xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision a729d2f093e9f7f13fa2e362a1a97bc2dc8f834a)
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 	/* The resources below are only valid with DV support. */
486 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
487 	/* Init port id action list. */
488 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
489 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
490 						   flow_dv_port_id_create_cb,
491 						   flow_dv_port_id_match_cb,
492 						   flow_dv_port_id_remove_cb,
493 						   flow_dv_port_id_clone_cb,
494 						 flow_dv_port_id_clone_free_cb);
495 	if (!sh->port_id_action_list)
496 		goto error;
497 	/* Init push vlan action list. */
498 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
499 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
500 						    flow_dv_push_vlan_create_cb,
501 						    flow_dv_push_vlan_match_cb,
502 						    flow_dv_push_vlan_remove_cb,
503 						    flow_dv_push_vlan_clone_cb,
504 					       flow_dv_push_vlan_clone_free_cb);
505 	if (!sh->push_vlan_action_list)
506 		goto error;
507 	/* Init sample action list. */
508 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
509 	sh->sample_action_list = mlx5_list_create(s, sh, true,
510 						  flow_dv_sample_create_cb,
511 						  flow_dv_sample_match_cb,
512 						  flow_dv_sample_remove_cb,
513 						  flow_dv_sample_clone_cb,
514 						  flow_dv_sample_clone_free_cb);
515 	if (!sh->sample_action_list)
516 		goto error;
517 	/* Init dest array action list. */
518 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
519 	sh->dest_array_list = mlx5_list_create(s, sh, true,
520 					       flow_dv_dest_array_create_cb,
521 					       flow_dv_dest_array_match_cb,
522 					       flow_dv_dest_array_remove_cb,
523 					       flow_dv_dest_array_clone_cb,
524 					      flow_dv_dest_array_clone_free_cb);
525 	if (!sh->dest_array_list)
526 		goto error;
527 	/* Init shared flex parsers list, no need lcore_share */
528 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
529 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
530 					       mlx5_flex_parser_create_cb,
531 					       mlx5_flex_parser_match_cb,
532 					       mlx5_flex_parser_remove_cb,
533 					       mlx5_flex_parser_clone_cb,
534 					       mlx5_flex_parser_clone_free_cb);
535 	if (!sh->flex_parsers_dv)
536 		goto error;
537 #endif
538 #ifdef HAVE_MLX5DV_DR
539 	void *domain;
540 
541 	/* Reference counter is zero, we should initialize structures. */
542 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
543 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
544 	if (!domain) {
545 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
546 		err = errno;
547 		goto error;
548 	}
549 	sh->rx_domain = domain;
550 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
551 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
552 	if (!domain) {
553 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
554 		err = errno;
555 		goto error;
556 	}
557 	sh->tx_domain = domain;
558 #ifdef HAVE_MLX5DV_DR_ESWITCH
559 	if (sh->config.dv_esw_en) {
560 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
561 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
562 		if (!domain) {
563 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
564 			err = errno;
565 			goto error;
566 		}
567 		sh->fdb_domain = domain;
568 	}
569 	/*
570 	 * The drop action is just some dummy placeholder in rdma-core. It
571 	 * does not belong to domains and has no any attributes, and, can be
572 	 * shared by the entire device.
573 	 */
574 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
575 	if (!sh->dr_drop_action) {
576 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
577 		err = errno;
578 		goto error;
579 	}
580 #endif
581 	if (!sh->tunnel_hub && sh->config.dv_miss_info)
582 		err = mlx5_alloc_tunnel_hub(sh);
583 	if (err) {
584 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
585 		goto error;
586 	}
587 	if (sh->config.reclaim_mode == MLX5_RCM_AGGR) {
588 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
589 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
590 		if (sh->fdb_domain)
591 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
592 	}
593 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
594 	if (!sh->config.allow_duplicate_pattern) {
595 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
596 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
597 #endif
598 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
599 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
600 		if (sh->fdb_domain)
601 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
602 	}
603 
604 	__mlx5_discovery_misc5_cap(priv);
605 #endif /* HAVE_MLX5DV_DR */
606 	sh->default_miss_action =
607 			mlx5_glue->dr_create_flow_action_default_miss();
608 	if (!sh->default_miss_action)
609 		DRV_LOG(WARNING, "Default miss action is not supported.");
610 	LIST_INIT(&sh->shared_rxqs);
611 	return 0;
612 error:
613 	/* Rollback the created objects. */
614 	if (sh->rx_domain) {
615 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
616 		sh->rx_domain = NULL;
617 	}
618 	if (sh->tx_domain) {
619 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
620 		sh->tx_domain = NULL;
621 	}
622 	if (sh->fdb_domain) {
623 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
624 		sh->fdb_domain = NULL;
625 	}
626 	if (sh->dr_drop_action) {
627 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
628 		sh->dr_drop_action = NULL;
629 	}
630 	if (sh->pop_vlan_action) {
631 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
632 		sh->pop_vlan_action = NULL;
633 	}
634 	if (sh->encaps_decaps) {
635 		mlx5_hlist_destroy(sh->encaps_decaps);
636 		sh->encaps_decaps = NULL;
637 	}
638 	if (sh->modify_cmds) {
639 		mlx5_hlist_destroy(sh->modify_cmds);
640 		sh->modify_cmds = NULL;
641 	}
642 	if (sh->tag_table) {
643 		/* tags should be destroyed with flow before. */
644 		mlx5_hlist_destroy(sh->tag_table);
645 		sh->tag_table = NULL;
646 	}
647 	if (sh->tunnel_hub) {
648 		mlx5_release_tunnel_hub(sh, priv->dev_port);
649 		sh->tunnel_hub = NULL;
650 	}
651 	mlx5_free_table_hash_list(priv);
652 	if (sh->port_id_action_list) {
653 		mlx5_list_destroy(sh->port_id_action_list);
654 		sh->port_id_action_list = NULL;
655 	}
656 	if (sh->push_vlan_action_list) {
657 		mlx5_list_destroy(sh->push_vlan_action_list);
658 		sh->push_vlan_action_list = NULL;
659 	}
660 	if (sh->sample_action_list) {
661 		mlx5_list_destroy(sh->sample_action_list);
662 		sh->sample_action_list = NULL;
663 	}
664 	if (sh->dest_array_list) {
665 		mlx5_list_destroy(sh->dest_array_list);
666 		sh->dest_array_list = NULL;
667 	}
668 	return err;
669 }
670 
671 /**
672  * Destroy DR related data within private structure.
673  *
674  * @param[in] priv
675  *   Pointer to the private device data structure.
676  */
677 void
678 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
679 {
680 	struct mlx5_dev_ctx_shared *sh = priv->sh;
681 
682 	MLX5_ASSERT(sh && sh->refcnt);
683 	if (sh->refcnt > 1)
684 		return;
685 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
686 #ifdef HAVE_MLX5DV_DR
687 	if (sh->rx_domain) {
688 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
689 		sh->rx_domain = NULL;
690 	}
691 	if (sh->tx_domain) {
692 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
693 		sh->tx_domain = NULL;
694 	}
695 #ifdef HAVE_MLX5DV_DR_ESWITCH
696 	if (sh->fdb_domain) {
697 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
698 		sh->fdb_domain = NULL;
699 	}
700 	if (sh->dr_drop_action) {
701 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
702 		sh->dr_drop_action = NULL;
703 	}
704 #endif
705 	if (sh->pop_vlan_action) {
706 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
707 		sh->pop_vlan_action = NULL;
708 	}
709 #endif /* HAVE_MLX5DV_DR */
710 	if (sh->default_miss_action)
711 		mlx5_glue->destroy_flow_action
712 				(sh->default_miss_action);
713 	if (sh->encaps_decaps) {
714 		mlx5_hlist_destroy(sh->encaps_decaps);
715 		sh->encaps_decaps = NULL;
716 	}
717 	if (sh->modify_cmds) {
718 		mlx5_hlist_destroy(sh->modify_cmds);
719 		sh->modify_cmds = NULL;
720 	}
721 	if (sh->tag_table) {
722 		/* tags should be destroyed with flow before. */
723 		mlx5_hlist_destroy(sh->tag_table);
724 		sh->tag_table = NULL;
725 	}
726 	if (sh->tunnel_hub) {
727 		mlx5_release_tunnel_hub(sh, priv->dev_port);
728 		sh->tunnel_hub = NULL;
729 	}
730 	mlx5_free_table_hash_list(priv);
731 	if (sh->port_id_action_list) {
732 		mlx5_list_destroy(sh->port_id_action_list);
733 		sh->port_id_action_list = NULL;
734 	}
735 	if (sh->push_vlan_action_list) {
736 		mlx5_list_destroy(sh->push_vlan_action_list);
737 		sh->push_vlan_action_list = NULL;
738 	}
739 	if (sh->sample_action_list) {
740 		mlx5_list_destroy(sh->sample_action_list);
741 		sh->sample_action_list = NULL;
742 	}
743 	if (sh->dest_array_list) {
744 		mlx5_list_destroy(sh->dest_array_list);
745 		sh->dest_array_list = NULL;
746 	}
747 }
748 
749 /**
750  * Initialize shared data between primary and secondary process.
751  *
752  * A memzone is reserved by primary process and secondary processes attach to
753  * the memzone.
754  *
755  * @return
756  *   0 on success, a negative errno value otherwise and rte_errno is set.
757  */
758 static int
759 mlx5_init_shared_data(void)
760 {
761 	const struct rte_memzone *mz;
762 	int ret = 0;
763 
764 	rte_spinlock_lock(&mlx5_shared_data_lock);
765 	if (mlx5_shared_data == NULL) {
766 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
767 			/* Allocate shared memory. */
768 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
769 						 sizeof(*mlx5_shared_data),
770 						 SOCKET_ID_ANY, 0);
771 			if (mz == NULL) {
772 				DRV_LOG(ERR,
773 					"Cannot allocate mlx5 shared data");
774 				ret = -rte_errno;
775 				goto error;
776 			}
777 			mlx5_shared_data = mz->addr;
778 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
779 			rte_spinlock_init(&mlx5_shared_data->lock);
780 		} else {
781 			/* Lookup allocated shared memory. */
782 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
783 			if (mz == NULL) {
784 				DRV_LOG(ERR,
785 					"Cannot attach mlx5 shared data");
786 				ret = -rte_errno;
787 				goto error;
788 			}
789 			mlx5_shared_data = mz->addr;
790 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
791 		}
792 	}
793 error:
794 	rte_spinlock_unlock(&mlx5_shared_data_lock);
795 	return ret;
796 }
797 
798 /**
799  * PMD global initialization.
800  *
801  * Independent from individual device, this function initializes global
802  * per-PMD data structures distinguishing primary and secondary processes.
803  * Hence, each initialization is called once per a process.
804  *
805  * @return
806  *   0 on success, a negative errno value otherwise and rte_errno is set.
807  */
808 static int
809 mlx5_init_once(void)
810 {
811 	struct mlx5_shared_data *sd;
812 	struct mlx5_local_data *ld = &mlx5_local_data;
813 	int ret = 0;
814 
815 	if (mlx5_init_shared_data())
816 		return -rte_errno;
817 	sd = mlx5_shared_data;
818 	MLX5_ASSERT(sd);
819 	rte_spinlock_lock(&sd->lock);
820 	switch (rte_eal_process_type()) {
821 	case RTE_PROC_PRIMARY:
822 		if (sd->init_done)
823 			break;
824 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
825 					   mlx5_mp_os_primary_handle);
826 		if (ret)
827 			goto out;
828 		sd->init_done = true;
829 		break;
830 	case RTE_PROC_SECONDARY:
831 		if (ld->init_done)
832 			break;
833 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
834 					     mlx5_mp_os_secondary_handle);
835 		if (ret)
836 			goto out;
837 		++sd->secondary_cnt;
838 		ld->init_done = true;
839 		break;
840 	default:
841 		break;
842 	}
843 out:
844 	rte_spinlock_unlock(&sd->lock);
845 	return ret;
846 }
847 
848 /**
849  * DR flow drop action support detect.
850  *
851  * @param dev
852  *   Pointer to rte_eth_dev structure.
853  *
854  */
855 static void
856 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
857 {
858 #ifdef HAVE_MLX5DV_DR
859 	struct mlx5_priv *priv = dev->data->dev_private;
860 
861 	if (!priv->sh->config.dv_flow_en || !priv->sh->dr_drop_action)
862 		return;
863 	/**
864 	 * DR supports drop action placeholder when it is supported;
865 	 * otherwise, use the queue drop action.
866 	 */
867 	if (!priv->sh->drop_action_check_flag) {
868 		if (!mlx5_flow_discover_dr_action_support(dev))
869 			priv->sh->dr_drop_action_en = 1;
870 		priv->sh->drop_action_check_flag = 1;
871 	}
872 	if (priv->sh->dr_drop_action_en)
873 		priv->root_drop_action = priv->sh->dr_drop_action;
874 	else
875 		priv->root_drop_action = priv->drop_queue.hrxq->action;
876 #endif
877 }
878 
879 static void
880 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
881 {
882 	struct mlx5_priv *priv = dev->data->dev_private;
883 	void *ctx = priv->sh->cdev->ctx;
884 
885 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
886 	if (!priv->q_counters) {
887 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
888 		struct ibv_wq *wq;
889 
890 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
891 			"by DevX - fall-back to use the kernel driver global "
892 			"queue counter.", dev->data->port_id);
893 		/* Create WQ by kernel and query its queue counter ID. */
894 		if (cq) {
895 			wq = mlx5_glue->create_wq(ctx,
896 						  &(struct ibv_wq_init_attr){
897 						    .wq_type = IBV_WQT_RQ,
898 						    .max_wr = 1,
899 						    .max_sge = 1,
900 						    .pd = priv->sh->cdev->pd,
901 						    .cq = cq,
902 						});
903 			if (wq) {
904 				/* Counter is assigned only on RDY state. */
905 				int ret = mlx5_glue->modify_wq(wq,
906 						 &(struct ibv_wq_attr){
907 						 .attr_mask = IBV_WQ_ATTR_STATE,
908 						 .wq_state = IBV_WQS_RDY,
909 						});
910 
911 				if (ret == 0)
912 					mlx5_devx_cmd_wq_query(wq,
913 							 &priv->counter_set_id);
914 				claim_zero(mlx5_glue->destroy_wq(wq));
915 			}
916 			claim_zero(mlx5_glue->destroy_cq(cq));
917 		}
918 	} else {
919 		priv->counter_set_id = priv->q_counters->id;
920 	}
921 	if (priv->counter_set_id == 0)
922 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
923 			"available.", dev->data->port_id);
924 }
925 
926 /**
927  * Check if representor spawn info match devargs.
928  *
929  * @param spawn
930  *   Verbs device parameters (name, port, switch_info) to spawn.
931  * @param eth_da
932  *   Device devargs to probe.
933  *
934  * @return
935  *   Match result.
936  */
937 static bool
938 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
939 		       struct rte_eth_devargs *eth_da)
940 {
941 	struct mlx5_switch_info *switch_info = &spawn->info;
942 	unsigned int p, f;
943 	uint16_t id;
944 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
945 						      eth_da->type);
946 
947 	switch (eth_da->type) {
948 	case RTE_ETH_REPRESENTOR_SF:
949 		if (!(spawn->info.port_name == -1 &&
950 		      switch_info->name_type ==
951 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
952 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
953 			rte_errno = EBUSY;
954 			return false;
955 		}
956 		break;
957 	case RTE_ETH_REPRESENTOR_VF:
958 		/* Allows HPF representor index -1 as exception. */
959 		if (!(spawn->info.port_name == -1 &&
960 		      switch_info->name_type ==
961 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
962 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
963 			rte_errno = EBUSY;
964 			return false;
965 		}
966 		break;
967 	case RTE_ETH_REPRESENTOR_NONE:
968 		rte_errno = EBUSY;
969 		return false;
970 	default:
971 		rte_errno = ENOTSUP;
972 		DRV_LOG(ERR, "unsupported representor type");
973 		return false;
974 	}
975 	/* Check representor ID: */
976 	for (p = 0; p < eth_da->nb_ports; ++p) {
977 		if (spawn->pf_bond < 0) {
978 			/* For non-LAG mode, allow and ignore pf. */
979 			switch_info->pf_num = eth_da->ports[p];
980 			repr_id = mlx5_representor_id_encode(switch_info,
981 							     eth_da->type);
982 		}
983 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
984 			id = MLX5_REPRESENTOR_ID
985 				(eth_da->ports[p], eth_da->type,
986 				 eth_da->representor_ports[f]);
987 			if (repr_id == id)
988 				return true;
989 		}
990 	}
991 	rte_errno = EBUSY;
992 	return false;
993 }
994 
995 /**
996  * Spawn an Ethernet device from Verbs information.
997  *
998  * @param dpdk_dev
999  *   Backing DPDK device.
1000  * @param spawn
1001  *   Verbs device parameters (name, port, switch_info) to spawn.
1002  * @param eth_da
1003  *   Device arguments.
1004  * @param mkvlist
1005  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1006  *
1007  * @return
1008  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1009  *   is set. The following errors are defined:
1010  *
1011  *   EBUSY: device is not supposed to be spawned.
1012  *   EEXIST: device is already spawned
1013  */
1014 static struct rte_eth_dev *
1015 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1016 	       struct mlx5_dev_spawn_data *spawn,
1017 	       struct rte_eth_devargs *eth_da,
1018 	       struct mlx5_kvargs_ctrl *mkvlist)
1019 {
1020 	const struct mlx5_switch_info *switch_info = &spawn->info;
1021 	struct mlx5_dev_ctx_shared *sh = NULL;
1022 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
1023 	struct rte_eth_dev *eth_dev = NULL;
1024 	struct mlx5_priv *priv = NULL;
1025 	int err = 0;
1026 	struct rte_ether_addr mac;
1027 	char name[RTE_ETH_NAME_MAX_LEN];
1028 	int own_domain_id = 0;
1029 	uint16_t port_id;
1030 	struct mlx5_port_info vport_info = { .query_flags = 0 };
1031 	int nl_rdma;
1032 	int i;
1033 
1034 	/* Determine if this port representor is supposed to be spawned. */
1035 	if (switch_info->representor && dpdk_dev->devargs &&
1036 	    !mlx5_representor_match(spawn, eth_da))
1037 		return NULL;
1038 	/* Build device name. */
1039 	if (spawn->pf_bond < 0) {
1040 		/* Single device. */
1041 		if (!switch_info->representor)
1042 			strlcpy(name, dpdk_dev->name, sizeof(name));
1043 		else
1044 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
1045 				 dpdk_dev->name,
1046 				 switch_info->name_type ==
1047 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1048 				 switch_info->port_name);
1049 	} else {
1050 		/* Bonding device. */
1051 		if (!switch_info->representor) {
1052 			err = snprintf(name, sizeof(name), "%s_%s",
1053 				       dpdk_dev->name, spawn->phys_dev_name);
1054 		} else {
1055 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1056 				dpdk_dev->name, spawn->phys_dev_name,
1057 				switch_info->ctrl_num,
1058 				switch_info->pf_num,
1059 				switch_info->name_type ==
1060 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1061 				switch_info->port_name);
1062 		}
1063 	}
1064 	if (err >= (int)sizeof(name))
1065 		DRV_LOG(WARNING, "device name overflow %s", name);
1066 	/* check if the device is already spawned */
1067 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1068 		/*
1069 		 * When device is already spawned, its devargs should be set
1070 		 * as used. otherwise, mlx5_kvargs_validate() will fail.
1071 		 */
1072 		if (mkvlist)
1073 			mlx5_port_args_set_used(name, port_id, mkvlist);
1074 		rte_errno = EEXIST;
1075 		return NULL;
1076 	}
1077 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1078 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1079 		struct mlx5_mp_id mp_id;
1080 
1081 		eth_dev = rte_eth_dev_attach_secondary(name);
1082 		if (eth_dev == NULL) {
1083 			DRV_LOG(ERR, "can not attach rte ethdev");
1084 			rte_errno = ENOMEM;
1085 			return NULL;
1086 		}
1087 		eth_dev->device = dpdk_dev;
1088 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1089 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1090 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1091 		err = mlx5_proc_priv_init(eth_dev);
1092 		if (err)
1093 			return NULL;
1094 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1095 		/* Receive command fd from primary process */
1096 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1097 		if (err < 0)
1098 			goto err_secondary;
1099 		/* Remap UAR for Tx queues. */
1100 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1101 		if (err)
1102 			goto err_secondary;
1103 		/*
1104 		 * Ethdev pointer is still required as input since
1105 		 * the primary device is not accessible from the
1106 		 * secondary process.
1107 		 */
1108 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1109 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1110 		return eth_dev;
1111 err_secondary:
1112 		mlx5_dev_close(eth_dev);
1113 		return NULL;
1114 	}
1115 	sh = mlx5_alloc_shared_dev_ctx(spawn, mkvlist);
1116 	if (!sh)
1117 		return NULL;
1118 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1119 	/* Check port status. */
1120 	if (spawn->phys_port <= UINT8_MAX) {
1121 		/* Legacy Verbs api only support u8 port number. */
1122 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1123 					    &port_attr);
1124 		if (err) {
1125 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1126 			goto error;
1127 		}
1128 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1129 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1130 			err = EINVAL;
1131 			goto error;
1132 		}
1133 	} else if (nl_rdma >= 0) {
1134 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1135 		err = mlx5_nl_port_state(nl_rdma,
1136 			spawn->phys_dev_name,
1137 			spawn->phys_port);
1138 		if (err < 0) {
1139 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1140 				strerror(rte_errno));
1141 			err = -rte_errno;
1142 			goto error;
1143 		}
1144 		port_attr.state = (enum ibv_port_state)err;
1145 	}
1146 	if (port_attr.state != IBV_PORT_ACTIVE)
1147 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1148 			mlx5_glue->port_state_str(port_attr.state),
1149 			port_attr.state);
1150 	/* Allocate private eth device data. */
1151 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1152 			   sizeof(*priv),
1153 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1154 	if (priv == NULL) {
1155 		DRV_LOG(ERR, "priv allocation failure");
1156 		err = ENOMEM;
1157 		goto error;
1158 	}
1159 	priv->sh = sh;
1160 	priv->dev_port = spawn->phys_port;
1161 	priv->pci_dev = spawn->pci_dev;
1162 	priv->mtu = RTE_ETHER_MTU;
1163 	/* Some internal functions rely on Netlink sockets, open them now. */
1164 	priv->nl_socket_rdma = nl_rdma;
1165 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1166 	priv->representor = !!switch_info->representor;
1167 	priv->master = !!switch_info->master;
1168 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1169 	priv->vport_meta_tag = 0;
1170 	priv->vport_meta_mask = 0;
1171 	priv->pf_bond = spawn->pf_bond;
1172 
1173 	DRV_LOG(DEBUG,
1174 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1175 		priv->dev_port, dpdk_dev->bus->name,
1176 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1177 		priv->master, priv->representor, priv->pf_bond);
1178 
1179 	/*
1180 	 * If we have E-Switch we should determine the vport attributes.
1181 	 * E-Switch may use either source vport field or reg_c[0] metadata
1182 	 * register to match on vport index. The engaged part of metadata
1183 	 * register is defined by mask.
1184 	 */
1185 	if (sh->esw_mode) {
1186 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1187 						 spawn->phys_port,
1188 						 &vport_info);
1189 		if (err) {
1190 			DRV_LOG(WARNING,
1191 				"Cannot query devx port %d on device %s",
1192 				spawn->phys_port, spawn->phys_dev_name);
1193 			vport_info.query_flags = 0;
1194 		}
1195 	}
1196 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1197 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1198 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1199 		if (!priv->vport_meta_mask) {
1200 			DRV_LOG(ERR,
1201 				"vport zero mask for port %d on bonding device %s",
1202 				spawn->phys_port, spawn->phys_dev_name);
1203 			err = ENOTSUP;
1204 			goto error;
1205 		}
1206 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1207 			DRV_LOG(ERR,
1208 				"Invalid vport tag for port %d on bonding device %s",
1209 				spawn->phys_port, spawn->phys_dev_name);
1210 			err = ENOTSUP;
1211 			goto error;
1212 		}
1213 	}
1214 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1215 		priv->vport_id = vport_info.vport_id;
1216 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1217 		DRV_LOG(ERR,
1218 			"Cannot deduce vport index for port %d on bonding device %s",
1219 			spawn->phys_port, spawn->phys_dev_name);
1220 		err = ENOTSUP;
1221 		goto error;
1222 	} else {
1223 		/*
1224 		 * Suppose vport index in compatible way. Kernel/rdma_core
1225 		 * support single E-Switch per PF configurations only and
1226 		 * vport_id field contains the vport index for associated VF,
1227 		 * which is deduced from representor port name.
1228 		 * For example, let's have the IB device port 10, it has
1229 		 * attached network device eth0, which has port name attribute
1230 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1231 		 * as 3 (2+1). This assigning schema should be changed if the
1232 		 * multiple E-Switch instances per PF configurations or/and PCI
1233 		 * subfunctions are added.
1234 		 */
1235 		priv->vport_id = switch_info->representor ?
1236 				 switch_info->port_name + 1 : -1;
1237 	}
1238 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1239 							  eth_da->type);
1240 	/*
1241 	 * Look for sibling devices in order to reuse their switch domain
1242 	 * if any, otherwise allocate one.
1243 	 */
1244 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1245 		const struct mlx5_priv *opriv =
1246 			rte_eth_devices[port_id].data->dev_private;
1247 
1248 		if (!opriv ||
1249 		    opriv->sh != priv->sh ||
1250 			opriv->domain_id ==
1251 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1252 			continue;
1253 		priv->domain_id = opriv->domain_id;
1254 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1255 			priv->dev_port, priv->domain_id);
1256 		break;
1257 	}
1258 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1259 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1260 		if (err) {
1261 			err = rte_errno;
1262 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1263 				strerror(rte_errno));
1264 			goto error;
1265 		}
1266 		own_domain_id = 1;
1267 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1268 			priv->dev_port, priv->domain_id);
1269 	}
1270 	if (sh->cdev->config.devx) {
1271 		struct mlx5_hca_attr *hca_attr = &sh->cdev->config.hca_attr;
1272 
1273 		sh->steering_format_version = hca_attr->steering_format_version;
1274 #if defined(HAVE_MLX5DV_DR) && \
1275 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1276 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1277 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1278 		    sh->config.dv_flow_en) {
1279 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1280 			/*
1281 			 * Meter needs two REG_C's for color match and pre-sfx
1282 			 * flow match. Here get the REG_C for color match.
1283 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1284 			 */
1285 			reg_c_mask &= 0xfc;
1286 			if (__builtin_popcount(reg_c_mask) < 1) {
1287 				priv->mtr_en = 0;
1288 				DRV_LOG(WARNING, "No available register for"
1289 					" meter.");
1290 			} else {
1291 				/*
1292 				 * The meter color register is used by the
1293 				 * flow-hit feature as well.
1294 				 * The flow-hit feature must use REG_C_3
1295 				 * Prefer REG_C_3 if it is available.
1296 				 */
1297 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1298 					priv->mtr_color_reg = REG_C_3;
1299 				else
1300 					priv->mtr_color_reg = ffs(reg_c_mask)
1301 							      - 1 + REG_C_0;
1302 				priv->mtr_en = 1;
1303 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1304 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1305 					priv->mtr_color_reg);
1306 			}
1307 		}
1308 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1309 			uint32_t log_obj_size =
1310 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1311 			if (log_obj_size >=
1312 			    hca_attr->qos.log_meter_aso_granularity &&
1313 			    log_obj_size <=
1314 			    hca_attr->qos.log_meter_aso_max_alloc)
1315 				sh->meter_aso_en = 1;
1316 		}
1317 		if (priv->mtr_en) {
1318 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1319 			if (err) {
1320 				err = -err;
1321 				goto error;
1322 			}
1323 		}
1324 		if (hca_attr->flow.tunnel_header_0_1)
1325 			sh->tunnel_header_0_1 = 1;
1326 #endif
1327 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1328 		if (hca_attr->flow_hit_aso && priv->mtr_color_reg == REG_C_3) {
1329 			sh->flow_hit_aso_en = 1;
1330 			err = mlx5_flow_aso_age_mng_init(sh);
1331 			if (err) {
1332 				err = -err;
1333 				goto error;
1334 			}
1335 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1336 		}
1337 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1338 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1339 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1340 		if (hca_attr->ct_offload && priv->mtr_color_reg == REG_C_3) {
1341 			err = mlx5_flow_aso_ct_mng_init(sh);
1342 			if (err) {
1343 				err = -err;
1344 				goto error;
1345 			}
1346 			DRV_LOG(DEBUG, "CT ASO is supported.");
1347 			sh->ct_aso_en = 1;
1348 		}
1349 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1350 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1351 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1352 		    sh->config.dv_flow_en) {
1353 			priv->sampler_en = 1;
1354 			DRV_LOG(DEBUG, "Sampler enabled!");
1355 		} else {
1356 			priv->sampler_en = 0;
1357 			if (!hca_attr->log_max_ft_sampler_num)
1358 				DRV_LOG(WARNING,
1359 					"No available register for sampler.");
1360 			else
1361 				DRV_LOG(DEBUG, "DV flow is not supported!");
1362 		}
1363 #endif
1364 	}
1365 	/* Process parameters and store port configuration on priv structure. */
1366 	err = mlx5_port_args_config(priv, mkvlist, &priv->config);
1367 	if (err) {
1368 		err = rte_errno;
1369 		DRV_LOG(ERR, "Failed to process port configure: %s",
1370 			strerror(rte_errno));
1371 		goto error;
1372 	}
1373 	eth_dev = rte_eth_dev_allocate(name);
1374 	if (eth_dev == NULL) {
1375 		DRV_LOG(ERR, "can not allocate rte ethdev");
1376 		err = ENOMEM;
1377 		goto error;
1378 	}
1379 	if (priv->representor) {
1380 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1381 		eth_dev->data->representor_id = priv->representor_id;
1382 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1383 			struct mlx5_priv *opriv =
1384 				rte_eth_devices[port_id].data->dev_private;
1385 			if (opriv &&
1386 			    opriv->master &&
1387 			    opriv->domain_id == priv->domain_id &&
1388 			    opriv->sh == priv->sh) {
1389 				eth_dev->data->backer_port_id = port_id;
1390 				break;
1391 			}
1392 		}
1393 		if (port_id >= RTE_MAX_ETHPORTS)
1394 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1395 	}
1396 	priv->mp_id.port_id = eth_dev->data->port_id;
1397 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1398 	/*
1399 	 * Store associated network device interface index. This index
1400 	 * is permanent throughout the lifetime of device. So, we may store
1401 	 * the ifindex here and use the cached value further.
1402 	 */
1403 	MLX5_ASSERT(spawn->ifindex);
1404 	priv->if_index = spawn->ifindex;
1405 	priv->lag_affinity_idx = sh->refcnt - 1;
1406 	eth_dev->data->dev_private = priv;
1407 	priv->dev_data = eth_dev->data;
1408 	eth_dev->data->mac_addrs = priv->mac;
1409 	eth_dev->device = dpdk_dev;
1410 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1411 	/* Configure the first MAC address by default. */
1412 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1413 		DRV_LOG(ERR,
1414 			"port %u cannot get MAC address, is mlx5_en"
1415 			" loaded? (errno: %s)",
1416 			eth_dev->data->port_id, strerror(rte_errno));
1417 		err = ENODEV;
1418 		goto error;
1419 	}
1420 	DRV_LOG(INFO,
1421 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1422 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1423 #ifdef RTE_LIBRTE_MLX5_DEBUG
1424 	{
1425 		char ifname[MLX5_NAMESIZE];
1426 
1427 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1428 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1429 				eth_dev->data->port_id, ifname);
1430 		else
1431 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1432 				eth_dev->data->port_id);
1433 	}
1434 #endif
1435 	/* Get actual MTU if possible. */
1436 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1437 	if (err) {
1438 		err = rte_errno;
1439 		goto error;
1440 	}
1441 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1442 		priv->mtu);
1443 	/* Initialize burst functions to prevent crashes before link-up. */
1444 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1445 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1446 	eth_dev->dev_ops = &mlx5_dev_ops;
1447 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1448 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1449 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1450 	/* Register MAC address. */
1451 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1452 	if (sh->dev_cap.vf && sh->config.vf_nl_en)
1453 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1454 				      mlx5_ifindex(eth_dev),
1455 				      eth_dev->data->mac_addrs,
1456 				      MLX5_MAX_MAC_ADDRESSES);
1457 	priv->ctrl_flows = 0;
1458 	rte_spinlock_init(&priv->flow_list_lock);
1459 	TAILQ_INIT(&priv->flow_meters);
1460 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1461 	if (!priv->mtr_profile_tbl)
1462 		goto error;
1463 	/* Bring Ethernet device up. */
1464 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1465 		eth_dev->data->port_id);
1466 	mlx5_set_link_up(eth_dev);
1467 	/*
1468 	 * Even though the interrupt handler is not installed yet,
1469 	 * interrupts will still trigger on the async_fd from
1470 	 * Verbs context returned by ibv_open_device().
1471 	 */
1472 	mlx5_link_update(eth_dev, 0);
1473 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1474 		icfg[i].release_mem_en = !!sh->config.reclaim_mode;
1475 		if (sh->config.reclaim_mode)
1476 			icfg[i].per_core_cache = 0;
1477 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1478 		if (!priv->flows[i])
1479 			goto error;
1480 	}
1481 	/* Create context for virtual machine VLAN workaround. */
1482 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1483 	if (sh->config.dv_flow_en) {
1484 		err = mlx5_alloc_shared_dr(priv);
1485 		if (err)
1486 			goto error;
1487 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1488 			goto error;
1489 	}
1490 	if (mlx5_devx_obj_ops_en(sh)) {
1491 		priv->obj_ops = devx_obj_ops;
1492 		mlx5_queue_counter_id_prepare(eth_dev);
1493 		priv->obj_ops.lb_dummy_queue_create =
1494 					mlx5_rxq_ibv_obj_dummy_lb_create;
1495 		priv->obj_ops.lb_dummy_queue_release =
1496 					mlx5_rxq_ibv_obj_dummy_lb_release;
1497 	} else if (spawn->max_port > UINT8_MAX) {
1498 		/* Verbs can't support ports larger than 255 by design. */
1499 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1500 		err = ENOTSUP;
1501 		goto error;
1502 	} else {
1503 		priv->obj_ops = ibv_obj_ops;
1504 	}
1505 	if (sh->config.tx_pp &&
1506 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1507 		/*
1508 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1509 		 * packet pacing and already checked above.
1510 		 * Hence, we should only make sure the SQs will be created
1511 		 * with DevX, not with Verbs.
1512 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1513 		 * with Clock Queue UAR as required for Tx scheduling.
1514 		 */
1515 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1516 		err = ENODEV;
1517 		goto error;
1518 	}
1519 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1520 	if (!priv->drop_queue.hrxq)
1521 		goto error;
1522 	/* Port representor shares the same max priority with pf port. */
1523 	if (!priv->sh->flow_priority_check_flag) {
1524 		/* Supported Verbs flow priority number detection. */
1525 		err = mlx5_flow_discover_priorities(eth_dev);
1526 		priv->sh->flow_max_priority = err;
1527 		priv->sh->flow_priority_check_flag = 1;
1528 	} else {
1529 		err = priv->sh->flow_max_priority;
1530 	}
1531 	if (err < 0) {
1532 		err = -err;
1533 		goto error;
1534 	}
1535 	mlx5_set_metadata_mask(eth_dev);
1536 	if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1537 	    !priv->sh->dv_regc0_mask) {
1538 		DRV_LOG(ERR, "metadata mode %u is not supported "
1539 			     "(no metadata reg_c[0] is available)",
1540 			     sh->config.dv_xmeta_en);
1541 			err = ENOTSUP;
1542 			goto error;
1543 	}
1544 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1545 				       mlx5_hrxq_create_cb,
1546 				       mlx5_hrxq_match_cb,
1547 				       mlx5_hrxq_remove_cb,
1548 				       mlx5_hrxq_clone_cb,
1549 				       mlx5_hrxq_clone_free_cb);
1550 	if (!priv->hrxqs)
1551 		goto error;
1552 	rte_rwlock_init(&priv->ind_tbls_lock);
1553 	/* Query availability of metadata reg_c's. */
1554 	if (!priv->sh->metadata_regc_check_flag) {
1555 		err = mlx5_flow_discover_mreg_c(eth_dev);
1556 		if (err < 0) {
1557 			err = -err;
1558 			goto error;
1559 		}
1560 	}
1561 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1562 		DRV_LOG(DEBUG,
1563 			"port %u extensive metadata register is not supported",
1564 			eth_dev->data->port_id);
1565 		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1566 			DRV_LOG(ERR, "metadata mode %u is not supported "
1567 				     "(no metadata registers available)",
1568 				     sh->config.dv_xmeta_en);
1569 			err = ENOTSUP;
1570 			goto error;
1571 		}
1572 	}
1573 	if (sh->config.dv_flow_en &&
1574 	    sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1575 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1576 	    priv->sh->dv_regc0_mask) {
1577 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1578 						      MLX5_FLOW_MREG_HTABLE_SZ,
1579 						      false, true, eth_dev,
1580 						      flow_dv_mreg_create_cb,
1581 						      flow_dv_mreg_match_cb,
1582 						      flow_dv_mreg_remove_cb,
1583 						      flow_dv_mreg_clone_cb,
1584 						    flow_dv_mreg_clone_free_cb);
1585 		if (!priv->mreg_cp_tbl) {
1586 			err = ENOMEM;
1587 			goto error;
1588 		}
1589 	}
1590 	rte_spinlock_init(&priv->shared_act_sl);
1591 	mlx5_flow_counter_mode_config(eth_dev);
1592 	mlx5_flow_drop_action_config(eth_dev);
1593 	if (sh->config.dv_flow_en)
1594 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1595 	return eth_dev;
1596 error:
1597 	if (priv) {
1598 		if (priv->mreg_cp_tbl)
1599 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1600 		if (priv->sh)
1601 			mlx5_os_free_shared_dr(priv);
1602 		if (priv->nl_socket_route >= 0)
1603 			close(priv->nl_socket_route);
1604 		if (priv->vmwa_context)
1605 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1606 		if (eth_dev && priv->drop_queue.hrxq)
1607 			mlx5_drop_action_destroy(eth_dev);
1608 		if (priv->mtr_profile_tbl)
1609 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1610 		if (own_domain_id)
1611 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1612 		if (priv->hrxqs)
1613 			mlx5_list_destroy(priv->hrxqs);
1614 		if (eth_dev && priv->flex_item_map)
1615 			mlx5_flex_item_port_cleanup(eth_dev);
1616 		mlx5_free(priv);
1617 		if (eth_dev != NULL)
1618 			eth_dev->data->dev_private = NULL;
1619 	}
1620 	if (eth_dev != NULL) {
1621 		/* mac_addrs must not be freed alone because part of
1622 		 * dev_private
1623 		 **/
1624 		eth_dev->data->mac_addrs = NULL;
1625 		rte_eth_dev_release_port(eth_dev);
1626 	}
1627 	if (sh)
1628 		mlx5_free_shared_dev_ctx(sh);
1629 	if (nl_rdma >= 0)
1630 		close(nl_rdma);
1631 	MLX5_ASSERT(err > 0);
1632 	rte_errno = err;
1633 	return NULL;
1634 }
1635 
1636 /**
1637  * Comparison callback to sort device data.
1638  *
1639  * This is meant to be used with qsort().
1640  *
1641  * @param a[in]
1642  *   Pointer to pointer to first data object.
1643  * @param b[in]
1644  *   Pointer to pointer to second data object.
1645  *
1646  * @return
1647  *   0 if both objects are equal, less than 0 if the first argument is less
1648  *   than the second, greater than 0 otherwise.
1649  */
1650 static int
1651 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1652 {
1653 	const struct mlx5_switch_info *si_a =
1654 		&((const struct mlx5_dev_spawn_data *)a)->info;
1655 	const struct mlx5_switch_info *si_b =
1656 		&((const struct mlx5_dev_spawn_data *)b)->info;
1657 	int ret;
1658 
1659 	/* Master device first. */
1660 	ret = si_b->master - si_a->master;
1661 	if (ret)
1662 		return ret;
1663 	/* Then representor devices. */
1664 	ret = si_b->representor - si_a->representor;
1665 	if (ret)
1666 		return ret;
1667 	/* Unidentified devices come last in no specific order. */
1668 	if (!si_a->representor)
1669 		return 0;
1670 	/* Order representors by name. */
1671 	return si_a->port_name - si_b->port_name;
1672 }
1673 
1674 /**
1675  * Match PCI information for possible slaves of bonding device.
1676  *
1677  * @param[in] ibdev_name
1678  *   Name of Infiniband device.
1679  * @param[in] pci_dev
1680  *   Pointer to primary PCI address structure to match.
1681  * @param[in] nl_rdma
1682  *   Netlink RDMA group socket handle.
1683  * @param[in] owner
1684  *   Representor owner PF index.
1685  * @param[out] bond_info
1686  *   Pointer to bonding information.
1687  *
1688  * @return
1689  *   negative value if no bonding device found, otherwise
1690  *   positive index of slave PF in bonding.
1691  */
1692 static int
1693 mlx5_device_bond_pci_match(const char *ibdev_name,
1694 			   const struct rte_pci_addr *pci_dev,
1695 			   int nl_rdma, uint16_t owner,
1696 			   struct mlx5_bond_info *bond_info)
1697 {
1698 	char ifname[IF_NAMESIZE + 1];
1699 	unsigned int ifindex;
1700 	unsigned int np, i;
1701 	FILE *bond_file = NULL, *file;
1702 	int pf = -1;
1703 	int ret;
1704 	uint8_t cur_guid[32] = {0};
1705 	uint8_t guid[32] = {0};
1706 
1707 	/*
1708 	 * Try to get master device name. If something goes wrong suppose
1709 	 * the lack of kernel support and no bonding devices.
1710 	 */
1711 	memset(bond_info, 0, sizeof(*bond_info));
1712 	if (nl_rdma < 0)
1713 		return -1;
1714 	if (!strstr(ibdev_name, "bond"))
1715 		return -1;
1716 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1717 	if (!np)
1718 		return -1;
1719 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1720 		return -1;
1721 	/*
1722 	 * The master device might not be on the predefined port(not on port
1723 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1724 	 * device ports and find master.
1725 	 */
1726 	for (i = 1; i <= np; ++i) {
1727 		/* Check whether Infiniband port is populated. */
1728 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1729 		if (!ifindex)
1730 			continue;
1731 		if (!if_indextoname(ifindex, ifname))
1732 			continue;
1733 		/* Try to read bonding slave names from sysfs. */
1734 		MKSTR(slaves,
1735 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1736 		bond_file = fopen(slaves, "r");
1737 		if (bond_file)
1738 			break;
1739 	}
1740 	if (!bond_file)
1741 		return -1;
1742 	/* Use safe format to check maximal buffer length. */
1743 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1744 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1745 		char tmp_str[IF_NAMESIZE + 32];
1746 		struct rte_pci_addr pci_addr;
1747 		struct mlx5_switch_info	info;
1748 		int ret;
1749 
1750 		/* Process slave interface names in the loop. */
1751 		snprintf(tmp_str, sizeof(tmp_str),
1752 			 "/sys/class/net/%s", ifname);
1753 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1754 			DRV_LOG(WARNING,
1755 				"Cannot get PCI address for netdev \"%s\".",
1756 				ifname);
1757 			continue;
1758 		}
1759 		/* Slave interface PCI address match found. */
1760 		snprintf(tmp_str, sizeof(tmp_str),
1761 			 "/sys/class/net/%s/phys_port_name", ifname);
1762 		file = fopen(tmp_str, "rb");
1763 		if (!file)
1764 			break;
1765 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1766 		if (fscanf(file, "%32s", tmp_str) == 1)
1767 			mlx5_translate_port_name(tmp_str, &info);
1768 		fclose(file);
1769 		/* Only process PF ports. */
1770 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1771 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1772 			continue;
1773 		/* Check max bonding member. */
1774 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1775 			DRV_LOG(WARNING, "bonding index out of range, "
1776 				"please increase MLX5_BOND_MAX_PORTS: %s",
1777 				tmp_str);
1778 			break;
1779 		}
1780 		/* Get ifindex. */
1781 		snprintf(tmp_str, sizeof(tmp_str),
1782 			 "/sys/class/net/%s/ifindex", ifname);
1783 		file = fopen(tmp_str, "rb");
1784 		if (!file)
1785 			break;
1786 		ret = fscanf(file, "%u", &ifindex);
1787 		fclose(file);
1788 		if (ret != 1)
1789 			break;
1790 		/* Save bonding info. */
1791 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1792 			sizeof(bond_info->ports[0].ifname));
1793 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1794 		bond_info->ports[info.port_name].ifindex = ifindex;
1795 		bond_info->n_port++;
1796 		/*
1797 		 * Under socket direct mode, bonding will use
1798 		 * system_image_guid as identification.
1799 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1800 		 * All bonding members should have the same guid even if driver
1801 		 * is using PCIe BDF.
1802 		 */
1803 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1804 		if (ret < 0)
1805 			break;
1806 		else if (ret > 0) {
1807 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1808 			    owner == info.port_name &&
1809 			    (owner != 0 || (owner == 0 &&
1810 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1811 				pf = info.port_name;
1812 		} else if (pci_dev->domain == pci_addr.domain &&
1813 		    pci_dev->bus == pci_addr.bus &&
1814 		    pci_dev->devid == pci_addr.devid &&
1815 		    ((pci_dev->function == 0 &&
1816 		      pci_dev->function + owner == pci_addr.function) ||
1817 		     (pci_dev->function == owner &&
1818 		      pci_addr.function == owner)))
1819 			pf = info.port_name;
1820 	}
1821 	if (pf >= 0) {
1822 		/* Get bond interface info */
1823 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1824 					   bond_info->ifname);
1825 		if (ret)
1826 			DRV_LOG(ERR, "unable to get bond info: %s",
1827 				strerror(rte_errno));
1828 		else
1829 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1830 				ifindex, bond_info->ifindex, bond_info->ifname);
1831 	}
1832 	if (owner == 0 && pf != 0) {
1833 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
1834 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1835 				pci_dev->function);
1836 	}
1837 	return pf;
1838 }
1839 
1840 /**
1841  * Register a PCI device within bonding.
1842  *
1843  * This function spawns Ethernet devices out of a given PCI device and
1844  * bonding owner PF index.
1845  *
1846  * @param[in] cdev
1847  *   Pointer to common mlx5 device structure.
1848  * @param[in] req_eth_da
1849  *   Requested ethdev device argument.
1850  * @param[in] owner_id
1851  *   Requested owner PF port ID within bonding device, default to 0.
1852  * @param[in, out] mkvlist
1853  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1854  *
1855  * @return
1856  *   0 on success, a negative errno value otherwise and rte_errno is set.
1857  */
1858 static int
1859 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
1860 		     struct rte_eth_devargs *req_eth_da,
1861 		     uint16_t owner_id, struct mlx5_kvargs_ctrl *mkvlist)
1862 {
1863 	struct ibv_device **ibv_list;
1864 	/*
1865 	 * Number of found IB Devices matching with requested PCI BDF.
1866 	 * nd != 1 means there are multiple IB devices over the same
1867 	 * PCI device and we have representors and master.
1868 	 */
1869 	unsigned int nd = 0;
1870 	/*
1871 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1872 	 * we have the single multiport IB device, and there may be
1873 	 * representors attached to some of found ports.
1874 	 */
1875 	unsigned int np = 0;
1876 	/*
1877 	 * Number of DPDK ethernet devices to Spawn - either over
1878 	 * multiple IB devices or multiple ports of single IB device.
1879 	 * Actually this is the number of iterations to spawn.
1880 	 */
1881 	unsigned int ns = 0;
1882 	/*
1883 	 * Bonding device
1884 	 *   < 0 - no bonding device (single one)
1885 	 *  >= 0 - bonding device (value is slave PF index)
1886 	 */
1887 	int bd = -1;
1888 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
1889 	struct mlx5_dev_spawn_data *list = NULL;
1890 	struct rte_eth_devargs eth_da = *req_eth_da;
1891 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1892 	struct mlx5_bond_info bond_info;
1893 	int ret = -1;
1894 
1895 	errno = 0;
1896 	ibv_list = mlx5_glue->get_device_list(&ret);
1897 	if (!ibv_list) {
1898 		rte_errno = errno ? errno : ENOSYS;
1899 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
1900 		return -rte_errno;
1901 	}
1902 	/*
1903 	 * First scan the list of all Infiniband devices to find
1904 	 * matching ones, gathering into the list.
1905 	 */
1906 	struct ibv_device *ibv_match[ret + 1];
1907 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1908 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1909 	unsigned int i;
1910 
1911 	while (ret-- > 0) {
1912 		struct rte_pci_addr pci_addr;
1913 
1914 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
1915 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
1916 						nl_rdma, owner_id, &bond_info);
1917 		if (bd >= 0) {
1918 			/*
1919 			 * Bonding device detected. Only one match is allowed,
1920 			 * the bonding is supported over multi-port IB device,
1921 			 * there should be no matches on representor PCI
1922 			 * functions or non VF LAG bonding devices with
1923 			 * specified address.
1924 			 */
1925 			if (nd) {
1926 				DRV_LOG(ERR,
1927 					"multiple PCI match on bonding device"
1928 					"\"%s\" found", ibv_list[ret]->name);
1929 				rte_errno = ENOENT;
1930 				ret = -rte_errno;
1931 				goto exit;
1932 			}
1933 			/* Amend owner pci address if owner PF ID specified. */
1934 			if (eth_da.nb_representor_ports)
1935 				owner_pci.function += owner_id;
1936 			DRV_LOG(INFO,
1937 				"PCI information matches for slave %d bonding device \"%s\"",
1938 				bd, ibv_list[ret]->name);
1939 			ibv_match[nd++] = ibv_list[ret];
1940 			break;
1941 		} else {
1942 			/* Bonding device not found. */
1943 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
1944 					      &pci_addr))
1945 				continue;
1946 			if (owner_pci.domain != pci_addr.domain ||
1947 			    owner_pci.bus != pci_addr.bus ||
1948 			    owner_pci.devid != pci_addr.devid ||
1949 			    owner_pci.function != pci_addr.function)
1950 				continue;
1951 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1952 				ibv_list[ret]->name);
1953 			ibv_match[nd++] = ibv_list[ret];
1954 		}
1955 	}
1956 	ibv_match[nd] = NULL;
1957 	if (!nd) {
1958 		/* No device matches, just complain and bail out. */
1959 		DRV_LOG(WARNING,
1960 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
1961 			" are kernel drivers loaded?",
1962 			owner_pci.domain, owner_pci.bus,
1963 			owner_pci.devid, owner_pci.function);
1964 		rte_errno = ENOENT;
1965 		ret = -rte_errno;
1966 		goto exit;
1967 	}
1968 	if (nd == 1) {
1969 		/*
1970 		 * Found single matching device may have multiple ports.
1971 		 * Each port may be representor, we have to check the port
1972 		 * number and check the representors existence.
1973 		 */
1974 		if (nl_rdma >= 0)
1975 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1976 		if (!np)
1977 			DRV_LOG(WARNING,
1978 				"Cannot get IB device \"%s\" ports number.",
1979 				ibv_match[0]->name);
1980 		if (bd >= 0 && !np) {
1981 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
1982 			rte_errno = ENOENT;
1983 			ret = -rte_errno;
1984 			goto exit;
1985 		}
1986 	}
1987 	/* Now we can determine the maximal amount of devices to be spawned. */
1988 	list = mlx5_malloc(MLX5_MEM_ZERO,
1989 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
1990 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1991 	if (!list) {
1992 		DRV_LOG(ERR, "Spawn data array allocation failure.");
1993 		rte_errno = ENOMEM;
1994 		ret = -rte_errno;
1995 		goto exit;
1996 	}
1997 	if (bd >= 0 || np > 1) {
1998 		/*
1999 		 * Single IB device with multiple ports found,
2000 		 * it may be E-Switch master device and representors.
2001 		 * We have to perform identification through the ports.
2002 		 */
2003 		MLX5_ASSERT(nl_rdma >= 0);
2004 		MLX5_ASSERT(ns == 0);
2005 		MLX5_ASSERT(nd == 1);
2006 		MLX5_ASSERT(np);
2007 		for (i = 1; i <= np; ++i) {
2008 			list[ns].bond_info = &bond_info;
2009 			list[ns].max_port = np;
2010 			list[ns].phys_port = i;
2011 			list[ns].phys_dev_name = ibv_match[0]->name;
2012 			list[ns].eth_dev = NULL;
2013 			list[ns].pci_dev = pci_dev;
2014 			list[ns].cdev = cdev;
2015 			list[ns].pf_bond = bd;
2016 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2017 							   ibv_match[0]->name,
2018 							   i);
2019 			if (!list[ns].ifindex) {
2020 				/*
2021 				 * No network interface index found for the
2022 				 * specified port, it means there is no
2023 				 * representor on this port. It's OK,
2024 				 * there can be disabled ports, for example
2025 				 * if sriov_numvfs < sriov_totalvfs.
2026 				 */
2027 				continue;
2028 			}
2029 			ret = -1;
2030 			if (nl_route >= 0)
2031 				ret = mlx5_nl_switch_info(nl_route,
2032 							  list[ns].ifindex,
2033 							  &list[ns].info);
2034 			if (ret || (!list[ns].info.representor &&
2035 				    !list[ns].info.master)) {
2036 				/*
2037 				 * We failed to recognize representors with
2038 				 * Netlink, let's try to perform the task
2039 				 * with sysfs.
2040 				 */
2041 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2042 							     &list[ns].info);
2043 			}
2044 			if (!ret && bd >= 0) {
2045 				switch (list[ns].info.name_type) {
2046 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2047 					if (np == 1) {
2048 						/*
2049 						 * Force standalone bonding
2050 						 * device for ROCE LAG
2051 						 * configurations.
2052 						 */
2053 						list[ns].info.master = 0;
2054 						list[ns].info.representor = 0;
2055 					}
2056 					if (list[ns].info.port_name == bd)
2057 						ns++;
2058 					break;
2059 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2060 					/* Fallthrough */
2061 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2062 					/* Fallthrough */
2063 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2064 					if (list[ns].info.pf_num == bd)
2065 						ns++;
2066 					break;
2067 				default:
2068 					break;
2069 				}
2070 				continue;
2071 			}
2072 			if (!ret && (list[ns].info.representor ^
2073 				     list[ns].info.master))
2074 				ns++;
2075 		}
2076 		if (!ns) {
2077 			DRV_LOG(ERR,
2078 				"Unable to recognize master/representors on the IB device with multiple ports.");
2079 			rte_errno = ENOENT;
2080 			ret = -rte_errno;
2081 			goto exit;
2082 		}
2083 	} else {
2084 		/*
2085 		 * The existence of several matching entries (nd > 1) means
2086 		 * port representors have been instantiated. No existing Verbs
2087 		 * call nor sysfs entries can tell them apart, this can only
2088 		 * be done through Netlink calls assuming kernel drivers are
2089 		 * recent enough to support them.
2090 		 *
2091 		 * In the event of identification failure through Netlink,
2092 		 * try again through sysfs, then:
2093 		 *
2094 		 * 1. A single IB device matches (nd == 1) with single
2095 		 *    port (np=0/1) and is not a representor, assume
2096 		 *    no switch support.
2097 		 *
2098 		 * 2. Otherwise no safe assumptions can be made;
2099 		 *    complain louder and bail out.
2100 		 */
2101 		for (i = 0; i != nd; ++i) {
2102 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2103 			list[ns].bond_info = NULL;
2104 			list[ns].max_port = 1;
2105 			list[ns].phys_port = 1;
2106 			list[ns].phys_dev_name = ibv_match[i]->name;
2107 			list[ns].eth_dev = NULL;
2108 			list[ns].pci_dev = pci_dev;
2109 			list[ns].cdev = cdev;
2110 			list[ns].pf_bond = -1;
2111 			list[ns].ifindex = 0;
2112 			if (nl_rdma >= 0)
2113 				list[ns].ifindex = mlx5_nl_ifindex
2114 							    (nl_rdma,
2115 							     ibv_match[i]->name,
2116 							     1);
2117 			if (!list[ns].ifindex) {
2118 				char ifname[IF_NAMESIZE];
2119 
2120 				/*
2121 				 * Netlink failed, it may happen with old
2122 				 * ib_core kernel driver (before 4.16).
2123 				 * We can assume there is old driver because
2124 				 * here we are processing single ports IB
2125 				 * devices. Let's try sysfs to retrieve
2126 				 * the ifindex. The method works for
2127 				 * master device only.
2128 				 */
2129 				if (nd > 1) {
2130 					/*
2131 					 * Multiple devices found, assume
2132 					 * representors, can not distinguish
2133 					 * master/representor and retrieve
2134 					 * ifindex via sysfs.
2135 					 */
2136 					continue;
2137 				}
2138 				ret = mlx5_get_ifname_sysfs
2139 					(ibv_match[i]->ibdev_path, ifname);
2140 				if (!ret)
2141 					list[ns].ifindex =
2142 						if_nametoindex(ifname);
2143 				if (!list[ns].ifindex) {
2144 					/*
2145 					 * No network interface index found
2146 					 * for the specified device, it means
2147 					 * there it is neither representor
2148 					 * nor master.
2149 					 */
2150 					continue;
2151 				}
2152 			}
2153 			ret = -1;
2154 			if (nl_route >= 0)
2155 				ret = mlx5_nl_switch_info(nl_route,
2156 							  list[ns].ifindex,
2157 							  &list[ns].info);
2158 			if (ret || (!list[ns].info.representor &&
2159 				    !list[ns].info.master)) {
2160 				/*
2161 				 * We failed to recognize representors with
2162 				 * Netlink, let's try to perform the task
2163 				 * with sysfs.
2164 				 */
2165 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2166 							     &list[ns].info);
2167 			}
2168 			if (!ret && (list[ns].info.representor ^
2169 				     list[ns].info.master)) {
2170 				ns++;
2171 			} else if ((nd == 1) &&
2172 				   !list[ns].info.representor &&
2173 				   !list[ns].info.master) {
2174 				/*
2175 				 * Single IB device with one physical port and
2176 				 * attached network device.
2177 				 * May be SRIOV is not enabled or there is no
2178 				 * representors.
2179 				 */
2180 				DRV_LOG(INFO, "No E-Switch support detected.");
2181 				ns++;
2182 				break;
2183 			}
2184 		}
2185 		if (!ns) {
2186 			DRV_LOG(ERR,
2187 				"Unable to recognize master/representors on the multiple IB devices.");
2188 			rte_errno = ENOENT;
2189 			ret = -rte_errno;
2190 			goto exit;
2191 		}
2192 		/*
2193 		 * New kernels may add the switch_id attribute for the case
2194 		 * there is no E-Switch and we wrongly recognized the only
2195 		 * device as master. Override this if there is the single
2196 		 * device with single port and new device name format present.
2197 		 */
2198 		if (nd == 1 &&
2199 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2200 			list[0].info.master = 0;
2201 			list[0].info.representor = 0;
2202 		}
2203 	}
2204 	MLX5_ASSERT(ns);
2205 	/*
2206 	 * Sort list to probe devices in natural order for users convenience
2207 	 * (i.e. master first, then representors from lowest to highest ID).
2208 	 */
2209 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2210 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2211 		/* Set devargs default values. */
2212 		if (eth_da.nb_mh_controllers == 0) {
2213 			eth_da.nb_mh_controllers = 1;
2214 			eth_da.mh_controllers[0] = 0;
2215 		}
2216 		if (eth_da.nb_ports == 0 && ns > 0) {
2217 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2218 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2219 					pci_dev->device.devargs->args);
2220 			eth_da.nb_ports = 1;
2221 			eth_da.ports[0] = list[0].info.pf_num;
2222 		}
2223 		if (eth_da.nb_representor_ports == 0) {
2224 			eth_da.nb_representor_ports = 1;
2225 			eth_da.representor_ports[0] = 0;
2226 		}
2227 	}
2228 	for (i = 0; i != ns; ++i) {
2229 		uint32_t restore;
2230 
2231 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i], &eth_da,
2232 						 mkvlist);
2233 		if (!list[i].eth_dev) {
2234 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2235 				break;
2236 			/* Device is disabled or already spawned. Ignore it. */
2237 			continue;
2238 		}
2239 		restore = list[i].eth_dev->data->dev_flags;
2240 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2241 		/**
2242 		 * Each representor has a dedicated interrupts vector.
2243 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2244 		 * representor eth_dev object because representor and PF
2245 		 * share the same PCI address.
2246 		 * Override representor device with a dedicated
2247 		 * interrupts handle here.
2248 		 * Representor interrupts handle is released in mlx5_dev_stop().
2249 		 */
2250 		if (list[i].info.representor) {
2251 			struct rte_intr_handle *intr_handle =
2252 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2253 			if (intr_handle == NULL) {
2254 				DRV_LOG(ERR,
2255 					"port %u failed to allocate memory for interrupt handler "
2256 					"Rx interrupts will not be supported",
2257 					i);
2258 				rte_errno = ENOMEM;
2259 				ret = -rte_errno;
2260 				goto exit;
2261 			}
2262 			list[i].eth_dev->intr_handle = intr_handle;
2263 		}
2264 		/* Restore non-PCI flags cleared by the above call. */
2265 		list[i].eth_dev->data->dev_flags |= restore;
2266 		rte_eth_dev_probing_finish(list[i].eth_dev);
2267 	}
2268 	if (i != ns) {
2269 		DRV_LOG(ERR,
2270 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2271 			" encountering an error: %s",
2272 			owner_pci.domain, owner_pci.bus,
2273 			owner_pci.devid, owner_pci.function,
2274 			strerror(rte_errno));
2275 		ret = -rte_errno;
2276 		/* Roll back. */
2277 		while (i--) {
2278 			if (!list[i].eth_dev)
2279 				continue;
2280 			mlx5_dev_close(list[i].eth_dev);
2281 			/* mac_addrs must not be freed because in dev_private */
2282 			list[i].eth_dev->data->mac_addrs = NULL;
2283 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2284 		}
2285 		/* Restore original error. */
2286 		rte_errno = -ret;
2287 	} else {
2288 		ret = 0;
2289 	}
2290 exit:
2291 	/*
2292 	 * Do the routine cleanup:
2293 	 * - close opened Netlink sockets
2294 	 * - free allocated spawn data array
2295 	 * - free the Infiniband device list
2296 	 */
2297 	if (nl_rdma >= 0)
2298 		close(nl_rdma);
2299 	if (nl_route >= 0)
2300 		close(nl_route);
2301 	if (list)
2302 		mlx5_free(list);
2303 	MLX5_ASSERT(ibv_list);
2304 	mlx5_glue->free_device_list(ibv_list);
2305 	return ret;
2306 }
2307 
2308 static int
2309 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2310 			  struct rte_eth_devargs *eth_da)
2311 {
2312 	int ret = 0;
2313 
2314 	if (dev->devargs == NULL)
2315 		return 0;
2316 	memset(eth_da, 0, sizeof(*eth_da));
2317 	/* Parse representor information first from class argument. */
2318 	if (dev->devargs->cls_str)
2319 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2320 	if (ret != 0) {
2321 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2322 			dev->devargs->cls_str);
2323 		return -rte_errno;
2324 	}
2325 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2326 		/* Parse legacy device argument */
2327 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2328 		if (ret) {
2329 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2330 				dev->devargs->args);
2331 			return -rte_errno;
2332 		}
2333 	}
2334 	return 0;
2335 }
2336 
2337 /**
2338  * Callback to register a PCI device.
2339  *
2340  * This function spawns Ethernet devices out of a given PCI device.
2341  *
2342  * @param[in] cdev
2343  *   Pointer to common mlx5 device structure.
2344  * @param[in, out] mkvlist
2345  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2346  *
2347  * @return
2348  *   0 on success, a negative errno value otherwise and rte_errno is set.
2349  */
2350 static int
2351 mlx5_os_pci_probe(struct mlx5_common_device *cdev,
2352 		  struct mlx5_kvargs_ctrl *mkvlist)
2353 {
2354 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2355 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2356 	int ret = 0;
2357 	uint16_t p;
2358 
2359 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2360 	if (ret != 0)
2361 		return ret;
2362 
2363 	if (eth_da.nb_ports > 0) {
2364 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2365 		for (p = 0; p < eth_da.nb_ports; p++) {
2366 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2367 						   eth_da.ports[p], mkvlist);
2368 			if (ret)
2369 				break;
2370 		}
2371 		if (ret) {
2372 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2373 				"aborted due to prodding failure of PF %u",
2374 				pci_dev->addr.domain, pci_dev->addr.bus,
2375 				pci_dev->addr.devid, pci_dev->addr.function,
2376 				eth_da.ports[p]);
2377 			mlx5_net_remove(cdev);
2378 		}
2379 	} else {
2380 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0, mkvlist);
2381 	}
2382 	return ret;
2383 }
2384 
2385 /* Probe a single SF device on auxiliary bus, no representor support. */
2386 static int
2387 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev,
2388 			struct mlx5_kvargs_ctrl *mkvlist)
2389 {
2390 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2391 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2392 	struct rte_device *dev = cdev->dev;
2393 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2394 	struct rte_eth_dev *eth_dev;
2395 	int ret = 0;
2396 
2397 	/* Parse ethdev devargs. */
2398 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2399 	if (ret != 0)
2400 		return ret;
2401 	/* Init spawn data. */
2402 	spawn.max_port = 1;
2403 	spawn.phys_port = 1;
2404 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2405 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2406 	if (ret < 0) {
2407 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2408 		return ret;
2409 	}
2410 	spawn.ifindex = ret;
2411 	spawn.cdev = cdev;
2412 	/* Spawn device. */
2413 	eth_dev = mlx5_dev_spawn(dev, &spawn, &eth_da, mkvlist);
2414 	if (eth_dev == NULL)
2415 		return -rte_errno;
2416 	/* Post create. */
2417 	eth_dev->intr_handle = adev->intr_handle;
2418 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2419 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2420 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2421 		eth_dev->data->numa_node = dev->numa_node;
2422 	}
2423 	rte_eth_dev_probing_finish(eth_dev);
2424 	return 0;
2425 }
2426 
2427 /**
2428  * Net class driver callback to probe a device.
2429  *
2430  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2431  *
2432  * @param[in] cdev
2433  *   Pointer to the common mlx5 device.
2434  * @param[in, out] mkvlist
2435  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2436  *
2437  * @return
2438  *   0 on success, a negative errno value otherwise and rte_errno is set.
2439  */
2440 int
2441 mlx5_os_net_probe(struct mlx5_common_device *cdev,
2442 		  struct mlx5_kvargs_ctrl *mkvlist)
2443 {
2444 	int ret;
2445 
2446 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2447 		mlx5_pmd_socket_init();
2448 	ret = mlx5_init_once();
2449 	if (ret) {
2450 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2451 			strerror(rte_errno));
2452 		return -rte_errno;
2453 	}
2454 	ret = mlx5_probe_again_args_validate(cdev, mkvlist);
2455 	if (ret) {
2456 		DRV_LOG(ERR, "Probe again parameters are not compatible : %s",
2457 			strerror(rte_errno));
2458 		return -rte_errno;
2459 	}
2460 	if (mlx5_dev_is_pci(cdev->dev))
2461 		return mlx5_os_pci_probe(cdev, mkvlist);
2462 	else
2463 		return mlx5_os_auxiliary_probe(cdev, mkvlist);
2464 }
2465 
2466 /**
2467  * Cleanup resources when the last device is closed.
2468  */
2469 void
2470 mlx5_os_net_cleanup(void)
2471 {
2472 	mlx5_pmd_socket_uninit();
2473 }
2474 
2475 /**
2476  * Install shared asynchronous device events handler.
2477  * This function is implemented to support event sharing
2478  * between multiple ports of single IB device.
2479  *
2480  * @param sh
2481  *   Pointer to mlx5_dev_ctx_shared object.
2482  */
2483 void
2484 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2485 {
2486 	int ret;
2487 	int flags;
2488 	struct ibv_context *ctx = sh->cdev->ctx;
2489 
2490 	sh->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2491 	if (sh->intr_handle == NULL) {
2492 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2493 		rte_errno = ENOMEM;
2494 		return;
2495 	}
2496 	rte_intr_fd_set(sh->intr_handle, -1);
2497 
2498 	flags = fcntl(ctx->async_fd, F_GETFL);
2499 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2500 	if (ret) {
2501 		DRV_LOG(INFO, "failed to change file descriptor async event"
2502 			" queue");
2503 	} else {
2504 		rte_intr_fd_set(sh->intr_handle, ctx->async_fd);
2505 		rte_intr_type_set(sh->intr_handle, RTE_INTR_HANDLE_EXT);
2506 		if (rte_intr_callback_register(sh->intr_handle,
2507 					mlx5_dev_interrupt_handler, sh)) {
2508 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2509 			rte_intr_fd_set(sh->intr_handle, -1);
2510 		}
2511 	}
2512 	if (sh->cdev->config.devx) {
2513 #ifdef HAVE_IBV_DEVX_ASYNC
2514 		sh->intr_handle_devx =
2515 			rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2516 		if (!sh->intr_handle_devx) {
2517 			DRV_LOG(ERR, "Fail to allocate intr_handle");
2518 			rte_errno = ENOMEM;
2519 			return;
2520 		}
2521 		rte_intr_fd_set(sh->intr_handle_devx, -1);
2522 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2523 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2524 		if (!devx_comp) {
2525 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2526 			return;
2527 		}
2528 		flags = fcntl(devx_comp->fd, F_GETFL);
2529 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2530 		if (ret) {
2531 			DRV_LOG(INFO, "failed to change file descriptor"
2532 				" devx comp");
2533 			return;
2534 		}
2535 		rte_intr_fd_set(sh->intr_handle_devx, devx_comp->fd);
2536 		rte_intr_type_set(sh->intr_handle_devx,
2537 					 RTE_INTR_HANDLE_EXT);
2538 		if (rte_intr_callback_register(sh->intr_handle_devx,
2539 					mlx5_dev_interrupt_handler_devx, sh)) {
2540 			DRV_LOG(INFO, "Fail to install the devx shared"
2541 				" interrupt.");
2542 			rte_intr_fd_set(sh->intr_handle_devx, -1);
2543 		}
2544 #endif /* HAVE_IBV_DEVX_ASYNC */
2545 	}
2546 }
2547 
2548 /**
2549  * Uninstall shared asynchronous device events handler.
2550  * This function is implemented to support event sharing
2551  * between multiple ports of single IB device.
2552  *
2553  * @param dev
2554  *   Pointer to mlx5_dev_ctx_shared object.
2555  */
2556 void
2557 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2558 {
2559 	if (rte_intr_fd_get(sh->intr_handle) >= 0)
2560 		mlx5_intr_callback_unregister(sh->intr_handle,
2561 					      mlx5_dev_interrupt_handler, sh);
2562 	rte_intr_instance_free(sh->intr_handle);
2563 #ifdef HAVE_IBV_DEVX_ASYNC
2564 	if (rte_intr_fd_get(sh->intr_handle_devx) >= 0)
2565 		rte_intr_callback_unregister(sh->intr_handle_devx,
2566 				  mlx5_dev_interrupt_handler_devx, sh);
2567 	rte_intr_instance_free(sh->intr_handle_devx);
2568 	if (sh->devx_comp)
2569 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2570 #endif
2571 }
2572 
2573 /**
2574  * Read statistics by a named counter.
2575  *
2576  * @param[in] priv
2577  *   Pointer to the private device data structure.
2578  * @param[in] ctr_name
2579  *   Pointer to the name of the statistic counter to read
2580  * @param[out] stat
2581  *   Pointer to read statistic value.
2582  * @return
2583  *   0 on success and stat is valud, 1 if failed to read the value
2584  *   rte_errno is set.
2585  *
2586  */
2587 int
2588 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2589 		      uint64_t *stat)
2590 {
2591 	int fd;
2592 
2593 	if (priv->sh) {
2594 		if (priv->q_counters != NULL &&
2595 		    strcmp(ctr_name, "out_of_buffer") == 0)
2596 			return mlx5_devx_cmd_queue_counter_query
2597 					(priv->q_counters, 0, (uint32_t *)stat);
2598 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2599 		      priv->sh->ibdev_path,
2600 		      priv->dev_port,
2601 		      ctr_name);
2602 		fd = open(path, O_RDONLY);
2603 		/*
2604 		 * in switchdev the file location is not per port
2605 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2606 		 */
2607 		if (fd == -1) {
2608 			MKSTR(path1, "%s/hw_counters/%s",
2609 			      priv->sh->ibdev_path,
2610 			      ctr_name);
2611 			fd = open(path1, O_RDONLY);
2612 		}
2613 		if (fd != -1) {
2614 			char buf[21] = {'\0'};
2615 			ssize_t n = read(fd, buf, sizeof(buf));
2616 
2617 			close(fd);
2618 			if (n != -1) {
2619 				*stat = strtoull(buf, NULL, 10);
2620 				return 0;
2621 			}
2622 		}
2623 	}
2624 	*stat = 0;
2625 	return 1;
2626 }
2627 
2628 /**
2629  * Remove a MAC address from device
2630  *
2631  * @param dev
2632  *   Pointer to Ethernet device structure.
2633  * @param index
2634  *   MAC address index.
2635  */
2636 void
2637 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2638 {
2639 	struct mlx5_priv *priv = dev->data->dev_private;
2640 	const int vf = priv->sh->dev_cap.vf;
2641 
2642 	if (vf)
2643 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2644 					mlx5_ifindex(dev), priv->mac_own,
2645 					&dev->data->mac_addrs[index], index);
2646 }
2647 
2648 /**
2649  * Adds a MAC address to the device
2650  *
2651  * @param dev
2652  *   Pointer to Ethernet device structure.
2653  * @param mac_addr
2654  *   MAC address to register.
2655  * @param index
2656  *   MAC address index.
2657  *
2658  * @return
2659  *   0 on success, a negative errno value otherwise
2660  */
2661 int
2662 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2663 		     uint32_t index)
2664 {
2665 	struct mlx5_priv *priv = dev->data->dev_private;
2666 	const int vf = priv->sh->dev_cap.vf;
2667 	int ret = 0;
2668 
2669 	if (vf)
2670 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2671 					   mlx5_ifindex(dev), priv->mac_own,
2672 					   mac, index);
2673 	return ret;
2674 }
2675 
2676 /**
2677  * Modify a VF MAC address
2678  *
2679  * @param priv
2680  *   Pointer to device private data.
2681  * @param mac_addr
2682  *   MAC address to modify into.
2683  * @param iface_idx
2684  *   Net device interface index
2685  * @param vf_index
2686  *   VF index
2687  *
2688  * @return
2689  *   0 on success, a negative errno value otherwise
2690  */
2691 int
2692 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2693 			   unsigned int iface_idx,
2694 			   struct rte_ether_addr *mac_addr,
2695 			   int vf_index)
2696 {
2697 	return mlx5_nl_vf_mac_addr_modify
2698 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2699 }
2700 
2701 /**
2702  * Set device promiscuous mode
2703  *
2704  * @param dev
2705  *   Pointer to Ethernet device structure.
2706  * @param enable
2707  *   0 - promiscuous is disabled, otherwise - enabled
2708  *
2709  * @return
2710  *   0 on success, a negative error value otherwise
2711  */
2712 int
2713 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2714 {
2715 	struct mlx5_priv *priv = dev->data->dev_private;
2716 
2717 	return mlx5_nl_promisc(priv->nl_socket_route,
2718 			       mlx5_ifindex(dev), !!enable);
2719 }
2720 
2721 /**
2722  * Set device promiscuous mode
2723  *
2724  * @param dev
2725  *   Pointer to Ethernet device structure.
2726  * @param enable
2727  *   0 - all multicase is disabled, otherwise - enabled
2728  *
2729  * @return
2730  *   0 on success, a negative error value otherwise
2731  */
2732 int
2733 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2734 {
2735 	struct mlx5_priv *priv = dev->data->dev_private;
2736 
2737 	return mlx5_nl_allmulti(priv->nl_socket_route,
2738 				mlx5_ifindex(dev), !!enable);
2739 }
2740 
2741 /**
2742  * Flush device MAC addresses
2743  *
2744  * @param dev
2745  *   Pointer to Ethernet device structure.
2746  *
2747  */
2748 void
2749 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2750 {
2751 	struct mlx5_priv *priv = dev->data->dev_private;
2752 
2753 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2754 			       dev->data->mac_addrs,
2755 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2756 }
2757