xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision d84c3cf7662c6abca83df803aaa5136b73d3831d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2020 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <linux/rtnetlink.h>
14 #include <linux/sockios.h>
15 #include <linux/ethtool.h>
16 #include <fcntl.h>
17 
18 #include <rte_malloc.h>
19 #include <ethdev_driver.h>
20 #include <ethdev_pci.h>
21 #include <rte_pci.h>
22 #include <rte_bus_pci.h>
23 #include <rte_bus_auxiliary.h>
24 #include <rte_common.h>
25 #include <rte_kvargs.h>
26 #include <rte_rwlock.h>
27 #include <rte_spinlock.h>
28 #include <rte_string_fns.h>
29 #include <rte_alarm.h>
30 #include <rte_eal_paging.h>
31 
32 #include <mlx5_glue.h>
33 #include <mlx5_devx_cmds.h>
34 #include <mlx5_common.h>
35 #include <mlx5_common_mp.h>
36 #include <mlx5_common_mr.h>
37 #include <mlx5_malloc.h>
38 
39 #include "mlx5_defs.h"
40 #include "mlx5.h"
41 #include "mlx5_common_os.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_rx.h"
45 #include "mlx5_tx.h"
46 #include "mlx5_autoconf.h"
47 #include "mlx5_flow.h"
48 #include "rte_pmd_mlx5.h"
49 #include "mlx5_verbs.h"
50 #include "mlx5_nl.h"
51 #include "mlx5_devx.h"
52 
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /* rte flow indexed pool configuration. */
71 static struct mlx5_indexed_pool_config icfg[] = {
72 	{
73 		.size = sizeof(struct rte_flow),
74 		.trunk_size = 64,
75 		.need_lock = 1,
76 		.release_mem_en = 0,
77 		.malloc = mlx5_malloc,
78 		.free = mlx5_free,
79 		.per_core_cache = 0,
80 		.type = "ctl_flow_ipool",
81 	},
82 	{
83 		.size = sizeof(struct rte_flow),
84 		.trunk_size = 64,
85 		.grow_trunk = 3,
86 		.grow_shift = 2,
87 		.need_lock = 1,
88 		.release_mem_en = 0,
89 		.malloc = mlx5_malloc,
90 		.free = mlx5_free,
91 		.per_core_cache = 1 << 14,
92 		.type = "rte_flow_ipool",
93 	},
94 	{
95 		.size = sizeof(struct rte_flow),
96 		.trunk_size = 64,
97 		.grow_trunk = 3,
98 		.grow_shift = 2,
99 		.need_lock = 1,
100 		.release_mem_en = 0,
101 		.malloc = mlx5_malloc,
102 		.free = mlx5_free,
103 		.per_core_cache = 0,
104 		.type = "mcp_flow_ipool",
105 	},
106 };
107 
108 /**
109  * Set the completion channel file descriptor interrupt as non-blocking.
110  *
111  * @param[in] rxq_obj
112  *   Pointer to RQ channel object, which includes the channel fd
113  *
114  * @param[out] fd
115  *   The file descriptor (representing the interrupt) used in this channel.
116  *
117  * @return
118  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
119  */
120 int
121 mlx5_os_set_nonblock_channel_fd(int fd)
122 {
123 	int flags;
124 
125 	flags = fcntl(fd, F_GETFL);
126 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
127 }
128 
129 /**
130  * Get mlx5 device attributes. The glue function query_device_ex() is called
131  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
132  * device attributes from the glue out parameter.
133  *
134  * @param sh
135  *   Pointer to shared device context.
136  *
137  * @return
138  *   0 on success, a negative errno value otherwise and rte_errno is set.
139  */
140 int
141 mlx5_os_capabilities_prepare(struct mlx5_dev_ctx_shared *sh)
142 {
143 	int err;
144 	struct mlx5_common_device *cdev = sh->cdev;
145 	struct mlx5_hca_attr *hca_attr = &cdev->config.hca_attr;
146 	struct ibv_device_attr_ex attr_ex = { .comp_mask = 0 };
147 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
148 
149 	err = mlx5_glue->query_device_ex(cdev->ctx, NULL, &attr_ex);
150 	if (err) {
151 		rte_errno = errno;
152 		return -rte_errno;
153 	}
154 #ifdef HAVE_IBV_MLX5_MOD_SWP
155 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
156 #endif
157 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
158 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
159 #endif
160 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
161 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
162 #endif
163 	err = mlx5_glue->dv_query_device(cdev->ctx, &dv_attr);
164 	if (err) {
165 		rte_errno = errno;
166 		return -rte_errno;
167 	}
168 	memset(&sh->dev_cap, 0, sizeof(struct mlx5_dev_cap));
169 	if (mlx5_dev_is_pci(cdev->dev))
170 		sh->dev_cap.vf = mlx5_dev_is_vf_pci(RTE_DEV_TO_PCI(cdev->dev));
171 	else
172 		sh->dev_cap.sf = 1;
173 	sh->dev_cap.max_qp_wr = attr_ex.orig_attr.max_qp_wr;
174 	sh->dev_cap.max_sge = attr_ex.orig_attr.max_sge;
175 	sh->dev_cap.max_cq = attr_ex.orig_attr.max_cq;
176 	sh->dev_cap.max_qp = attr_ex.orig_attr.max_qp;
177 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
178 	sh->dev_cap.dest_tir = 1;
179 #endif
180 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) && defined(HAVE_MLX5DV_DR)
181 	DRV_LOG(DEBUG, "DV flow is supported.");
182 	sh->dev_cap.dv_flow_en = 1;
183 #endif
184 #ifdef HAVE_MLX5DV_DR_ESWITCH
185 	if (hca_attr->eswitch_manager && sh->dev_cap.dv_flow_en && sh->esw_mode)
186 		sh->dev_cap.dv_esw_en = 1;
187 #endif
188 	/*
189 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
190 	 * as all ConnectX-5 devices.
191 	 */
192 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
193 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
194 			DRV_LOG(DEBUG, "Enhanced MPW is supported.");
195 			sh->dev_cap.mps = MLX5_MPW_ENHANCED;
196 		} else {
197 			DRV_LOG(DEBUG, "MPW is supported.");
198 			sh->dev_cap.mps = MLX5_MPW;
199 		}
200 	} else {
201 		DRV_LOG(DEBUG, "MPW isn't supported.");
202 		sh->dev_cap.mps = MLX5_MPW_DISABLED;
203 	}
204 #if (RTE_CACHE_LINE_SIZE == 128)
205 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)
206 		sh->dev_cap.cqe_comp = 1;
207 	DRV_LOG(DEBUG, "Rx CQE 128B compression is %ssupported.",
208 		sh->dev_cap.cqe_comp ? "" : "not ");
209 #else
210 	sh->dev_cap.cqe_comp = 1;
211 #endif
212 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
213 	sh->dev_cap.mpls_en =
214 		((dv_attr.tunnel_offloads_caps &
215 		  MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
216 		 (dv_attr.tunnel_offloads_caps &
217 		  MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
218 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported.",
219 		sh->dev_cap.mpls_en ? "" : "not ");
220 #else
221 	DRV_LOG(WARNING,
222 		"MPLS over GRE/UDP tunnel offloading disabled due to old OFED/rdma-core version or firmware configuration");
223 #endif
224 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
225 	sh->dev_cap.hw_padding = !!attr_ex.rx_pad_end_addr_align;
226 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
227 	sh->dev_cap.hw_padding = !!(attr_ex.device_cap_flags_ex &
228 				    IBV_DEVICE_PCI_WRITE_END_PADDING);
229 #endif
230 	sh->dev_cap.hw_csum =
231 		!!(attr_ex.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
232 	DRV_LOG(DEBUG, "Checksum offloading is %ssupported.",
233 		sh->dev_cap.hw_csum ? "" : "not ");
234 	sh->dev_cap.hw_vlan_strip = !!(attr_ex.raw_packet_caps &
235 				       IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
236 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported.",
237 		(sh->dev_cap.hw_vlan_strip ? "" : "not "));
238 	sh->dev_cap.hw_fcs_strip = !!(attr_ex.raw_packet_caps &
239 				      IBV_RAW_PACKET_CAP_SCATTER_FCS);
240 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
241 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
242 	DRV_LOG(DEBUG, "Counters are not supported.");
243 #endif
244 	/*
245 	 * DPDK doesn't support larger/variable indirection tables.
246 	 * Once DPDK supports it, take max size from device attr.
247 	 */
248 	sh->dev_cap.ind_table_max_size =
249 			RTE_MIN(attr_ex.rss_caps.max_rwq_indirection_table_size,
250 				(unsigned int)RTE_ETH_RSS_RETA_SIZE_512);
251 	DRV_LOG(DEBUG, "Maximum Rx indirection table size is %u",
252 		sh->dev_cap.ind_table_max_size);
253 	sh->dev_cap.tso = (attr_ex.tso_caps.max_tso > 0 &&
254 			   (attr_ex.tso_caps.supported_qpts &
255 			    (1 << IBV_QPT_RAW_PACKET)));
256 	if (sh->dev_cap.tso)
257 		sh->dev_cap.tso_max_payload_sz = attr_ex.tso_caps.max_tso;
258 	strlcpy(sh->dev_cap.fw_ver, attr_ex.orig_attr.fw_ver,
259 		sizeof(sh->dev_cap.fw_ver));
260 #ifdef HAVE_IBV_MLX5_MOD_SWP
261 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
262 		sh->dev_cap.swp = dv_attr.sw_parsing_caps.sw_parsing_offloads &
263 				  (MLX5_SW_PARSING_CAP |
264 				   MLX5_SW_PARSING_CSUM_CAP |
265 				   MLX5_SW_PARSING_TSO_CAP);
266 	DRV_LOG(DEBUG, "SWP support: %u", sh->dev_cap.swp);
267 #endif
268 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
269 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
270 		struct mlx5dv_striding_rq_caps *strd_rq_caps =
271 				&dv_attr.striding_rq_caps;
272 
273 		sh->dev_cap.mprq.enabled = 1;
274 		sh->dev_cap.mprq.log_min_stride_size =
275 			strd_rq_caps->min_single_stride_log_num_of_bytes;
276 		sh->dev_cap.mprq.log_max_stride_size =
277 			strd_rq_caps->max_single_stride_log_num_of_bytes;
278 		sh->dev_cap.mprq.log_min_stride_num =
279 			strd_rq_caps->min_single_wqe_log_num_of_strides;
280 		sh->dev_cap.mprq.log_max_stride_num =
281 			strd_rq_caps->max_single_wqe_log_num_of_strides;
282 		sh->dev_cap.mprq.log_min_stride_wqe_size =
283 					cdev->config.devx ?
284 					hca_attr->log_min_stride_wqe_sz :
285 					MLX5_MPRQ_LOG_MIN_STRIDE_WQE_SIZE;
286 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %u",
287 			sh->dev_cap.mprq.log_min_stride_size);
288 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %u",
289 			sh->dev_cap.mprq.log_max_stride_size);
290 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %u",
291 			sh->dev_cap.mprq.log_min_stride_num);
292 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %u",
293 			sh->dev_cap.mprq.log_max_stride_num);
294 		DRV_LOG(DEBUG, "\tmin_stride_wqe_log_size: %u",
295 			sh->dev_cap.mprq.log_min_stride_wqe_size);
296 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
297 			strd_rq_caps->supported_qpts);
298 		DRV_LOG(DEBUG, "Device supports Multi-Packet RQ.");
299 	}
300 #endif
301 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
302 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
303 		sh->dev_cap.tunnel_en = dv_attr.tunnel_offloads_caps &
304 					(MLX5_TUNNELED_OFFLOADS_VXLAN_CAP |
305 					 MLX5_TUNNELED_OFFLOADS_GRE_CAP |
306 					 MLX5_TUNNELED_OFFLOADS_GENEVE_CAP);
307 	}
308 	if (sh->dev_cap.tunnel_en) {
309 		DRV_LOG(DEBUG, "Tunnel offloading is supported for %s%s%s",
310 			sh->dev_cap.tunnel_en &
311 			MLX5_TUNNELED_OFFLOADS_VXLAN_CAP ? "[VXLAN]" : "",
312 			sh->dev_cap.tunnel_en &
313 			MLX5_TUNNELED_OFFLOADS_GRE_CAP ? "[GRE]" : "",
314 			sh->dev_cap.tunnel_en &
315 			MLX5_TUNNELED_OFFLOADS_GENEVE_CAP ? "[GENEVE]" : "");
316 	} else {
317 		DRV_LOG(DEBUG, "Tunnel offloading is not supported.");
318 	}
319 #else
320 	DRV_LOG(WARNING,
321 		"Tunnel offloading disabled due to old OFED/rdma-core version");
322 #endif
323 	if (!sh->cdev->config.devx)
324 		return 0;
325 	/* Check capabilities for Packet Pacing. */
326 	DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz.",
327 		hca_attr->dev_freq_khz);
328 	DRV_LOG(DEBUG, "Packet pacing is %ssupported.",
329 		hca_attr->qos.packet_pacing ? "" : "not ");
330 	DRV_LOG(DEBUG, "Cross channel ops are %ssupported.",
331 		hca_attr->cross_channel ? "" : "not ");
332 	DRV_LOG(DEBUG, "WQE index ignore is %ssupported.",
333 		hca_attr->wqe_index_ignore ? "" : "not ");
334 	DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported.",
335 		hca_attr->non_wire_sq ? "" : "not ");
336 	DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
337 		hca_attr->log_max_static_sq_wq ? "" : "not ",
338 		hca_attr->log_max_static_sq_wq);
339 	DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported.",
340 		hca_attr->qos.wqe_rate_pp ? "" : "not ");
341 	sh->dev_cap.txpp_en = hca_attr->qos.packet_pacing;
342 	if (!hca_attr->cross_channel) {
343 		DRV_LOG(DEBUG,
344 			"Cross channel operations are required for packet pacing.");
345 		sh->dev_cap.txpp_en = 0;
346 	}
347 	if (!hca_attr->wqe_index_ignore) {
348 		DRV_LOG(DEBUG,
349 			"WQE index ignore feature is required for packet pacing.");
350 		sh->dev_cap.txpp_en = 0;
351 	}
352 	if (!hca_attr->non_wire_sq) {
353 		DRV_LOG(DEBUG,
354 			"Non-wire SQ feature is required for packet pacing.");
355 		sh->dev_cap.txpp_en = 0;
356 	}
357 	if (!hca_attr->log_max_static_sq_wq) {
358 		DRV_LOG(DEBUG,
359 			"Static WQE SQ feature is required for packet pacing.");
360 		sh->dev_cap.txpp_en = 0;
361 	}
362 	if (!hca_attr->qos.wqe_rate_pp) {
363 		DRV_LOG(DEBUG,
364 			"WQE rate mode is required for packet pacing.");
365 		sh->dev_cap.txpp_en = 0;
366 	}
367 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
368 	DRV_LOG(DEBUG,
369 		"DevX does not provide UAR offset, can't create queues for packet pacing.");
370 	sh->dev_cap.txpp_en = 0;
371 #endif
372 	/* Check for LRO support. */
373 	if (mlx5_devx_obj_ops_en(sh) && hca_attr->lro_cap) {
374 		/* TBD check tunnel lro caps. */
375 		sh->dev_cap.lro_supported = 1;
376 		DRV_LOG(DEBUG, "Device supports LRO.");
377 		DRV_LOG(DEBUG,
378 			"LRO minimal size of TCP segment required for coalescing is %d bytes.",
379 			hca_attr->lro_min_mss_size);
380 	}
381 	sh->dev_cap.scatter_fcs_w_decap_disable =
382 					hca_attr->scatter_fcs_w_decap_disable;
383 	sh->dev_cap.rq_delay_drop_en = hca_attr->rq_delay_drop;
384 	mlx5_rt_timestamp_config(sh, hca_attr);
385 	return 0;
386 }
387 
388 /**
389  * Detect misc5 support or not
390  *
391  * @param[in] priv
392  *   Device private data pointer
393  */
394 #ifdef HAVE_MLX5DV_DR
395 static void
396 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
397 {
398 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
399 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
400 	 * Case: IPv4--->UDP--->VxLAN--->vni
401 	 */
402 	void *tbl;
403 	struct mlx5_flow_dv_match_params matcher_mask;
404 	void *match_m;
405 	void *matcher;
406 	void *headers_m;
407 	void *misc5_m;
408 	uint32_t *tunnel_header_m;
409 	struct mlx5dv_flow_matcher_attr dv_attr;
410 
411 	memset(&matcher_mask, 0, sizeof(matcher_mask));
412 	matcher_mask.size = sizeof(matcher_mask.buf);
413 	match_m = matcher_mask.buf;
414 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
415 	misc5_m = MLX5_ADDR_OF(fte_match_param,
416 			       match_m, misc_parameters_5);
417 	tunnel_header_m = (uint32_t *)
418 				MLX5_ADDR_OF(fte_match_set_misc5,
419 				misc5_m, tunnel_header_1);
420 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
421 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
422 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
423 	*tunnel_header_m = 0xffffff;
424 
425 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
426 	if (!tbl) {
427 		DRV_LOG(INFO, "No SW steering support");
428 		return;
429 	}
430 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
431 	dv_attr.match_mask = (void *)&matcher_mask,
432 	dv_attr.match_criteria_enable =
433 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
434 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
435 	dv_attr.priority = 3;
436 #ifdef HAVE_MLX5DV_DR_ESWITCH
437 	void *misc2_m;
438 	if (priv->sh->config.dv_esw_en) {
439 		/* FDB enabled reg_c_0 */
440 		dv_attr.match_criteria_enable |=
441 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
442 		misc2_m = MLX5_ADDR_OF(fte_match_param,
443 				       match_m, misc_parameters_2);
444 		MLX5_SET(fte_match_set_misc2, misc2_m,
445 			 metadata_reg_c_0, 0xffff);
446 	}
447 #endif
448 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
449 						    &dv_attr, tbl);
450 	if (matcher) {
451 		priv->sh->misc5_cap = 1;
452 		mlx5_glue->dv_destroy_flow_matcher(matcher);
453 	}
454 	mlx5_glue->dr_destroy_flow_tbl(tbl);
455 #else
456 	RTE_SET_USED(priv);
457 #endif
458 }
459 #endif
460 
461 /**
462  * Initialize DR related data within private structure.
463  * Routine checks the reference counter and does actual
464  * resources creation/initialization only if counter is zero.
465  *
466  * @param[in] priv
467  *   Pointer to the private device data structure.
468  *
469  * @return
470  *   Zero on success, positive error code otherwise.
471  */
472 static int
473 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
474 {
475 	struct mlx5_dev_ctx_shared *sh = priv->sh;
476 	char s[MLX5_NAME_SIZE] __rte_unused;
477 	int err;
478 
479 	MLX5_ASSERT(sh && sh->refcnt);
480 	if (sh->refcnt > 1)
481 		return 0;
482 	err = mlx5_alloc_table_hash_list(priv);
483 	if (err)
484 		goto error;
485 	if (priv->sh->config.dv_flow_en == 2)
486 		return 0;
487 	/* The resources below are only valid with DV support. */
488 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
489 	/* Init port id action list. */
490 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
491 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
492 						   flow_dv_port_id_create_cb,
493 						   flow_dv_port_id_match_cb,
494 						   flow_dv_port_id_remove_cb,
495 						   flow_dv_port_id_clone_cb,
496 						 flow_dv_port_id_clone_free_cb);
497 	if (!sh->port_id_action_list)
498 		goto error;
499 	/* Init push vlan action list. */
500 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
501 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
502 						    flow_dv_push_vlan_create_cb,
503 						    flow_dv_push_vlan_match_cb,
504 						    flow_dv_push_vlan_remove_cb,
505 						    flow_dv_push_vlan_clone_cb,
506 					       flow_dv_push_vlan_clone_free_cb);
507 	if (!sh->push_vlan_action_list)
508 		goto error;
509 	/* Init sample action list. */
510 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
511 	sh->sample_action_list = mlx5_list_create(s, sh, true,
512 						  flow_dv_sample_create_cb,
513 						  flow_dv_sample_match_cb,
514 						  flow_dv_sample_remove_cb,
515 						  flow_dv_sample_clone_cb,
516 						  flow_dv_sample_clone_free_cb);
517 	if (!sh->sample_action_list)
518 		goto error;
519 	/* Init dest array action list. */
520 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
521 	sh->dest_array_list = mlx5_list_create(s, sh, true,
522 					       flow_dv_dest_array_create_cb,
523 					       flow_dv_dest_array_match_cb,
524 					       flow_dv_dest_array_remove_cb,
525 					       flow_dv_dest_array_clone_cb,
526 					      flow_dv_dest_array_clone_free_cb);
527 	if (!sh->dest_array_list)
528 		goto error;
529 	/* Init shared flex parsers list, no need lcore_share */
530 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
531 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
532 					       mlx5_flex_parser_create_cb,
533 					       mlx5_flex_parser_match_cb,
534 					       mlx5_flex_parser_remove_cb,
535 					       mlx5_flex_parser_clone_cb,
536 					       mlx5_flex_parser_clone_free_cb);
537 	if (!sh->flex_parsers_dv)
538 		goto error;
539 #endif
540 #ifdef HAVE_MLX5DV_DR
541 	void *domain;
542 
543 	/* Reference counter is zero, we should initialize structures. */
544 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
545 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
546 	if (!domain) {
547 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
548 		err = errno;
549 		goto error;
550 	}
551 	sh->rx_domain = domain;
552 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
553 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
554 	if (!domain) {
555 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
556 		err = errno;
557 		goto error;
558 	}
559 	sh->tx_domain = domain;
560 #ifdef HAVE_MLX5DV_DR_ESWITCH
561 	if (sh->config.dv_esw_en) {
562 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
563 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
564 		if (!domain) {
565 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
566 			err = errno;
567 			goto error;
568 		}
569 		sh->fdb_domain = domain;
570 	}
571 	/*
572 	 * The drop action is just some dummy placeholder in rdma-core. It
573 	 * does not belong to domains and has no any attributes, and, can be
574 	 * shared by the entire device.
575 	 */
576 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
577 	if (!sh->dr_drop_action) {
578 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
579 		err = errno;
580 		goto error;
581 	}
582 #endif
583 	if (!sh->tunnel_hub && sh->config.dv_miss_info)
584 		err = mlx5_alloc_tunnel_hub(sh);
585 	if (err) {
586 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
587 		goto error;
588 	}
589 	if (sh->config.reclaim_mode == MLX5_RCM_AGGR) {
590 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
591 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
592 		if (sh->fdb_domain)
593 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
594 	}
595 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
596 	if (!sh->config.allow_duplicate_pattern) {
597 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
598 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
599 #endif
600 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
601 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
602 		if (sh->fdb_domain)
603 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
604 	}
605 
606 	__mlx5_discovery_misc5_cap(priv);
607 #endif /* HAVE_MLX5DV_DR */
608 	sh->default_miss_action =
609 			mlx5_glue->dr_create_flow_action_default_miss();
610 	if (!sh->default_miss_action)
611 		DRV_LOG(WARNING, "Default miss action is not supported.");
612 	LIST_INIT(&sh->shared_rxqs);
613 	return 0;
614 error:
615 	/* Rollback the created objects. */
616 	if (sh->rx_domain) {
617 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
618 		sh->rx_domain = NULL;
619 	}
620 	if (sh->tx_domain) {
621 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
622 		sh->tx_domain = NULL;
623 	}
624 	if (sh->fdb_domain) {
625 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
626 		sh->fdb_domain = NULL;
627 	}
628 	if (sh->dr_drop_action) {
629 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
630 		sh->dr_drop_action = NULL;
631 	}
632 	if (sh->pop_vlan_action) {
633 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
634 		sh->pop_vlan_action = NULL;
635 	}
636 	if (sh->encaps_decaps) {
637 		mlx5_hlist_destroy(sh->encaps_decaps);
638 		sh->encaps_decaps = NULL;
639 	}
640 	if (sh->modify_cmds) {
641 		mlx5_hlist_destroy(sh->modify_cmds);
642 		sh->modify_cmds = NULL;
643 	}
644 	if (sh->tag_table) {
645 		/* tags should be destroyed with flow before. */
646 		mlx5_hlist_destroy(sh->tag_table);
647 		sh->tag_table = NULL;
648 	}
649 	if (sh->tunnel_hub) {
650 		mlx5_release_tunnel_hub(sh, priv->dev_port);
651 		sh->tunnel_hub = NULL;
652 	}
653 	mlx5_free_table_hash_list(priv);
654 	if (sh->port_id_action_list) {
655 		mlx5_list_destroy(sh->port_id_action_list);
656 		sh->port_id_action_list = NULL;
657 	}
658 	if (sh->push_vlan_action_list) {
659 		mlx5_list_destroy(sh->push_vlan_action_list);
660 		sh->push_vlan_action_list = NULL;
661 	}
662 	if (sh->sample_action_list) {
663 		mlx5_list_destroy(sh->sample_action_list);
664 		sh->sample_action_list = NULL;
665 	}
666 	if (sh->dest_array_list) {
667 		mlx5_list_destroy(sh->dest_array_list);
668 		sh->dest_array_list = NULL;
669 	}
670 	return err;
671 }
672 
673 /**
674  * Destroy DR related data within private structure.
675  *
676  * @param[in] priv
677  *   Pointer to the private device data structure.
678  */
679 void
680 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
681 {
682 	struct mlx5_dev_ctx_shared *sh = priv->sh;
683 
684 	MLX5_ASSERT(sh && sh->refcnt);
685 	if (sh->refcnt > 1)
686 		return;
687 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
688 #ifdef HAVE_MLX5DV_DR
689 	if (sh->rx_domain) {
690 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
691 		sh->rx_domain = NULL;
692 	}
693 	if (sh->tx_domain) {
694 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
695 		sh->tx_domain = NULL;
696 	}
697 #ifdef HAVE_MLX5DV_DR_ESWITCH
698 	if (sh->fdb_domain) {
699 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
700 		sh->fdb_domain = NULL;
701 	}
702 	if (sh->dr_drop_action) {
703 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
704 		sh->dr_drop_action = NULL;
705 	}
706 #endif
707 	if (sh->pop_vlan_action) {
708 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
709 		sh->pop_vlan_action = NULL;
710 	}
711 #endif /* HAVE_MLX5DV_DR */
712 	if (sh->default_miss_action)
713 		mlx5_glue->destroy_flow_action
714 				(sh->default_miss_action);
715 	if (sh->encaps_decaps) {
716 		mlx5_hlist_destroy(sh->encaps_decaps);
717 		sh->encaps_decaps = NULL;
718 	}
719 	if (sh->modify_cmds) {
720 		mlx5_hlist_destroy(sh->modify_cmds);
721 		sh->modify_cmds = NULL;
722 	}
723 	if (sh->tag_table) {
724 		/* tags should be destroyed with flow before. */
725 		mlx5_hlist_destroy(sh->tag_table);
726 		sh->tag_table = NULL;
727 	}
728 	if (sh->tunnel_hub) {
729 		mlx5_release_tunnel_hub(sh, priv->dev_port);
730 		sh->tunnel_hub = NULL;
731 	}
732 	mlx5_free_table_hash_list(priv);
733 	if (sh->port_id_action_list) {
734 		mlx5_list_destroy(sh->port_id_action_list);
735 		sh->port_id_action_list = NULL;
736 	}
737 	if (sh->push_vlan_action_list) {
738 		mlx5_list_destroy(sh->push_vlan_action_list);
739 		sh->push_vlan_action_list = NULL;
740 	}
741 	if (sh->sample_action_list) {
742 		mlx5_list_destroy(sh->sample_action_list);
743 		sh->sample_action_list = NULL;
744 	}
745 	if (sh->dest_array_list) {
746 		mlx5_list_destroy(sh->dest_array_list);
747 		sh->dest_array_list = NULL;
748 	}
749 }
750 
751 /**
752  * Initialize shared data between primary and secondary process.
753  *
754  * A memzone is reserved by primary process and secondary processes attach to
755  * the memzone.
756  *
757  * @return
758  *   0 on success, a negative errno value otherwise and rte_errno is set.
759  */
760 static int
761 mlx5_init_shared_data(void)
762 {
763 	const struct rte_memzone *mz;
764 	int ret = 0;
765 
766 	rte_spinlock_lock(&mlx5_shared_data_lock);
767 	if (mlx5_shared_data == NULL) {
768 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
769 			/* Allocate shared memory. */
770 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
771 						 sizeof(*mlx5_shared_data),
772 						 SOCKET_ID_ANY, 0);
773 			if (mz == NULL) {
774 				DRV_LOG(ERR,
775 					"Cannot allocate mlx5 shared data");
776 				ret = -rte_errno;
777 				goto error;
778 			}
779 			mlx5_shared_data = mz->addr;
780 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
781 			rte_spinlock_init(&mlx5_shared_data->lock);
782 		} else {
783 			/* Lookup allocated shared memory. */
784 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
785 			if (mz == NULL) {
786 				DRV_LOG(ERR,
787 					"Cannot attach mlx5 shared data");
788 				ret = -rte_errno;
789 				goto error;
790 			}
791 			mlx5_shared_data = mz->addr;
792 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
793 		}
794 	}
795 error:
796 	rte_spinlock_unlock(&mlx5_shared_data_lock);
797 	return ret;
798 }
799 
800 /**
801  * PMD global initialization.
802  *
803  * Independent from individual device, this function initializes global
804  * per-PMD data structures distinguishing primary and secondary processes.
805  * Hence, each initialization is called once per a process.
806  *
807  * @return
808  *   0 on success, a negative errno value otherwise and rte_errno is set.
809  */
810 static int
811 mlx5_init_once(void)
812 {
813 	struct mlx5_shared_data *sd;
814 	struct mlx5_local_data *ld = &mlx5_local_data;
815 	int ret = 0;
816 
817 	if (mlx5_init_shared_data())
818 		return -rte_errno;
819 	sd = mlx5_shared_data;
820 	MLX5_ASSERT(sd);
821 	rte_spinlock_lock(&sd->lock);
822 	switch (rte_eal_process_type()) {
823 	case RTE_PROC_PRIMARY:
824 		if (sd->init_done)
825 			break;
826 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
827 					   mlx5_mp_os_primary_handle);
828 		if (ret)
829 			goto out;
830 		sd->init_done = true;
831 		break;
832 	case RTE_PROC_SECONDARY:
833 		if (ld->init_done)
834 			break;
835 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
836 					     mlx5_mp_os_secondary_handle);
837 		if (ret)
838 			goto out;
839 		++sd->secondary_cnt;
840 		ld->init_done = true;
841 		break;
842 	default:
843 		break;
844 	}
845 out:
846 	rte_spinlock_unlock(&sd->lock);
847 	return ret;
848 }
849 
850 /**
851  * DR flow drop action support detect.
852  *
853  * @param dev
854  *   Pointer to rte_eth_dev structure.
855  *
856  */
857 static void
858 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
859 {
860 #ifdef HAVE_MLX5DV_DR
861 	struct mlx5_priv *priv = dev->data->dev_private;
862 
863 	if (!priv->sh->config.dv_flow_en || !priv->sh->dr_drop_action)
864 		return;
865 	/**
866 	 * DR supports drop action placeholder when it is supported;
867 	 * otherwise, use the queue drop action.
868 	 */
869 	if (!priv->sh->drop_action_check_flag) {
870 		if (!mlx5_flow_discover_dr_action_support(dev))
871 			priv->sh->dr_drop_action_en = 1;
872 		priv->sh->drop_action_check_flag = 1;
873 	}
874 	if (priv->sh->dr_drop_action_en)
875 		priv->root_drop_action = priv->sh->dr_drop_action;
876 	else
877 		priv->root_drop_action = priv->drop_queue.hrxq->action;
878 #endif
879 }
880 
881 static void
882 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
883 {
884 	struct mlx5_priv *priv = dev->data->dev_private;
885 	void *ctx = priv->sh->cdev->ctx;
886 
887 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
888 	if (!priv->q_counters) {
889 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
890 		struct ibv_wq *wq;
891 
892 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
893 			"by DevX - fall-back to use the kernel driver global "
894 			"queue counter.", dev->data->port_id);
895 		/* Create WQ by kernel and query its queue counter ID. */
896 		if (cq) {
897 			wq = mlx5_glue->create_wq(ctx,
898 						  &(struct ibv_wq_init_attr){
899 						    .wq_type = IBV_WQT_RQ,
900 						    .max_wr = 1,
901 						    .max_sge = 1,
902 						    .pd = priv->sh->cdev->pd,
903 						    .cq = cq,
904 						});
905 			if (wq) {
906 				/* Counter is assigned only on RDY state. */
907 				int ret = mlx5_glue->modify_wq(wq,
908 						 &(struct ibv_wq_attr){
909 						 .attr_mask = IBV_WQ_ATTR_STATE,
910 						 .wq_state = IBV_WQS_RDY,
911 						});
912 
913 				if (ret == 0)
914 					mlx5_devx_cmd_wq_query(wq,
915 							 &priv->counter_set_id);
916 				claim_zero(mlx5_glue->destroy_wq(wq));
917 			}
918 			claim_zero(mlx5_glue->destroy_cq(cq));
919 		}
920 	} else {
921 		priv->counter_set_id = priv->q_counters->id;
922 	}
923 	if (priv->counter_set_id == 0)
924 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
925 			"available.", dev->data->port_id);
926 }
927 
928 /**
929  * Check if representor spawn info match devargs.
930  *
931  * @param spawn
932  *   Verbs device parameters (name, port, switch_info) to spawn.
933  * @param eth_da
934  *   Device devargs to probe.
935  *
936  * @return
937  *   Match result.
938  */
939 static bool
940 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
941 		       struct rte_eth_devargs *eth_da)
942 {
943 	struct mlx5_switch_info *switch_info = &spawn->info;
944 	unsigned int p, f;
945 	uint16_t id;
946 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
947 						      eth_da->type);
948 
949 	switch (eth_da->type) {
950 	case RTE_ETH_REPRESENTOR_SF:
951 		if (!(spawn->info.port_name == -1 &&
952 		      switch_info->name_type ==
953 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
954 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
955 			rte_errno = EBUSY;
956 			return false;
957 		}
958 		break;
959 	case RTE_ETH_REPRESENTOR_VF:
960 		/* Allows HPF representor index -1 as exception. */
961 		if (!(spawn->info.port_name == -1 &&
962 		      switch_info->name_type ==
963 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
964 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
965 			rte_errno = EBUSY;
966 			return false;
967 		}
968 		break;
969 	case RTE_ETH_REPRESENTOR_NONE:
970 		rte_errno = EBUSY;
971 		return false;
972 	default:
973 		rte_errno = ENOTSUP;
974 		DRV_LOG(ERR, "unsupported representor type");
975 		return false;
976 	}
977 	/* Check representor ID: */
978 	for (p = 0; p < eth_da->nb_ports; ++p) {
979 		if (spawn->pf_bond < 0) {
980 			/* For non-LAG mode, allow and ignore pf. */
981 			switch_info->pf_num = eth_da->ports[p];
982 			repr_id = mlx5_representor_id_encode(switch_info,
983 							     eth_da->type);
984 		}
985 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
986 			id = MLX5_REPRESENTOR_ID
987 				(eth_da->ports[p], eth_da->type,
988 				 eth_da->representor_ports[f]);
989 			if (repr_id == id)
990 				return true;
991 		}
992 	}
993 	rte_errno = EBUSY;
994 	return false;
995 }
996 
997 /**
998  * Spawn an Ethernet device from Verbs information.
999  *
1000  * @param dpdk_dev
1001  *   Backing DPDK device.
1002  * @param spawn
1003  *   Verbs device parameters (name, port, switch_info) to spawn.
1004  * @param eth_da
1005  *   Device arguments.
1006  * @param mkvlist
1007  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1008  *
1009  * @return
1010  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1011  *   is set. The following errors are defined:
1012  *
1013  *   EBUSY: device is not supposed to be spawned.
1014  *   EEXIST: device is already spawned
1015  */
1016 static struct rte_eth_dev *
1017 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1018 	       struct mlx5_dev_spawn_data *spawn,
1019 	       struct rte_eth_devargs *eth_da,
1020 	       struct mlx5_kvargs_ctrl *mkvlist)
1021 {
1022 	const struct mlx5_switch_info *switch_info = &spawn->info;
1023 	struct mlx5_dev_ctx_shared *sh = NULL;
1024 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
1025 	struct rte_eth_dev *eth_dev = NULL;
1026 	struct mlx5_priv *priv = NULL;
1027 	int err = 0;
1028 	struct rte_ether_addr mac;
1029 	char name[RTE_ETH_NAME_MAX_LEN];
1030 	int own_domain_id = 0;
1031 	uint16_t port_id;
1032 	struct mlx5_port_info vport_info = { .query_flags = 0 };
1033 	int nl_rdma;
1034 	int i;
1035 
1036 	/* Determine if this port representor is supposed to be spawned. */
1037 	if (switch_info->representor && dpdk_dev->devargs &&
1038 	    !mlx5_representor_match(spawn, eth_da))
1039 		return NULL;
1040 	/* Build device name. */
1041 	if (spawn->pf_bond < 0) {
1042 		/* Single device. */
1043 		if (!switch_info->representor)
1044 			strlcpy(name, dpdk_dev->name, sizeof(name));
1045 		else
1046 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
1047 				 dpdk_dev->name,
1048 				 switch_info->name_type ==
1049 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1050 				 switch_info->port_name);
1051 	} else {
1052 		/* Bonding device. */
1053 		if (!switch_info->representor) {
1054 			err = snprintf(name, sizeof(name), "%s_%s",
1055 				       dpdk_dev->name, spawn->phys_dev_name);
1056 		} else {
1057 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1058 				dpdk_dev->name, spawn->phys_dev_name,
1059 				switch_info->ctrl_num,
1060 				switch_info->pf_num,
1061 				switch_info->name_type ==
1062 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1063 				switch_info->port_name);
1064 		}
1065 	}
1066 	if (err >= (int)sizeof(name))
1067 		DRV_LOG(WARNING, "device name overflow %s", name);
1068 	/* check if the device is already spawned */
1069 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1070 		/*
1071 		 * When device is already spawned, its devargs should be set
1072 		 * as used. otherwise, mlx5_kvargs_validate() will fail.
1073 		 */
1074 		if (mkvlist)
1075 			mlx5_port_args_set_used(name, port_id, mkvlist);
1076 		rte_errno = EEXIST;
1077 		return NULL;
1078 	}
1079 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1080 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1081 		struct mlx5_mp_id mp_id;
1082 
1083 		eth_dev = rte_eth_dev_attach_secondary(name);
1084 		if (eth_dev == NULL) {
1085 			DRV_LOG(ERR, "can not attach rte ethdev");
1086 			rte_errno = ENOMEM;
1087 			return NULL;
1088 		}
1089 		eth_dev->device = dpdk_dev;
1090 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1091 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1092 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1093 		err = mlx5_proc_priv_init(eth_dev);
1094 		if (err)
1095 			return NULL;
1096 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1097 		/* Receive command fd from primary process */
1098 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1099 		if (err < 0)
1100 			goto err_secondary;
1101 		/* Remap UAR for Tx queues. */
1102 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1103 		if (err)
1104 			goto err_secondary;
1105 		/*
1106 		 * Ethdev pointer is still required as input since
1107 		 * the primary device is not accessible from the
1108 		 * secondary process.
1109 		 */
1110 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1111 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1112 		return eth_dev;
1113 err_secondary:
1114 		mlx5_dev_close(eth_dev);
1115 		return NULL;
1116 	}
1117 	sh = mlx5_alloc_shared_dev_ctx(spawn, mkvlist);
1118 	if (!sh)
1119 		return NULL;
1120 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1121 	/* Check port status. */
1122 	if (spawn->phys_port <= UINT8_MAX) {
1123 		/* Legacy Verbs api only support u8 port number. */
1124 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1125 					    &port_attr);
1126 		if (err) {
1127 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1128 			goto error;
1129 		}
1130 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1131 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1132 			err = EINVAL;
1133 			goto error;
1134 		}
1135 	} else if (nl_rdma >= 0) {
1136 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1137 		err = mlx5_nl_port_state(nl_rdma,
1138 			spawn->phys_dev_name,
1139 			spawn->phys_port);
1140 		if (err < 0) {
1141 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1142 				strerror(rte_errno));
1143 			err = -rte_errno;
1144 			goto error;
1145 		}
1146 		port_attr.state = (enum ibv_port_state)err;
1147 	}
1148 	if (port_attr.state != IBV_PORT_ACTIVE)
1149 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1150 			mlx5_glue->port_state_str(port_attr.state),
1151 			port_attr.state);
1152 	/* Allocate private eth device data. */
1153 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1154 			   sizeof(*priv),
1155 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1156 	if (priv == NULL) {
1157 		DRV_LOG(ERR, "priv allocation failure");
1158 		err = ENOMEM;
1159 		goto error;
1160 	}
1161 	priv->sh = sh;
1162 	priv->dev_port = spawn->phys_port;
1163 	priv->pci_dev = spawn->pci_dev;
1164 	priv->mtu = RTE_ETHER_MTU;
1165 	/* Some internal functions rely on Netlink sockets, open them now. */
1166 	priv->nl_socket_rdma = nl_rdma;
1167 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1168 	priv->representor = !!switch_info->representor;
1169 	priv->master = !!switch_info->master;
1170 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1171 	priv->vport_meta_tag = 0;
1172 	priv->vport_meta_mask = 0;
1173 	priv->pf_bond = spawn->pf_bond;
1174 
1175 	DRV_LOG(DEBUG,
1176 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1177 		priv->dev_port, dpdk_dev->bus->name,
1178 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1179 		priv->master, priv->representor, priv->pf_bond);
1180 
1181 	/*
1182 	 * If we have E-Switch we should determine the vport attributes.
1183 	 * E-Switch may use either source vport field or reg_c[0] metadata
1184 	 * register to match on vport index. The engaged part of metadata
1185 	 * register is defined by mask.
1186 	 */
1187 	if (sh->esw_mode) {
1188 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1189 						 spawn->phys_port,
1190 						 &vport_info);
1191 		if (err) {
1192 			DRV_LOG(WARNING,
1193 				"Cannot query devx port %d on device %s",
1194 				spawn->phys_port, spawn->phys_dev_name);
1195 			vport_info.query_flags = 0;
1196 		}
1197 	}
1198 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1199 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1200 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1201 		if (!priv->vport_meta_mask) {
1202 			DRV_LOG(ERR,
1203 				"vport zero mask for port %d on bonding device %s",
1204 				spawn->phys_port, spawn->phys_dev_name);
1205 			err = ENOTSUP;
1206 			goto error;
1207 		}
1208 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1209 			DRV_LOG(ERR,
1210 				"Invalid vport tag for port %d on bonding device %s",
1211 				spawn->phys_port, spawn->phys_dev_name);
1212 			err = ENOTSUP;
1213 			goto error;
1214 		}
1215 	}
1216 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1217 		priv->vport_id = vport_info.vport_id;
1218 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1219 		DRV_LOG(ERR,
1220 			"Cannot deduce vport index for port %d on bonding device %s",
1221 			spawn->phys_port, spawn->phys_dev_name);
1222 		err = ENOTSUP;
1223 		goto error;
1224 	} else {
1225 		/*
1226 		 * Suppose vport index in compatible way. Kernel/rdma_core
1227 		 * support single E-Switch per PF configurations only and
1228 		 * vport_id field contains the vport index for associated VF,
1229 		 * which is deduced from representor port name.
1230 		 * For example, let's have the IB device port 10, it has
1231 		 * attached network device eth0, which has port name attribute
1232 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1233 		 * as 3 (2+1). This assigning schema should be changed if the
1234 		 * multiple E-Switch instances per PF configurations or/and PCI
1235 		 * subfunctions are added.
1236 		 */
1237 		priv->vport_id = switch_info->representor ?
1238 				 switch_info->port_name + 1 : -1;
1239 	}
1240 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1241 							  eth_da->type);
1242 	/*
1243 	 * Look for sibling devices in order to reuse their switch domain
1244 	 * if any, otherwise allocate one.
1245 	 */
1246 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1247 		const struct mlx5_priv *opriv =
1248 			rte_eth_devices[port_id].data->dev_private;
1249 
1250 		if (!opriv ||
1251 		    opriv->sh != priv->sh ||
1252 			opriv->domain_id ==
1253 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1254 			continue;
1255 		priv->domain_id = opriv->domain_id;
1256 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1257 			priv->dev_port, priv->domain_id);
1258 		break;
1259 	}
1260 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1261 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1262 		if (err) {
1263 			err = rte_errno;
1264 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1265 				strerror(rte_errno));
1266 			goto error;
1267 		}
1268 		own_domain_id = 1;
1269 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1270 			priv->dev_port, priv->domain_id);
1271 	}
1272 	if (sh->cdev->config.devx) {
1273 		struct mlx5_hca_attr *hca_attr = &sh->cdev->config.hca_attr;
1274 
1275 		sh->steering_format_version = hca_attr->steering_format_version;
1276 #if defined(HAVE_MLX5DV_DR) && \
1277 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1278 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1279 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1280 		    sh->config.dv_flow_en) {
1281 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1282 			/*
1283 			 * Meter needs two REG_C's for color match and pre-sfx
1284 			 * flow match. Here get the REG_C for color match.
1285 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1286 			 */
1287 			reg_c_mask &= 0xfc;
1288 			if (__builtin_popcount(reg_c_mask) < 1) {
1289 				priv->mtr_en = 0;
1290 				DRV_LOG(WARNING, "No available register for"
1291 					" meter.");
1292 			} else {
1293 				/*
1294 				 * The meter color register is used by the
1295 				 * flow-hit feature as well.
1296 				 * The flow-hit feature must use REG_C_3
1297 				 * Prefer REG_C_3 if it is available.
1298 				 */
1299 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1300 					priv->mtr_color_reg = REG_C_3;
1301 				else
1302 					priv->mtr_color_reg = ffs(reg_c_mask)
1303 							      - 1 + REG_C_0;
1304 				priv->mtr_en = 1;
1305 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1306 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1307 					priv->mtr_color_reg);
1308 			}
1309 		}
1310 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1311 			uint32_t log_obj_size =
1312 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1313 			if (log_obj_size >=
1314 			    hca_attr->qos.log_meter_aso_granularity &&
1315 			    log_obj_size <=
1316 			    hca_attr->qos.log_meter_aso_max_alloc)
1317 				sh->meter_aso_en = 1;
1318 		}
1319 		if (priv->mtr_en) {
1320 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1321 			if (err) {
1322 				err = -err;
1323 				goto error;
1324 			}
1325 		}
1326 		if (hca_attr->flow.tunnel_header_0_1)
1327 			sh->tunnel_header_0_1 = 1;
1328 #endif
1329 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1330 		if (hca_attr->flow_hit_aso && priv->mtr_color_reg == REG_C_3) {
1331 			sh->flow_hit_aso_en = 1;
1332 			err = mlx5_flow_aso_age_mng_init(sh);
1333 			if (err) {
1334 				err = -err;
1335 				goto error;
1336 			}
1337 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1338 		}
1339 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1340 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1341 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1342 		if (hca_attr->ct_offload && priv->mtr_color_reg == REG_C_3) {
1343 			err = mlx5_flow_aso_ct_mng_init(sh);
1344 			if (err) {
1345 				err = -err;
1346 				goto error;
1347 			}
1348 			DRV_LOG(DEBUG, "CT ASO is supported.");
1349 			sh->ct_aso_en = 1;
1350 		}
1351 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1352 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1353 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1354 		    sh->config.dv_flow_en) {
1355 			priv->sampler_en = 1;
1356 			DRV_LOG(DEBUG, "Sampler enabled!");
1357 		} else {
1358 			priv->sampler_en = 0;
1359 			if (!hca_attr->log_max_ft_sampler_num)
1360 				DRV_LOG(WARNING,
1361 					"No available register for sampler.");
1362 			else
1363 				DRV_LOG(DEBUG, "DV flow is not supported!");
1364 		}
1365 #endif
1366 	}
1367 	/* Process parameters and store port configuration on priv structure. */
1368 	err = mlx5_port_args_config(priv, mkvlist, &priv->config);
1369 	if (err) {
1370 		err = rte_errno;
1371 		DRV_LOG(ERR, "Failed to process port configure: %s",
1372 			strerror(rte_errno));
1373 		goto error;
1374 	}
1375 	eth_dev = rte_eth_dev_allocate(name);
1376 	if (eth_dev == NULL) {
1377 		DRV_LOG(ERR, "can not allocate rte ethdev");
1378 		err = ENOMEM;
1379 		goto error;
1380 	}
1381 	if (priv->representor) {
1382 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1383 		eth_dev->data->representor_id = priv->representor_id;
1384 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1385 			struct mlx5_priv *opriv =
1386 				rte_eth_devices[port_id].data->dev_private;
1387 			if (opriv &&
1388 			    opriv->master &&
1389 			    opriv->domain_id == priv->domain_id &&
1390 			    opriv->sh == priv->sh) {
1391 				eth_dev->data->backer_port_id = port_id;
1392 				break;
1393 			}
1394 		}
1395 		if (port_id >= RTE_MAX_ETHPORTS)
1396 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1397 	}
1398 	priv->mp_id.port_id = eth_dev->data->port_id;
1399 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1400 	/*
1401 	 * Store associated network device interface index. This index
1402 	 * is permanent throughout the lifetime of device. So, we may store
1403 	 * the ifindex here and use the cached value further.
1404 	 */
1405 	MLX5_ASSERT(spawn->ifindex);
1406 	priv->if_index = spawn->ifindex;
1407 	priv->lag_affinity_idx = sh->refcnt - 1;
1408 	eth_dev->data->dev_private = priv;
1409 	priv->dev_data = eth_dev->data;
1410 	eth_dev->data->mac_addrs = priv->mac;
1411 	eth_dev->device = dpdk_dev;
1412 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1413 	/* Configure the first MAC address by default. */
1414 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1415 		DRV_LOG(ERR,
1416 			"port %u cannot get MAC address, is mlx5_en"
1417 			" loaded? (errno: %s)",
1418 			eth_dev->data->port_id, strerror(rte_errno));
1419 		err = ENODEV;
1420 		goto error;
1421 	}
1422 	DRV_LOG(INFO,
1423 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1424 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1425 #ifdef RTE_LIBRTE_MLX5_DEBUG
1426 	{
1427 		char ifname[MLX5_NAMESIZE];
1428 
1429 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1430 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1431 				eth_dev->data->port_id, ifname);
1432 		else
1433 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1434 				eth_dev->data->port_id);
1435 	}
1436 #endif
1437 	/* Get actual MTU if possible. */
1438 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1439 	if (err) {
1440 		err = rte_errno;
1441 		goto error;
1442 	}
1443 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1444 		priv->mtu);
1445 	/* Initialize burst functions to prevent crashes before link-up. */
1446 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1447 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1448 	eth_dev->dev_ops = &mlx5_dev_ops;
1449 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1450 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1451 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1452 	/* Register MAC address. */
1453 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1454 	if (sh->dev_cap.vf && sh->config.vf_nl_en)
1455 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1456 				      mlx5_ifindex(eth_dev),
1457 				      eth_dev->data->mac_addrs,
1458 				      MLX5_MAX_MAC_ADDRESSES);
1459 	priv->ctrl_flows = 0;
1460 	rte_spinlock_init(&priv->flow_list_lock);
1461 	TAILQ_INIT(&priv->flow_meters);
1462 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1463 	if (!priv->mtr_profile_tbl)
1464 		goto error;
1465 	/* Bring Ethernet device up. */
1466 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1467 		eth_dev->data->port_id);
1468 	mlx5_set_link_up(eth_dev);
1469 	/*
1470 	 * Even though the interrupt handler is not installed yet,
1471 	 * interrupts will still trigger on the async_fd from
1472 	 * Verbs context returned by ibv_open_device().
1473 	 */
1474 	mlx5_link_update(eth_dev, 0);
1475 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1476 		icfg[i].release_mem_en = !!sh->config.reclaim_mode;
1477 		if (sh->config.reclaim_mode)
1478 			icfg[i].per_core_cache = 0;
1479 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1480 		if (!priv->flows[i])
1481 			goto error;
1482 	}
1483 	/* Create context for virtual machine VLAN workaround. */
1484 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1485 	if (sh->config.dv_flow_en) {
1486 		err = mlx5_alloc_shared_dr(priv);
1487 		if (err)
1488 			goto error;
1489 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1490 			goto error;
1491 	}
1492 	if (mlx5_devx_obj_ops_en(sh)) {
1493 		priv->obj_ops = devx_obj_ops;
1494 		mlx5_queue_counter_id_prepare(eth_dev);
1495 		priv->obj_ops.lb_dummy_queue_create =
1496 					mlx5_rxq_ibv_obj_dummy_lb_create;
1497 		priv->obj_ops.lb_dummy_queue_release =
1498 					mlx5_rxq_ibv_obj_dummy_lb_release;
1499 	} else if (spawn->max_port > UINT8_MAX) {
1500 		/* Verbs can't support ports larger than 255 by design. */
1501 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1502 		err = ENOTSUP;
1503 		goto error;
1504 	} else {
1505 		priv->obj_ops = ibv_obj_ops;
1506 	}
1507 	if (sh->config.tx_pp &&
1508 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1509 		/*
1510 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1511 		 * packet pacing and already checked above.
1512 		 * Hence, we should only make sure the SQs will be created
1513 		 * with DevX, not with Verbs.
1514 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1515 		 * with Clock Queue UAR as required for Tx scheduling.
1516 		 */
1517 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1518 		err = ENODEV;
1519 		goto error;
1520 	}
1521 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1522 	if (!priv->drop_queue.hrxq)
1523 		goto error;
1524 	if (priv->sh->config.dv_flow_en == 2)
1525 		return eth_dev;
1526 	/* Port representor shares the same max priority with pf port. */
1527 	if (!priv->sh->flow_priority_check_flag) {
1528 		/* Supported Verbs flow priority number detection. */
1529 		err = mlx5_flow_discover_priorities(eth_dev);
1530 		priv->sh->flow_max_priority = err;
1531 		priv->sh->flow_priority_check_flag = 1;
1532 	} else {
1533 		err = priv->sh->flow_max_priority;
1534 	}
1535 	if (err < 0) {
1536 		err = -err;
1537 		goto error;
1538 	}
1539 	mlx5_set_metadata_mask(eth_dev);
1540 	if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1541 	    !priv->sh->dv_regc0_mask) {
1542 		DRV_LOG(ERR, "metadata mode %u is not supported "
1543 			     "(no metadata reg_c[0] is available)",
1544 			     sh->config.dv_xmeta_en);
1545 			err = ENOTSUP;
1546 			goto error;
1547 	}
1548 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1549 				       mlx5_hrxq_create_cb,
1550 				       mlx5_hrxq_match_cb,
1551 				       mlx5_hrxq_remove_cb,
1552 				       mlx5_hrxq_clone_cb,
1553 				       mlx5_hrxq_clone_free_cb);
1554 	if (!priv->hrxqs)
1555 		goto error;
1556 	rte_rwlock_init(&priv->ind_tbls_lock);
1557 	/* Query availability of metadata reg_c's. */
1558 	if (!priv->sh->metadata_regc_check_flag) {
1559 		err = mlx5_flow_discover_mreg_c(eth_dev);
1560 		if (err < 0) {
1561 			err = -err;
1562 			goto error;
1563 		}
1564 	}
1565 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1566 		DRV_LOG(DEBUG,
1567 			"port %u extensive metadata register is not supported",
1568 			eth_dev->data->port_id);
1569 		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1570 			DRV_LOG(ERR, "metadata mode %u is not supported "
1571 				     "(no metadata registers available)",
1572 				     sh->config.dv_xmeta_en);
1573 			err = ENOTSUP;
1574 			goto error;
1575 		}
1576 	}
1577 	if (sh->config.dv_flow_en &&
1578 	    sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1579 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1580 	    priv->sh->dv_regc0_mask) {
1581 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1582 						      MLX5_FLOW_MREG_HTABLE_SZ,
1583 						      false, true, eth_dev,
1584 						      flow_dv_mreg_create_cb,
1585 						      flow_dv_mreg_match_cb,
1586 						      flow_dv_mreg_remove_cb,
1587 						      flow_dv_mreg_clone_cb,
1588 						    flow_dv_mreg_clone_free_cb);
1589 		if (!priv->mreg_cp_tbl) {
1590 			err = ENOMEM;
1591 			goto error;
1592 		}
1593 	}
1594 	rte_spinlock_init(&priv->shared_act_sl);
1595 	mlx5_flow_counter_mode_config(eth_dev);
1596 	mlx5_flow_drop_action_config(eth_dev);
1597 	if (sh->config.dv_flow_en)
1598 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1599 	return eth_dev;
1600 error:
1601 	if (priv) {
1602 		if (priv->mreg_cp_tbl)
1603 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1604 		if (priv->sh)
1605 			mlx5_os_free_shared_dr(priv);
1606 		if (priv->nl_socket_route >= 0)
1607 			close(priv->nl_socket_route);
1608 		if (priv->vmwa_context)
1609 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1610 		if (eth_dev && priv->drop_queue.hrxq)
1611 			mlx5_drop_action_destroy(eth_dev);
1612 		if (priv->mtr_profile_tbl)
1613 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1614 		if (own_domain_id)
1615 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1616 		if (priv->hrxqs)
1617 			mlx5_list_destroy(priv->hrxqs);
1618 		if (eth_dev && priv->flex_item_map)
1619 			mlx5_flex_item_port_cleanup(eth_dev);
1620 		mlx5_free(priv);
1621 		if (eth_dev != NULL)
1622 			eth_dev->data->dev_private = NULL;
1623 	}
1624 	if (eth_dev != NULL) {
1625 		/* mac_addrs must not be freed alone because part of
1626 		 * dev_private
1627 		 **/
1628 		eth_dev->data->mac_addrs = NULL;
1629 		rte_eth_dev_release_port(eth_dev);
1630 	}
1631 	if (sh)
1632 		mlx5_free_shared_dev_ctx(sh);
1633 	if (nl_rdma >= 0)
1634 		close(nl_rdma);
1635 	MLX5_ASSERT(err > 0);
1636 	rte_errno = err;
1637 	return NULL;
1638 }
1639 
1640 /**
1641  * Comparison callback to sort device data.
1642  *
1643  * This is meant to be used with qsort().
1644  *
1645  * @param a[in]
1646  *   Pointer to pointer to first data object.
1647  * @param b[in]
1648  *   Pointer to pointer to second data object.
1649  *
1650  * @return
1651  *   0 if both objects are equal, less than 0 if the first argument is less
1652  *   than the second, greater than 0 otherwise.
1653  */
1654 static int
1655 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1656 {
1657 	const struct mlx5_switch_info *si_a =
1658 		&((const struct mlx5_dev_spawn_data *)a)->info;
1659 	const struct mlx5_switch_info *si_b =
1660 		&((const struct mlx5_dev_spawn_data *)b)->info;
1661 	int ret;
1662 
1663 	/* Master device first. */
1664 	ret = si_b->master - si_a->master;
1665 	if (ret)
1666 		return ret;
1667 	/* Then representor devices. */
1668 	ret = si_b->representor - si_a->representor;
1669 	if (ret)
1670 		return ret;
1671 	/* Unidentified devices come last in no specific order. */
1672 	if (!si_a->representor)
1673 		return 0;
1674 	/* Order representors by name. */
1675 	return si_a->port_name - si_b->port_name;
1676 }
1677 
1678 /**
1679  * Match PCI information for possible slaves of bonding device.
1680  *
1681  * @param[in] ibdev_name
1682  *   Name of Infiniband device.
1683  * @param[in] pci_dev
1684  *   Pointer to primary PCI address structure to match.
1685  * @param[in] nl_rdma
1686  *   Netlink RDMA group socket handle.
1687  * @param[in] owner
1688  *   Representor owner PF index.
1689  * @param[out] bond_info
1690  *   Pointer to bonding information.
1691  *
1692  * @return
1693  *   negative value if no bonding device found, otherwise
1694  *   positive index of slave PF in bonding.
1695  */
1696 static int
1697 mlx5_device_bond_pci_match(const char *ibdev_name,
1698 			   const struct rte_pci_addr *pci_dev,
1699 			   int nl_rdma, uint16_t owner,
1700 			   struct mlx5_bond_info *bond_info)
1701 {
1702 	char ifname[IF_NAMESIZE + 1];
1703 	unsigned int ifindex;
1704 	unsigned int np, i;
1705 	FILE *bond_file = NULL, *file;
1706 	int pf = -1;
1707 	int ret;
1708 	uint8_t cur_guid[32] = {0};
1709 	uint8_t guid[32] = {0};
1710 
1711 	/*
1712 	 * Try to get master device name. If something goes wrong suppose
1713 	 * the lack of kernel support and no bonding devices.
1714 	 */
1715 	memset(bond_info, 0, sizeof(*bond_info));
1716 	if (nl_rdma < 0)
1717 		return -1;
1718 	if (!strstr(ibdev_name, "bond"))
1719 		return -1;
1720 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1721 	if (!np)
1722 		return -1;
1723 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1724 		return -1;
1725 	/*
1726 	 * The master device might not be on the predefined port(not on port
1727 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1728 	 * device ports and find master.
1729 	 */
1730 	for (i = 1; i <= np; ++i) {
1731 		/* Check whether Infiniband port is populated. */
1732 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1733 		if (!ifindex)
1734 			continue;
1735 		if (!if_indextoname(ifindex, ifname))
1736 			continue;
1737 		/* Try to read bonding slave names from sysfs. */
1738 		MKSTR(slaves,
1739 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1740 		bond_file = fopen(slaves, "r");
1741 		if (bond_file)
1742 			break;
1743 	}
1744 	if (!bond_file)
1745 		return -1;
1746 	/* Use safe format to check maximal buffer length. */
1747 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1748 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1749 		char tmp_str[IF_NAMESIZE + 32];
1750 		struct rte_pci_addr pci_addr;
1751 		struct mlx5_switch_info	info;
1752 		int ret;
1753 
1754 		/* Process slave interface names in the loop. */
1755 		snprintf(tmp_str, sizeof(tmp_str),
1756 			 "/sys/class/net/%s", ifname);
1757 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1758 			DRV_LOG(WARNING,
1759 				"Cannot get PCI address for netdev \"%s\".",
1760 				ifname);
1761 			continue;
1762 		}
1763 		/* Slave interface PCI address match found. */
1764 		snprintf(tmp_str, sizeof(tmp_str),
1765 			 "/sys/class/net/%s/phys_port_name", ifname);
1766 		file = fopen(tmp_str, "rb");
1767 		if (!file)
1768 			break;
1769 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1770 		if (fscanf(file, "%32s", tmp_str) == 1)
1771 			mlx5_translate_port_name(tmp_str, &info);
1772 		fclose(file);
1773 		/* Only process PF ports. */
1774 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1775 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1776 			continue;
1777 		/* Check max bonding member. */
1778 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1779 			DRV_LOG(WARNING, "bonding index out of range, "
1780 				"please increase MLX5_BOND_MAX_PORTS: %s",
1781 				tmp_str);
1782 			break;
1783 		}
1784 		/* Get ifindex. */
1785 		snprintf(tmp_str, sizeof(tmp_str),
1786 			 "/sys/class/net/%s/ifindex", ifname);
1787 		file = fopen(tmp_str, "rb");
1788 		if (!file)
1789 			break;
1790 		ret = fscanf(file, "%u", &ifindex);
1791 		fclose(file);
1792 		if (ret != 1)
1793 			break;
1794 		/* Save bonding info. */
1795 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1796 			sizeof(bond_info->ports[0].ifname));
1797 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1798 		bond_info->ports[info.port_name].ifindex = ifindex;
1799 		bond_info->n_port++;
1800 		/*
1801 		 * Under socket direct mode, bonding will use
1802 		 * system_image_guid as identification.
1803 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1804 		 * All bonding members should have the same guid even if driver
1805 		 * is using PCIe BDF.
1806 		 */
1807 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1808 		if (ret < 0)
1809 			break;
1810 		else if (ret > 0) {
1811 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1812 			    owner == info.port_name &&
1813 			    (owner != 0 || (owner == 0 &&
1814 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1815 				pf = info.port_name;
1816 		} else if (pci_dev->domain == pci_addr.domain &&
1817 		    pci_dev->bus == pci_addr.bus &&
1818 		    pci_dev->devid == pci_addr.devid &&
1819 		    ((pci_dev->function == 0 &&
1820 		      pci_dev->function + owner == pci_addr.function) ||
1821 		     (pci_dev->function == owner &&
1822 		      pci_addr.function == owner)))
1823 			pf = info.port_name;
1824 	}
1825 	if (pf >= 0) {
1826 		/* Get bond interface info */
1827 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1828 					   bond_info->ifname);
1829 		if (ret)
1830 			DRV_LOG(ERR, "unable to get bond info: %s",
1831 				strerror(rte_errno));
1832 		else
1833 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1834 				ifindex, bond_info->ifindex, bond_info->ifname);
1835 	}
1836 	if (owner == 0 && pf != 0) {
1837 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
1838 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1839 				pci_dev->function);
1840 	}
1841 	return pf;
1842 }
1843 
1844 /**
1845  * Register a PCI device within bonding.
1846  *
1847  * This function spawns Ethernet devices out of a given PCI device and
1848  * bonding owner PF index.
1849  *
1850  * @param[in] cdev
1851  *   Pointer to common mlx5 device structure.
1852  * @param[in] req_eth_da
1853  *   Requested ethdev device argument.
1854  * @param[in] owner_id
1855  *   Requested owner PF port ID within bonding device, default to 0.
1856  * @param[in, out] mkvlist
1857  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1858  *
1859  * @return
1860  *   0 on success, a negative errno value otherwise and rte_errno is set.
1861  */
1862 static int
1863 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
1864 		     struct rte_eth_devargs *req_eth_da,
1865 		     uint16_t owner_id, struct mlx5_kvargs_ctrl *mkvlist)
1866 {
1867 	struct ibv_device **ibv_list;
1868 	/*
1869 	 * Number of found IB Devices matching with requested PCI BDF.
1870 	 * nd != 1 means there are multiple IB devices over the same
1871 	 * PCI device and we have representors and master.
1872 	 */
1873 	unsigned int nd = 0;
1874 	/*
1875 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1876 	 * we have the single multiport IB device, and there may be
1877 	 * representors attached to some of found ports.
1878 	 */
1879 	unsigned int np = 0;
1880 	/*
1881 	 * Number of DPDK ethernet devices to Spawn - either over
1882 	 * multiple IB devices or multiple ports of single IB device.
1883 	 * Actually this is the number of iterations to spawn.
1884 	 */
1885 	unsigned int ns = 0;
1886 	/*
1887 	 * Bonding device
1888 	 *   < 0 - no bonding device (single one)
1889 	 *  >= 0 - bonding device (value is slave PF index)
1890 	 */
1891 	int bd = -1;
1892 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
1893 	struct mlx5_dev_spawn_data *list = NULL;
1894 	struct rte_eth_devargs eth_da = *req_eth_da;
1895 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1896 	struct mlx5_bond_info bond_info;
1897 	int ret = -1;
1898 
1899 	errno = 0;
1900 	ibv_list = mlx5_glue->get_device_list(&ret);
1901 	if (!ibv_list) {
1902 		rte_errno = errno ? errno : ENOSYS;
1903 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
1904 		return -rte_errno;
1905 	}
1906 	/*
1907 	 * First scan the list of all Infiniband devices to find
1908 	 * matching ones, gathering into the list.
1909 	 */
1910 	struct ibv_device *ibv_match[ret + 1];
1911 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1912 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1913 	unsigned int i;
1914 
1915 	while (ret-- > 0) {
1916 		struct rte_pci_addr pci_addr;
1917 
1918 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
1919 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
1920 						nl_rdma, owner_id, &bond_info);
1921 		if (bd >= 0) {
1922 			/*
1923 			 * Bonding device detected. Only one match is allowed,
1924 			 * the bonding is supported over multi-port IB device,
1925 			 * there should be no matches on representor PCI
1926 			 * functions or non VF LAG bonding devices with
1927 			 * specified address.
1928 			 */
1929 			if (nd) {
1930 				DRV_LOG(ERR,
1931 					"multiple PCI match on bonding device"
1932 					"\"%s\" found", ibv_list[ret]->name);
1933 				rte_errno = ENOENT;
1934 				ret = -rte_errno;
1935 				goto exit;
1936 			}
1937 			/* Amend owner pci address if owner PF ID specified. */
1938 			if (eth_da.nb_representor_ports)
1939 				owner_pci.function += owner_id;
1940 			DRV_LOG(INFO,
1941 				"PCI information matches for slave %d bonding device \"%s\"",
1942 				bd, ibv_list[ret]->name);
1943 			ibv_match[nd++] = ibv_list[ret];
1944 			break;
1945 		} else {
1946 			/* Bonding device not found. */
1947 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
1948 					      &pci_addr))
1949 				continue;
1950 			if (owner_pci.domain != pci_addr.domain ||
1951 			    owner_pci.bus != pci_addr.bus ||
1952 			    owner_pci.devid != pci_addr.devid ||
1953 			    owner_pci.function != pci_addr.function)
1954 				continue;
1955 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1956 				ibv_list[ret]->name);
1957 			ibv_match[nd++] = ibv_list[ret];
1958 		}
1959 	}
1960 	ibv_match[nd] = NULL;
1961 	if (!nd) {
1962 		/* No device matches, just complain and bail out. */
1963 		DRV_LOG(WARNING,
1964 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
1965 			" are kernel drivers loaded?",
1966 			owner_pci.domain, owner_pci.bus,
1967 			owner_pci.devid, owner_pci.function);
1968 		rte_errno = ENOENT;
1969 		ret = -rte_errno;
1970 		goto exit;
1971 	}
1972 	if (nd == 1) {
1973 		/*
1974 		 * Found single matching device may have multiple ports.
1975 		 * Each port may be representor, we have to check the port
1976 		 * number and check the representors existence.
1977 		 */
1978 		if (nl_rdma >= 0)
1979 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1980 		if (!np)
1981 			DRV_LOG(WARNING,
1982 				"Cannot get IB device \"%s\" ports number.",
1983 				ibv_match[0]->name);
1984 		if (bd >= 0 && !np) {
1985 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
1986 			rte_errno = ENOENT;
1987 			ret = -rte_errno;
1988 			goto exit;
1989 		}
1990 	}
1991 	/* Now we can determine the maximal amount of devices to be spawned. */
1992 	list = mlx5_malloc(MLX5_MEM_ZERO,
1993 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
1994 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1995 	if (!list) {
1996 		DRV_LOG(ERR, "Spawn data array allocation failure.");
1997 		rte_errno = ENOMEM;
1998 		ret = -rte_errno;
1999 		goto exit;
2000 	}
2001 	if (bd >= 0 || np > 1) {
2002 		/*
2003 		 * Single IB device with multiple ports found,
2004 		 * it may be E-Switch master device and representors.
2005 		 * We have to perform identification through the ports.
2006 		 */
2007 		MLX5_ASSERT(nl_rdma >= 0);
2008 		MLX5_ASSERT(ns == 0);
2009 		MLX5_ASSERT(nd == 1);
2010 		MLX5_ASSERT(np);
2011 		for (i = 1; i <= np; ++i) {
2012 			list[ns].bond_info = &bond_info;
2013 			list[ns].max_port = np;
2014 			list[ns].phys_port = i;
2015 			list[ns].phys_dev_name = ibv_match[0]->name;
2016 			list[ns].eth_dev = NULL;
2017 			list[ns].pci_dev = pci_dev;
2018 			list[ns].cdev = cdev;
2019 			list[ns].pf_bond = bd;
2020 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2021 							   ibv_match[0]->name,
2022 							   i);
2023 			if (!list[ns].ifindex) {
2024 				/*
2025 				 * No network interface index found for the
2026 				 * specified port, it means there is no
2027 				 * representor on this port. It's OK,
2028 				 * there can be disabled ports, for example
2029 				 * if sriov_numvfs < sriov_totalvfs.
2030 				 */
2031 				continue;
2032 			}
2033 			ret = -1;
2034 			if (nl_route >= 0)
2035 				ret = mlx5_nl_switch_info(nl_route,
2036 							  list[ns].ifindex,
2037 							  &list[ns].info);
2038 			if (ret || (!list[ns].info.representor &&
2039 				    !list[ns].info.master)) {
2040 				/*
2041 				 * We failed to recognize representors with
2042 				 * Netlink, let's try to perform the task
2043 				 * with sysfs.
2044 				 */
2045 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2046 							     &list[ns].info);
2047 			}
2048 			if (!ret && bd >= 0) {
2049 				switch (list[ns].info.name_type) {
2050 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2051 					if (np == 1) {
2052 						/*
2053 						 * Force standalone bonding
2054 						 * device for ROCE LAG
2055 						 * configurations.
2056 						 */
2057 						list[ns].info.master = 0;
2058 						list[ns].info.representor = 0;
2059 					}
2060 					if (list[ns].info.port_name == bd)
2061 						ns++;
2062 					break;
2063 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2064 					/* Fallthrough */
2065 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2066 					/* Fallthrough */
2067 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2068 					if (list[ns].info.pf_num == bd)
2069 						ns++;
2070 					break;
2071 				default:
2072 					break;
2073 				}
2074 				continue;
2075 			}
2076 			if (!ret && (list[ns].info.representor ^
2077 				     list[ns].info.master))
2078 				ns++;
2079 		}
2080 		if (!ns) {
2081 			DRV_LOG(ERR,
2082 				"Unable to recognize master/representors on the IB device with multiple ports.");
2083 			rte_errno = ENOENT;
2084 			ret = -rte_errno;
2085 			goto exit;
2086 		}
2087 	} else {
2088 		/*
2089 		 * The existence of several matching entries (nd > 1) means
2090 		 * port representors have been instantiated. No existing Verbs
2091 		 * call nor sysfs entries can tell them apart, this can only
2092 		 * be done through Netlink calls assuming kernel drivers are
2093 		 * recent enough to support them.
2094 		 *
2095 		 * In the event of identification failure through Netlink,
2096 		 * try again through sysfs, then:
2097 		 *
2098 		 * 1. A single IB device matches (nd == 1) with single
2099 		 *    port (np=0/1) and is not a representor, assume
2100 		 *    no switch support.
2101 		 *
2102 		 * 2. Otherwise no safe assumptions can be made;
2103 		 *    complain louder and bail out.
2104 		 */
2105 		for (i = 0; i != nd; ++i) {
2106 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2107 			list[ns].bond_info = NULL;
2108 			list[ns].max_port = 1;
2109 			list[ns].phys_port = 1;
2110 			list[ns].phys_dev_name = ibv_match[i]->name;
2111 			list[ns].eth_dev = NULL;
2112 			list[ns].pci_dev = pci_dev;
2113 			list[ns].cdev = cdev;
2114 			list[ns].pf_bond = -1;
2115 			list[ns].ifindex = 0;
2116 			if (nl_rdma >= 0)
2117 				list[ns].ifindex = mlx5_nl_ifindex
2118 							    (nl_rdma,
2119 							     ibv_match[i]->name,
2120 							     1);
2121 			if (!list[ns].ifindex) {
2122 				char ifname[IF_NAMESIZE];
2123 
2124 				/*
2125 				 * Netlink failed, it may happen with old
2126 				 * ib_core kernel driver (before 4.16).
2127 				 * We can assume there is old driver because
2128 				 * here we are processing single ports IB
2129 				 * devices. Let's try sysfs to retrieve
2130 				 * the ifindex. The method works for
2131 				 * master device only.
2132 				 */
2133 				if (nd > 1) {
2134 					/*
2135 					 * Multiple devices found, assume
2136 					 * representors, can not distinguish
2137 					 * master/representor and retrieve
2138 					 * ifindex via sysfs.
2139 					 */
2140 					continue;
2141 				}
2142 				ret = mlx5_get_ifname_sysfs
2143 					(ibv_match[i]->ibdev_path, ifname);
2144 				if (!ret)
2145 					list[ns].ifindex =
2146 						if_nametoindex(ifname);
2147 				if (!list[ns].ifindex) {
2148 					/*
2149 					 * No network interface index found
2150 					 * for the specified device, it means
2151 					 * there it is neither representor
2152 					 * nor master.
2153 					 */
2154 					continue;
2155 				}
2156 			}
2157 			ret = -1;
2158 			if (nl_route >= 0)
2159 				ret = mlx5_nl_switch_info(nl_route,
2160 							  list[ns].ifindex,
2161 							  &list[ns].info);
2162 			if (ret || (!list[ns].info.representor &&
2163 				    !list[ns].info.master)) {
2164 				/*
2165 				 * We failed to recognize representors with
2166 				 * Netlink, let's try to perform the task
2167 				 * with sysfs.
2168 				 */
2169 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2170 							     &list[ns].info);
2171 			}
2172 			if (!ret && (list[ns].info.representor ^
2173 				     list[ns].info.master)) {
2174 				ns++;
2175 			} else if ((nd == 1) &&
2176 				   !list[ns].info.representor &&
2177 				   !list[ns].info.master) {
2178 				/*
2179 				 * Single IB device with one physical port and
2180 				 * attached network device.
2181 				 * May be SRIOV is not enabled or there is no
2182 				 * representors.
2183 				 */
2184 				DRV_LOG(INFO, "No E-Switch support detected.");
2185 				ns++;
2186 				break;
2187 			}
2188 		}
2189 		if (!ns) {
2190 			DRV_LOG(ERR,
2191 				"Unable to recognize master/representors on the multiple IB devices.");
2192 			rte_errno = ENOENT;
2193 			ret = -rte_errno;
2194 			goto exit;
2195 		}
2196 		/*
2197 		 * New kernels may add the switch_id attribute for the case
2198 		 * there is no E-Switch and we wrongly recognized the only
2199 		 * device as master. Override this if there is the single
2200 		 * device with single port and new device name format present.
2201 		 */
2202 		if (nd == 1 &&
2203 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2204 			list[0].info.master = 0;
2205 			list[0].info.representor = 0;
2206 		}
2207 	}
2208 	MLX5_ASSERT(ns);
2209 	/*
2210 	 * Sort list to probe devices in natural order for users convenience
2211 	 * (i.e. master first, then representors from lowest to highest ID).
2212 	 */
2213 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2214 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2215 		/* Set devargs default values. */
2216 		if (eth_da.nb_mh_controllers == 0) {
2217 			eth_da.nb_mh_controllers = 1;
2218 			eth_da.mh_controllers[0] = 0;
2219 		}
2220 		if (eth_da.nb_ports == 0 && ns > 0) {
2221 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2222 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2223 					pci_dev->device.devargs->args);
2224 			eth_da.nb_ports = 1;
2225 			eth_da.ports[0] = list[0].info.pf_num;
2226 		}
2227 		if (eth_da.nb_representor_ports == 0) {
2228 			eth_da.nb_representor_ports = 1;
2229 			eth_da.representor_ports[0] = 0;
2230 		}
2231 	}
2232 	for (i = 0; i != ns; ++i) {
2233 		uint32_t restore;
2234 
2235 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i], &eth_da,
2236 						 mkvlist);
2237 		if (!list[i].eth_dev) {
2238 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2239 				break;
2240 			/* Device is disabled or already spawned. Ignore it. */
2241 			continue;
2242 		}
2243 		restore = list[i].eth_dev->data->dev_flags;
2244 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2245 		/**
2246 		 * Each representor has a dedicated interrupts vector.
2247 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2248 		 * representor eth_dev object because representor and PF
2249 		 * share the same PCI address.
2250 		 * Override representor device with a dedicated
2251 		 * interrupts handle here.
2252 		 * Representor interrupts handle is released in mlx5_dev_stop().
2253 		 */
2254 		if (list[i].info.representor) {
2255 			struct rte_intr_handle *intr_handle =
2256 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2257 			if (intr_handle == NULL) {
2258 				DRV_LOG(ERR,
2259 					"port %u failed to allocate memory for interrupt handler "
2260 					"Rx interrupts will not be supported",
2261 					i);
2262 				rte_errno = ENOMEM;
2263 				ret = -rte_errno;
2264 				goto exit;
2265 			}
2266 			list[i].eth_dev->intr_handle = intr_handle;
2267 		}
2268 		/* Restore non-PCI flags cleared by the above call. */
2269 		list[i].eth_dev->data->dev_flags |= restore;
2270 		rte_eth_dev_probing_finish(list[i].eth_dev);
2271 	}
2272 	if (i != ns) {
2273 		DRV_LOG(ERR,
2274 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2275 			" encountering an error: %s",
2276 			owner_pci.domain, owner_pci.bus,
2277 			owner_pci.devid, owner_pci.function,
2278 			strerror(rte_errno));
2279 		ret = -rte_errno;
2280 		/* Roll back. */
2281 		while (i--) {
2282 			if (!list[i].eth_dev)
2283 				continue;
2284 			mlx5_dev_close(list[i].eth_dev);
2285 			/* mac_addrs must not be freed because in dev_private */
2286 			list[i].eth_dev->data->mac_addrs = NULL;
2287 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2288 		}
2289 		/* Restore original error. */
2290 		rte_errno = -ret;
2291 	} else {
2292 		ret = 0;
2293 	}
2294 exit:
2295 	/*
2296 	 * Do the routine cleanup:
2297 	 * - close opened Netlink sockets
2298 	 * - free allocated spawn data array
2299 	 * - free the Infiniband device list
2300 	 */
2301 	if (nl_rdma >= 0)
2302 		close(nl_rdma);
2303 	if (nl_route >= 0)
2304 		close(nl_route);
2305 	if (list)
2306 		mlx5_free(list);
2307 	MLX5_ASSERT(ibv_list);
2308 	mlx5_glue->free_device_list(ibv_list);
2309 	return ret;
2310 }
2311 
2312 static int
2313 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2314 			  struct rte_eth_devargs *eth_da)
2315 {
2316 	int ret = 0;
2317 
2318 	if (dev->devargs == NULL)
2319 		return 0;
2320 	memset(eth_da, 0, sizeof(*eth_da));
2321 	/* Parse representor information first from class argument. */
2322 	if (dev->devargs->cls_str)
2323 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2324 	if (ret != 0) {
2325 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2326 			dev->devargs->cls_str);
2327 		return -rte_errno;
2328 	}
2329 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2330 		/* Parse legacy device argument */
2331 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2332 		if (ret) {
2333 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2334 				dev->devargs->args);
2335 			return -rte_errno;
2336 		}
2337 	}
2338 	return 0;
2339 }
2340 
2341 /**
2342  * Callback to register a PCI device.
2343  *
2344  * This function spawns Ethernet devices out of a given PCI device.
2345  *
2346  * @param[in] cdev
2347  *   Pointer to common mlx5 device structure.
2348  * @param[in, out] mkvlist
2349  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2350  *
2351  * @return
2352  *   0 on success, a negative errno value otherwise and rte_errno is set.
2353  */
2354 static int
2355 mlx5_os_pci_probe(struct mlx5_common_device *cdev,
2356 		  struct mlx5_kvargs_ctrl *mkvlist)
2357 {
2358 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2359 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2360 	int ret = 0;
2361 	uint16_t p;
2362 
2363 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2364 	if (ret != 0)
2365 		return ret;
2366 
2367 	if (eth_da.nb_ports > 0) {
2368 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2369 		for (p = 0; p < eth_da.nb_ports; p++) {
2370 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2371 						   eth_da.ports[p], mkvlist);
2372 			if (ret)
2373 				break;
2374 		}
2375 		if (ret) {
2376 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2377 				"aborted due to prodding failure of PF %u",
2378 				pci_dev->addr.domain, pci_dev->addr.bus,
2379 				pci_dev->addr.devid, pci_dev->addr.function,
2380 				eth_da.ports[p]);
2381 			mlx5_net_remove(cdev);
2382 		}
2383 	} else {
2384 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0, mkvlist);
2385 	}
2386 	return ret;
2387 }
2388 
2389 /* Probe a single SF device on auxiliary bus, no representor support. */
2390 static int
2391 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev,
2392 			struct mlx5_kvargs_ctrl *mkvlist)
2393 {
2394 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2395 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2396 	struct rte_device *dev = cdev->dev;
2397 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2398 	struct rte_eth_dev *eth_dev;
2399 	int ret = 0;
2400 
2401 	/* Parse ethdev devargs. */
2402 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2403 	if (ret != 0)
2404 		return ret;
2405 	/* Init spawn data. */
2406 	spawn.max_port = 1;
2407 	spawn.phys_port = 1;
2408 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2409 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2410 	if (ret < 0) {
2411 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2412 		return ret;
2413 	}
2414 	spawn.ifindex = ret;
2415 	spawn.cdev = cdev;
2416 	/* Spawn device. */
2417 	eth_dev = mlx5_dev_spawn(dev, &spawn, &eth_da, mkvlist);
2418 	if (eth_dev == NULL)
2419 		return -rte_errno;
2420 	/* Post create. */
2421 	eth_dev->intr_handle = adev->intr_handle;
2422 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2423 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2424 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2425 		eth_dev->data->numa_node = dev->numa_node;
2426 	}
2427 	rte_eth_dev_probing_finish(eth_dev);
2428 	return 0;
2429 }
2430 
2431 /**
2432  * Net class driver callback to probe a device.
2433  *
2434  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2435  *
2436  * @param[in] cdev
2437  *   Pointer to the common mlx5 device.
2438  * @param[in, out] mkvlist
2439  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2440  *
2441  * @return
2442  *   0 on success, a negative errno value otherwise and rte_errno is set.
2443  */
2444 int
2445 mlx5_os_net_probe(struct mlx5_common_device *cdev,
2446 		  struct mlx5_kvargs_ctrl *mkvlist)
2447 {
2448 	int ret;
2449 
2450 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2451 		mlx5_pmd_socket_init();
2452 	ret = mlx5_init_once();
2453 	if (ret) {
2454 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2455 			strerror(rte_errno));
2456 		return -rte_errno;
2457 	}
2458 	ret = mlx5_probe_again_args_validate(cdev, mkvlist);
2459 	if (ret) {
2460 		DRV_LOG(ERR, "Probe again parameters are not compatible : %s",
2461 			strerror(rte_errno));
2462 		return -rte_errno;
2463 	}
2464 	if (mlx5_dev_is_pci(cdev->dev))
2465 		return mlx5_os_pci_probe(cdev, mkvlist);
2466 	else
2467 		return mlx5_os_auxiliary_probe(cdev, mkvlist);
2468 }
2469 
2470 /**
2471  * Cleanup resources when the last device is closed.
2472  */
2473 void
2474 mlx5_os_net_cleanup(void)
2475 {
2476 	mlx5_pmd_socket_uninit();
2477 }
2478 
2479 /**
2480  * Install shared asynchronous device events handler.
2481  * This function is implemented to support event sharing
2482  * between multiple ports of single IB device.
2483  *
2484  * @param sh
2485  *   Pointer to mlx5_dev_ctx_shared object.
2486  */
2487 void
2488 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2489 {
2490 	int ret;
2491 	int flags;
2492 	struct ibv_context *ctx = sh->cdev->ctx;
2493 
2494 	sh->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2495 	if (sh->intr_handle == NULL) {
2496 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2497 		rte_errno = ENOMEM;
2498 		return;
2499 	}
2500 	rte_intr_fd_set(sh->intr_handle, -1);
2501 
2502 	flags = fcntl(ctx->async_fd, F_GETFL);
2503 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2504 	if (ret) {
2505 		DRV_LOG(INFO, "failed to change file descriptor async event"
2506 			" queue");
2507 	} else {
2508 		rte_intr_fd_set(sh->intr_handle, ctx->async_fd);
2509 		rte_intr_type_set(sh->intr_handle, RTE_INTR_HANDLE_EXT);
2510 		if (rte_intr_callback_register(sh->intr_handle,
2511 					mlx5_dev_interrupt_handler, sh)) {
2512 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2513 			rte_intr_fd_set(sh->intr_handle, -1);
2514 		}
2515 	}
2516 	if (sh->cdev->config.devx) {
2517 #ifdef HAVE_IBV_DEVX_ASYNC
2518 		sh->intr_handle_devx =
2519 			rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2520 		if (!sh->intr_handle_devx) {
2521 			DRV_LOG(ERR, "Fail to allocate intr_handle");
2522 			rte_errno = ENOMEM;
2523 			return;
2524 		}
2525 		rte_intr_fd_set(sh->intr_handle_devx, -1);
2526 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2527 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2528 		if (!devx_comp) {
2529 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2530 			return;
2531 		}
2532 		flags = fcntl(devx_comp->fd, F_GETFL);
2533 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2534 		if (ret) {
2535 			DRV_LOG(INFO, "failed to change file descriptor"
2536 				" devx comp");
2537 			return;
2538 		}
2539 		rte_intr_fd_set(sh->intr_handle_devx, devx_comp->fd);
2540 		rte_intr_type_set(sh->intr_handle_devx,
2541 					 RTE_INTR_HANDLE_EXT);
2542 		if (rte_intr_callback_register(sh->intr_handle_devx,
2543 					mlx5_dev_interrupt_handler_devx, sh)) {
2544 			DRV_LOG(INFO, "Fail to install the devx shared"
2545 				" interrupt.");
2546 			rte_intr_fd_set(sh->intr_handle_devx, -1);
2547 		}
2548 #endif /* HAVE_IBV_DEVX_ASYNC */
2549 	}
2550 }
2551 
2552 /**
2553  * Uninstall shared asynchronous device events handler.
2554  * This function is implemented to support event sharing
2555  * between multiple ports of single IB device.
2556  *
2557  * @param dev
2558  *   Pointer to mlx5_dev_ctx_shared object.
2559  */
2560 void
2561 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2562 {
2563 	if (rte_intr_fd_get(sh->intr_handle) >= 0)
2564 		mlx5_intr_callback_unregister(sh->intr_handle,
2565 					      mlx5_dev_interrupt_handler, sh);
2566 	rte_intr_instance_free(sh->intr_handle);
2567 #ifdef HAVE_IBV_DEVX_ASYNC
2568 	if (rte_intr_fd_get(sh->intr_handle_devx) >= 0)
2569 		rte_intr_callback_unregister(sh->intr_handle_devx,
2570 				  mlx5_dev_interrupt_handler_devx, sh);
2571 	rte_intr_instance_free(sh->intr_handle_devx);
2572 	if (sh->devx_comp)
2573 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2574 #endif
2575 }
2576 
2577 /**
2578  * Read statistics by a named counter.
2579  *
2580  * @param[in] priv
2581  *   Pointer to the private device data structure.
2582  * @param[in] ctr_name
2583  *   Pointer to the name of the statistic counter to read
2584  * @param[out] stat
2585  *   Pointer to read statistic value.
2586  * @return
2587  *   0 on success and stat is valud, 1 if failed to read the value
2588  *   rte_errno is set.
2589  *
2590  */
2591 int
2592 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2593 		      uint64_t *stat)
2594 {
2595 	int fd;
2596 
2597 	if (priv->sh) {
2598 		if (priv->q_counters != NULL &&
2599 		    strcmp(ctr_name, "out_of_buffer") == 0)
2600 			return mlx5_devx_cmd_queue_counter_query
2601 					(priv->q_counters, 0, (uint32_t *)stat);
2602 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2603 		      priv->sh->ibdev_path,
2604 		      priv->dev_port,
2605 		      ctr_name);
2606 		fd = open(path, O_RDONLY);
2607 		/*
2608 		 * in switchdev the file location is not per port
2609 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2610 		 */
2611 		if (fd == -1) {
2612 			MKSTR(path1, "%s/hw_counters/%s",
2613 			      priv->sh->ibdev_path,
2614 			      ctr_name);
2615 			fd = open(path1, O_RDONLY);
2616 		}
2617 		if (fd != -1) {
2618 			char buf[21] = {'\0'};
2619 			ssize_t n = read(fd, buf, sizeof(buf));
2620 
2621 			close(fd);
2622 			if (n != -1) {
2623 				*stat = strtoull(buf, NULL, 10);
2624 				return 0;
2625 			}
2626 		}
2627 	}
2628 	*stat = 0;
2629 	return 1;
2630 }
2631 
2632 /**
2633  * Remove a MAC address from device
2634  *
2635  * @param dev
2636  *   Pointer to Ethernet device structure.
2637  * @param index
2638  *   MAC address index.
2639  */
2640 void
2641 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2642 {
2643 	struct mlx5_priv *priv = dev->data->dev_private;
2644 	const int vf = priv->sh->dev_cap.vf;
2645 
2646 	if (vf)
2647 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2648 					mlx5_ifindex(dev), priv->mac_own,
2649 					&dev->data->mac_addrs[index], index);
2650 }
2651 
2652 /**
2653  * Adds a MAC address to the device
2654  *
2655  * @param dev
2656  *   Pointer to Ethernet device structure.
2657  * @param mac_addr
2658  *   MAC address to register.
2659  * @param index
2660  *   MAC address index.
2661  *
2662  * @return
2663  *   0 on success, a negative errno value otherwise
2664  */
2665 int
2666 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2667 		     uint32_t index)
2668 {
2669 	struct mlx5_priv *priv = dev->data->dev_private;
2670 	const int vf = priv->sh->dev_cap.vf;
2671 	int ret = 0;
2672 
2673 	if (vf)
2674 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2675 					   mlx5_ifindex(dev), priv->mac_own,
2676 					   mac, index);
2677 	return ret;
2678 }
2679 
2680 /**
2681  * Modify a VF MAC address
2682  *
2683  * @param priv
2684  *   Pointer to device private data.
2685  * @param mac_addr
2686  *   MAC address to modify into.
2687  * @param iface_idx
2688  *   Net device interface index
2689  * @param vf_index
2690  *   VF index
2691  *
2692  * @return
2693  *   0 on success, a negative errno value otherwise
2694  */
2695 int
2696 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2697 			   unsigned int iface_idx,
2698 			   struct rte_ether_addr *mac_addr,
2699 			   int vf_index)
2700 {
2701 	return mlx5_nl_vf_mac_addr_modify
2702 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2703 }
2704 
2705 /**
2706  * Set device promiscuous mode
2707  *
2708  * @param dev
2709  *   Pointer to Ethernet device structure.
2710  * @param enable
2711  *   0 - promiscuous is disabled, otherwise - enabled
2712  *
2713  * @return
2714  *   0 on success, a negative error value otherwise
2715  */
2716 int
2717 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2718 {
2719 	struct mlx5_priv *priv = dev->data->dev_private;
2720 
2721 	return mlx5_nl_promisc(priv->nl_socket_route,
2722 			       mlx5_ifindex(dev), !!enable);
2723 }
2724 
2725 /**
2726  * Set device promiscuous mode
2727  *
2728  * @param dev
2729  *   Pointer to Ethernet device structure.
2730  * @param enable
2731  *   0 - all multicase is disabled, otherwise - enabled
2732  *
2733  * @return
2734  *   0 on success, a negative error value otherwise
2735  */
2736 int
2737 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2738 {
2739 	struct mlx5_priv *priv = dev->data->dev_private;
2740 
2741 	return mlx5_nl_allmulti(priv->nl_socket_route,
2742 				mlx5_ifindex(dev), !!enable);
2743 }
2744 
2745 /**
2746  * Flush device MAC addresses
2747  *
2748  * @param dev
2749  *   Pointer to Ethernet device structure.
2750  *
2751  */
2752 void
2753 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2754 {
2755 	struct mlx5_priv *priv = dev->data->dev_private;
2756 
2757 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2758 			       dev->data->mac_addrs,
2759 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2760 }
2761