a MAC address to the device
2414  *
2415  * @param dev
2416  *   Pointer to Ethernet device structure.
2417  * @param mac_addr
2418  *   MAC address to register.
2419  * @param index
2420  *   MAC address index.
2421  *
2422  * @return
2423  *   0 on success, a negative errno value otherwise
2424  */
2425 int
2426 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2427 		     uint32_t index)
2428 {
2429 	struct mlx5_priv *priv = dev->data->dev_private;
2430 	const int vf = priv->config.vf;
2431 	int ret = 0;
2432 
2433 	if (vf)
2434 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2435 					   mlx5_ifindex(dev), priv->mac_own,
2436 					   mac, index);
2437 	return ret;
2438 }
2439 
2440 /**
2441  * Modify a VF MAC address
2442  *
2443  * @param priv
2444  *   Pointer to device private data.
2445  * @param mac_addr
2446  *   MAC address to modify into.
2447  * @param iface_idx
2448  *   Net device interface index
2449  * @param vf_index
2450  *   VF index
2451  *
2452  * @return
2453  *   0 on success, a negative errno value otherwise
2454  */
2455 int
2456 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2457 			   unsigned int iface_idx,
2458 			   struct rte_ether_addr *mac_addr,
2459 			   int vf_index)
2460 {
2461 	return mlx5_nl_vf_mac_addr_modify
2462 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2463 }
2464 
2465 const struct eth_dev_ops mlx5_os_dev_ops = {
2466 	.dev_configure = mlx5_dev_configure,
2467 	.dev_start = mlx5_dev_start,
2468 	.dev_stop = mlx5_dev_stop,
2469 	.dev_set_link_down = mlx5_set_link_down,
2470 	.dev_set_link_up = mlx5_set_link_up,
2471 	.dev_close = mlx5_dev_close,
2472 	.promiscuous_enable = mlx5_promiscuous_enable,
2473 	.promiscuous_disable = mlx5_promiscuous_disable,
2474 	.allmulticast_enable = mlx5_allmulticast_enable,
2475 	.allmulticast_disable = mlx5_allmulticast_disable,
2476 	.link_update = mlx5_link_update,
2477 	.stats_get = mlx5_stats_get,
2478 	.stats_reset = mlx5_stats_reset,
2479 	.xstats_get = mlx5_xstats_get,
2480 	.xstats_reset = mlx5_xstats_reset,
2481 	.xstats_get_names = mlx5_xstats_get_names,
2482 	.fw_version_get = mlx5_fw_version_get,
2483 	.dev_infos_get = mlx5_dev_infos_get,
2484 	.read_clock = mlx5_txpp_read_clock,
2485 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2486 	.vlan_filter_set = mlx5_vlan_filter_set,
2487 	.rx_queue_setup = mlx5_rx_queue_setup,
2488 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2489 	.tx_queue_setup = mlx5_tx_queue_setup,
2490 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2491 	.rx_queue_release = mlx5_rx_queue_release,
2492 	.tx_queue_release = mlx5_tx_queue_release,
2493 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2494 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2495 	.mac_addr_remove = mlx5_mac_addr_remove,
2496 	.mac_addr_add = mlx5_mac_addr_add,
2497 	.mac_addr_set = mlx5_mac_addr_set,
2498 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2499 	.mtu_set = mlx5_dev_set_mtu,
2500 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2501 	.vlan_offload_set = mlx5_vlan_offload_set,
2502 	.reta_update = mlx5_dev_rss_reta_update,
2503 	.reta_query = mlx5_dev_rss_reta_query,
2504 	.rss_hash_update = mlx5_rss_hash_update,
2505 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2506 	.filter_ctrl = mlx5_dev_filter_ctrl,
2507 	.rx_descriptor_status = mlx5_rx_descriptor_status,
2508 	.tx_descriptor_status = mlx5_tx_descriptor_status,
2509 	.rxq_info_get = mlx5_rxq_info_get,
2510 	.txq_info_get = mlx5_txq_info_get,
2511 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2512 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2513 	.rx_queue_count = mlx5_rx_queue_count,
2514 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2515 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2516 	.is_removed = mlx5_is_removed,
2517 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2518 	.get_module_info = mlx5_get_module_info,
2519 	.get_module_eeprom = mlx5_get_module_eeprom,
2520 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2521 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2522 };
2523 
2524 /* Available operations from secondary process. */
2525 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2526 	.stats_get = mlx5_stats_get,
2527 	.stats_reset = mlx5_stats_reset,
2528 	.xstats_get = mlx5_xstats_get,
2529 	.xstats_reset = mlx5_xstats_reset,
2530 	.xstats_get_names = mlx5_xstats_get_names,
2531 	.fw_version_get = mlx5_fw_version_get,
2532 	.dev_infos_get = mlx5_dev_infos_get,
2533 	.read_clock = mlx5_txpp_read_clock,
2534 	.rx_descriptor_status = mlx5_rx_descriptor_status,
2535 	.tx_descriptor_status = mlx5_tx_descriptor_status,
2536 	.rxq_info_get = mlx5_rxq_info_get,
2537 	.txq_info_get = mlx5_txq_info_get,
2538 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2539 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2540 	.get_module_info = mlx5_get_module_info,
2541 	.get_module_eeprom = mlx5_get_module_eeprom,
2542 };
2543 
2544 /* Available operations in flow isolated mode. */
2545 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2546 	.dev_configure = mlx5_dev_configure,
2547 	.dev_start = mlx5_dev_start,
2548 	.dev_stop = mlx5_dev_stop,
2549 	.dev_set_link_down = mlx5_set_link_down,
2550 	.dev_set_link_up = mlx5_set_link_up,
2551 	.dev_close = mlx5_dev_close,
2552 	.promiscuous_enable = mlx5_promiscuous_enable,
2553 	.promiscuous_disable = mlx5_promiscuous_disable,
2554 	.allmulticast_enable = mlx5_allmulticast_enable,
2555 	.allmulticast_disable = mlx5_allmulticast_disable,
2556 	.link_update = mlx5_link_update,
2557 	.stats_get = mlx5_stats_get,
2558 	.stats_reset = mlx5_stats_reset,
2559 	.xstats_get = mlx5_xstats_get,
2560 	.xstats_reset = mlx5_xstats_reset,
2561 	.xstats_get_names = mlx5_xstats_get_names,
2562 	.fw_version_get = mlx5_fw_version_get,
2563 	.dev_infos_get = mlx5_dev_infos_get,
2564 	.read_clock = mlx5_txpp_read_clock,
2565 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2566 	.vlan_filter_set = mlx5_vlan_filter_set,
2567 	.rx_queue_setup = mlx5_rx_queue_setup,
2568 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2569 	.tx_queue_setup = mlx5_tx_queue_setup,
2570 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2571 	.rx_queue_release = mlx5_rx_queue_release,
2572 	.tx_queue_release = mlx5_tx_queue_release,
2573 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2574 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2575 	.mac_addr_remove = mlx5_mac_addr_remove,
2576 	.mac_addr_add = mlx5_mac_addr_add,
2577 	.mac_addr_set = mlx5_mac_addr_set,
2578 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2579 	.mtu_set = mlx5_dev_set_mtu,
2580 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2581 	.vlan_offload_set = mlx5_vlan_offload_set,
2582 	.filter_ctrl = mlx5_dev_filter_ctrl,
2583 	.rx_descriptor_status = mlx5_rx_descriptor_status,
2584 	.tx_descriptor_status = mlx5_tx_descriptor_status,
2585 	.rxq_info_get = mlx5_rxq_info_get,
2586 	.txq_info_get = mlx5_txq_info_get,
2587 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2588 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2589 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2590 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2591 	.is_removed = mlx5_is_removed,
2592 	.get_module_info = mlx5_get_module_info,
2593 	.get_module_eeprom = mlx5_get_module_eeprom,
2594 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2595 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2596 };
2597