xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 87d396163c005deb8d9f72ec0977f19e5edd8f47)
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 <bus_driver.h>
23 #include <bus_pci_driver.h>
24 #include <bus_auxiliary_driver.h>
25 #include <rte_common.h>
26 #include <rte_kvargs.h>
27 #include <rte_rwlock.h>
28 #include <rte_spinlock.h>
29 #include <rte_string_fns.h>
30 #include <rte_alarm.h>
31 #include <rte_eal_paging.h>
32 
33 #include <mlx5_glue.h>
34 #include <mlx5_devx_cmds.h>
35 #include <mlx5_common.h>
36 #include <mlx5_common_mp.h>
37 #include <mlx5_common_mr.h>
38 #include <mlx5_malloc.h>
39 
40 #include "mlx5_defs.h"
41 #include "mlx5.h"
42 #include "mlx5_common_os.h"
43 #include "mlx5_utils.h"
44 #include "mlx5_rxtx.h"
45 #include "mlx5_rx.h"
46 #include "mlx5_tx.h"
47 #include "mlx5_autoconf.h"
48 #include "mlx5_flow.h"
49 #include "rte_pmd_mlx5.h"
50 #include "mlx5_verbs.h"
51 #include "mlx5_nl.h"
52 #include "mlx5_devx.h"
53 
54 #ifndef HAVE_IBV_MLX5_MOD_MPW
55 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
56 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
57 #endif
58 
59 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
60 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
61 #endif
62 
63 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
64 
65 /* Spinlock for mlx5_shared_data allocation. */
66 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
67 
68 /* Process local data for secondary processes. */
69 static struct mlx5_local_data mlx5_local_data;
70 
71 /* rte flow indexed pool configuration. */
72 static struct mlx5_indexed_pool_config icfg[] = {
73 	{
74 		.size = sizeof(struct rte_flow),
75 		.trunk_size = 64,
76 		.need_lock = 1,
77 		.release_mem_en = 0,
78 		.malloc = mlx5_malloc,
79 		.free = mlx5_free,
80 		.per_core_cache = 0,
81 		.type = "ctl_flow_ipool",
82 	},
83 	{
84 		.size = sizeof(struct rte_flow),
85 		.trunk_size = 64,
86 		.grow_trunk = 3,
87 		.grow_shift = 2,
88 		.need_lock = 1,
89 		.release_mem_en = 0,
90 		.malloc = mlx5_malloc,
91 		.free = mlx5_free,
92 		.per_core_cache = 1 << 14,
93 		.type = "rte_flow_ipool",
94 	},
95 	{
96 		.size = sizeof(struct rte_flow),
97 		.trunk_size = 64,
98 		.grow_trunk = 3,
99 		.grow_shift = 2,
100 		.need_lock = 1,
101 		.release_mem_en = 0,
102 		.malloc = mlx5_malloc,
103 		.free = mlx5_free,
104 		.per_core_cache = 0,
105 		.type = "mcp_flow_ipool",
106 	},
107 };
108 
109 /**
110  * Set the completion channel file descriptor interrupt as non-blocking.
111  *
112  * @param[in] rxq_obj
113  *   Pointer to RQ channel object, which includes the channel fd
114  *
115  * @param[out] fd
116  *   The file descriptor (representing the interrupt) used in this channel.
117  *
118  * @return
119  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
120  */
121 int
122 mlx5_os_set_nonblock_channel_fd(int fd)
123 {
124 	int flags;
125 
126 	flags = fcntl(fd, F_GETFL);
127 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
128 }
129 
130 /**
131  * Get mlx5 device attributes. The glue function query_device_ex() is called
132  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
133  * device attributes from the glue out parameter.
134  *
135  * @param sh
136  *   Pointer to shared device context.
137  *
138  * @return
139  *   0 on success, a negative errno value otherwise and rte_errno is set.
140  */
141 int
142 mlx5_os_capabilities_prepare(struct mlx5_dev_ctx_shared *sh)
143 {
144 	int err;
145 	struct mlx5_common_device *cdev = sh->cdev;
146 	struct mlx5_hca_attr *hca_attr = &cdev->config.hca_attr;
147 	struct ibv_device_attr_ex attr_ex = { .comp_mask = 0 };
148 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
149 
150 	err = mlx5_glue->query_device_ex(cdev->ctx, NULL, &attr_ex);
151 	if (err) {
152 		rte_errno = errno;
153 		return -rte_errno;
154 	}
155 #ifdef HAVE_IBV_MLX5_MOD_SWP
156 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
157 #endif
158 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
159 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
160 #endif
161 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
162 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
163 #endif
164 	err = mlx5_glue->dv_query_device(cdev->ctx, &dv_attr);
165 	if (err) {
166 		rte_errno = errno;
167 		return -rte_errno;
168 	}
169 	memset(&sh->dev_cap, 0, sizeof(struct mlx5_dev_cap));
170 	if (mlx5_dev_is_pci(cdev->dev))
171 		sh->dev_cap.vf = mlx5_dev_is_vf_pci(RTE_DEV_TO_PCI(cdev->dev));
172 	else
173 		sh->dev_cap.sf = 1;
174 	sh->dev_cap.max_qp_wr = attr_ex.orig_attr.max_qp_wr;
175 	sh->dev_cap.max_sge = attr_ex.orig_attr.max_sge;
176 	sh->dev_cap.max_cq = attr_ex.orig_attr.max_cq;
177 	sh->dev_cap.max_qp = attr_ex.orig_attr.max_qp;
178 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
179 	sh->dev_cap.dest_tir = 1;
180 #endif
181 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) && defined(HAVE_MLX5DV_DR)
182 	DRV_LOG(DEBUG, "DV flow is supported.");
183 	sh->dev_cap.dv_flow_en = 1;
184 #endif
185 #ifdef HAVE_MLX5DV_DR_ESWITCH
186 	if (hca_attr->eswitch_manager && sh->dev_cap.dv_flow_en && sh->esw_mode)
187 		sh->dev_cap.dv_esw_en = 1;
188 #endif
189 	/*
190 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
191 	 * as all ConnectX-5 devices.
192 	 */
193 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
194 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
195 			DRV_LOG(DEBUG, "Enhanced MPW is supported.");
196 			sh->dev_cap.mps = MLX5_MPW_ENHANCED;
197 		} else {
198 			DRV_LOG(DEBUG, "MPW is supported.");
199 			sh->dev_cap.mps = MLX5_MPW;
200 		}
201 	} else {
202 		DRV_LOG(DEBUG, "MPW isn't supported.");
203 		sh->dev_cap.mps = MLX5_MPW_DISABLED;
204 	}
205 #if (RTE_CACHE_LINE_SIZE == 128)
206 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)
207 		sh->dev_cap.cqe_comp = 1;
208 	DRV_LOG(DEBUG, "Rx CQE 128B compression is %ssupported.",
209 		sh->dev_cap.cqe_comp ? "" : "not ");
210 #else
211 	sh->dev_cap.cqe_comp = 1;
212 #endif
213 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
214 	sh->dev_cap.mpls_en =
215 		((dv_attr.tunnel_offloads_caps &
216 		  MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
217 		 (dv_attr.tunnel_offloads_caps &
218 		  MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
219 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported.",
220 		sh->dev_cap.mpls_en ? "" : "not ");
221 #else
222 	DRV_LOG(WARNING,
223 		"MPLS over GRE/UDP tunnel offloading disabled due to old OFED/rdma-core version or firmware configuration");
224 #endif
225 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
226 	sh->dev_cap.hw_padding = !!attr_ex.rx_pad_end_addr_align;
227 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
228 	sh->dev_cap.hw_padding = !!(attr_ex.device_cap_flags_ex &
229 				    IBV_DEVICE_PCI_WRITE_END_PADDING);
230 #endif
231 	sh->dev_cap.hw_csum =
232 		!!(attr_ex.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
233 	DRV_LOG(DEBUG, "Checksum offloading is %ssupported.",
234 		sh->dev_cap.hw_csum ? "" : "not ");
235 	sh->dev_cap.hw_vlan_strip = !!(attr_ex.raw_packet_caps &
236 				       IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
237 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported.",
238 		(sh->dev_cap.hw_vlan_strip ? "" : "not "));
239 	sh->dev_cap.hw_fcs_strip = !!(attr_ex.raw_packet_caps &
240 				      IBV_RAW_PACKET_CAP_SCATTER_FCS);
241 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
242 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
243 	DRV_LOG(DEBUG, "Counters are not supported.");
244 #endif
245 	/*
246 	 * DPDK doesn't support larger/variable indirection tables.
247 	 * Once DPDK supports it, take max size from device attr.
248 	 */
249 	sh->dev_cap.ind_table_max_size =
250 			RTE_MIN(attr_ex.rss_caps.max_rwq_indirection_table_size,
251 				(unsigned int)RTE_ETH_RSS_RETA_SIZE_512);
252 	DRV_LOG(DEBUG, "Maximum Rx indirection table size is %u",
253 		sh->dev_cap.ind_table_max_size);
254 	sh->dev_cap.tso = (attr_ex.tso_caps.max_tso > 0 &&
255 			   (attr_ex.tso_caps.supported_qpts &
256 			    (1 << IBV_QPT_RAW_PACKET)));
257 	if (sh->dev_cap.tso)
258 		sh->dev_cap.tso_max_payload_sz = attr_ex.tso_caps.max_tso;
259 	strlcpy(sh->dev_cap.fw_ver, attr_ex.orig_attr.fw_ver,
260 		sizeof(sh->dev_cap.fw_ver));
261 #ifdef HAVE_IBV_MLX5_MOD_SWP
262 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
263 		sh->dev_cap.swp = dv_attr.sw_parsing_caps.sw_parsing_offloads &
264 				  (MLX5_SW_PARSING_CAP |
265 				   MLX5_SW_PARSING_CSUM_CAP |
266 				   MLX5_SW_PARSING_TSO_CAP);
267 	DRV_LOG(DEBUG, "SWP support: %u", sh->dev_cap.swp);
268 #endif
269 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
270 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
271 		struct mlx5dv_striding_rq_caps *strd_rq_caps =
272 				&dv_attr.striding_rq_caps;
273 
274 		sh->dev_cap.mprq.enabled = 1;
275 		sh->dev_cap.mprq.log_min_stride_size =
276 			strd_rq_caps->min_single_stride_log_num_of_bytes;
277 		sh->dev_cap.mprq.log_max_stride_size =
278 			strd_rq_caps->max_single_stride_log_num_of_bytes;
279 		sh->dev_cap.mprq.log_min_stride_num =
280 			strd_rq_caps->min_single_wqe_log_num_of_strides;
281 		sh->dev_cap.mprq.log_max_stride_num =
282 			strd_rq_caps->max_single_wqe_log_num_of_strides;
283 		sh->dev_cap.mprq.log_min_stride_wqe_size =
284 					cdev->config.devx ?
285 					hca_attr->log_min_stride_wqe_sz :
286 					MLX5_MPRQ_LOG_MIN_STRIDE_WQE_SIZE;
287 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %u",
288 			sh->dev_cap.mprq.log_min_stride_size);
289 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %u",
290 			sh->dev_cap.mprq.log_max_stride_size);
291 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %u",
292 			sh->dev_cap.mprq.log_min_stride_num);
293 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %u",
294 			sh->dev_cap.mprq.log_max_stride_num);
295 		DRV_LOG(DEBUG, "\tmin_stride_wqe_log_size: %u",
296 			sh->dev_cap.mprq.log_min_stride_wqe_size);
297 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
298 			strd_rq_caps->supported_qpts);
299 		DRV_LOG(DEBUG, "Device supports Multi-Packet RQ.");
300 	}
301 #endif
302 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
303 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
304 		sh->dev_cap.tunnel_en = dv_attr.tunnel_offloads_caps &
305 					(MLX5_TUNNELED_OFFLOADS_VXLAN_CAP |
306 					 MLX5_TUNNELED_OFFLOADS_GRE_CAP |
307 					 MLX5_TUNNELED_OFFLOADS_GENEVE_CAP);
308 	}
309 	if (sh->dev_cap.tunnel_en) {
310 		DRV_LOG(DEBUG, "Tunnel offloading is supported for %s%s%s",
311 			sh->dev_cap.tunnel_en &
312 			MLX5_TUNNELED_OFFLOADS_VXLAN_CAP ? "[VXLAN]" : "",
313 			sh->dev_cap.tunnel_en &
314 			MLX5_TUNNELED_OFFLOADS_GRE_CAP ? "[GRE]" : "",
315 			sh->dev_cap.tunnel_en &
316 			MLX5_TUNNELED_OFFLOADS_GENEVE_CAP ? "[GENEVE]" : "");
317 	} else {
318 		DRV_LOG(DEBUG, "Tunnel offloading is not supported.");
319 	}
320 #else
321 	DRV_LOG(WARNING,
322 		"Tunnel offloading disabled due to old OFED/rdma-core version");
323 #endif
324 	if (!sh->cdev->config.devx)
325 		return 0;
326 	/* Check capabilities for Packet Pacing. */
327 	DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz.",
328 		hca_attr->dev_freq_khz);
329 	DRV_LOG(DEBUG, "Packet pacing is %ssupported.",
330 		hca_attr->qos.packet_pacing ? "" : "not ");
331 	DRV_LOG(DEBUG, "Cross channel ops are %ssupported.",
332 		hca_attr->cross_channel ? "" : "not ");
333 	DRV_LOG(DEBUG, "WQE index ignore is %ssupported.",
334 		hca_attr->wqe_index_ignore ? "" : "not ");
335 	DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported.",
336 		hca_attr->non_wire_sq ? "" : "not ");
337 	DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
338 		hca_attr->log_max_static_sq_wq ? "" : "not ",
339 		hca_attr->log_max_static_sq_wq);
340 	DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported.",
341 		hca_attr->qos.wqe_rate_pp ? "" : "not ");
342 	sh->dev_cap.txpp_en = hca_attr->qos.packet_pacing;
343 	if (!hca_attr->cross_channel) {
344 		DRV_LOG(DEBUG,
345 			"Cross channel operations are required for packet pacing.");
346 		sh->dev_cap.txpp_en = 0;
347 	}
348 	if (!hca_attr->wqe_index_ignore) {
349 		DRV_LOG(DEBUG,
350 			"WQE index ignore feature is required for packet pacing.");
351 		sh->dev_cap.txpp_en = 0;
352 	}
353 	if (!hca_attr->non_wire_sq) {
354 		DRV_LOG(DEBUG,
355 			"Non-wire SQ feature is required for packet pacing.");
356 		sh->dev_cap.txpp_en = 0;
357 	}
358 	if (!hca_attr->log_max_static_sq_wq) {
359 		DRV_LOG(DEBUG,
360 			"Static WQE SQ feature is required for packet pacing.");
361 		sh->dev_cap.txpp_en = 0;
362 	}
363 	if (!hca_attr->qos.wqe_rate_pp) {
364 		DRV_LOG(DEBUG,
365 			"WQE rate mode is required for packet pacing.");
366 		sh->dev_cap.txpp_en = 0;
367 	}
368 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
369 	DRV_LOG(DEBUG,
370 		"DevX does not provide UAR offset, can't create queues for packet pacing.");
371 	sh->dev_cap.txpp_en = 0;
372 #endif
373 	/* Check for LRO support. */
374 	if (mlx5_devx_obj_ops_en(sh) && hca_attr->lro_cap) {
375 		/* TBD check tunnel lro caps. */
376 		sh->dev_cap.lro_supported = 1;
377 		DRV_LOG(DEBUG, "Device supports LRO.");
378 		DRV_LOG(DEBUG,
379 			"LRO minimal size of TCP segment required for coalescing is %d bytes.",
380 			hca_attr->lro_min_mss_size);
381 	}
382 	sh->dev_cap.scatter_fcs_w_decap_disable =
383 					hca_attr->scatter_fcs_w_decap_disable;
384 	sh->dev_cap.rq_delay_drop_en = hca_attr->rq_delay_drop;
385 	mlx5_rt_timestamp_config(sh, hca_attr);
386 	return 0;
387 }
388 
389 /**
390  * Detect misc5 support or not
391  *
392  * @param[in] priv
393  *   Device private data pointer
394  */
395 #ifdef HAVE_MLX5DV_DR
396 static void
397 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
398 {
399 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
400 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
401 	 * Case: IPv4--->UDP--->VxLAN--->vni
402 	 */
403 	void *tbl;
404 	struct mlx5_flow_dv_match_params matcher_mask;
405 	void *match_m;
406 	void *matcher;
407 	void *headers_m;
408 	void *misc5_m;
409 	uint32_t *tunnel_header_m;
410 	struct mlx5dv_flow_matcher_attr dv_attr;
411 
412 	memset(&matcher_mask, 0, sizeof(matcher_mask));
413 	matcher_mask.size = sizeof(matcher_mask.buf);
414 	match_m = matcher_mask.buf;
415 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
416 	misc5_m = MLX5_ADDR_OF(fte_match_param,
417 			       match_m, misc_parameters_5);
418 	tunnel_header_m = (uint32_t *)
419 				MLX5_ADDR_OF(fte_match_set_misc5,
420 				misc5_m, tunnel_header_1);
421 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
422 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
423 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
424 	*tunnel_header_m = 0xffffff;
425 
426 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
427 	if (!tbl) {
428 		DRV_LOG(INFO, "No SW steering support");
429 		return;
430 	}
431 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
432 	dv_attr.match_mask = (void *)&matcher_mask,
433 	dv_attr.match_criteria_enable =
434 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
435 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
436 	dv_attr.priority = 3;
437 #ifdef HAVE_MLX5DV_DR_ESWITCH
438 	void *misc2_m;
439 	if (priv->sh->config.dv_esw_en) {
440 		/* FDB enabled reg_c_0 */
441 		dv_attr.match_criteria_enable |=
442 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
443 		misc2_m = MLX5_ADDR_OF(fte_match_param,
444 				       match_m, misc_parameters_2);
445 		MLX5_SET(fte_match_set_misc2, misc2_m,
446 			 metadata_reg_c_0, 0xffff);
447 	}
448 #endif
449 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
450 						    &dv_attr, tbl);
451 	if (matcher) {
452 		priv->sh->misc5_cap = 1;
453 		mlx5_glue->dv_destroy_flow_matcher(matcher);
454 	}
455 	mlx5_glue->dr_destroy_flow_tbl(tbl);
456 #else
457 	RTE_SET_USED(priv);
458 #endif
459 }
460 #endif
461 
462 /**
463  * Initialize DR related data within private structure.
464  * Routine checks the reference counter and does actual
465  * resources creation/initialization only if counter is zero.
466  *
467  * @param[in] priv
468  *   Pointer to the private device data structure.
469  *
470  * @return
471  *   Zero on success, positive error code otherwise.
472  */
473 static int
474 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
475 {
476 	struct mlx5_dev_ctx_shared *sh = priv->sh;
477 	char s[MLX5_NAME_SIZE] __rte_unused;
478 	int err;
479 
480 	MLX5_ASSERT(sh && sh->refcnt);
481 	if (sh->refcnt > 1)
482 		return 0;
483 	err = mlx5_alloc_table_hash_list(priv);
484 	if (err)
485 		goto error;
486 	if (priv->sh->config.dv_flow_en == 2)
487 		return 0;
488 	/* The resources below are only valid with DV support. */
489 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
490 	/* Init port id action list. */
491 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
492 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
493 						   flow_dv_port_id_create_cb,
494 						   flow_dv_port_id_match_cb,
495 						   flow_dv_port_id_remove_cb,
496 						   flow_dv_port_id_clone_cb,
497 						 flow_dv_port_id_clone_free_cb);
498 	if (!sh->port_id_action_list)
499 		goto error;
500 	/* Init push vlan action list. */
501 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
502 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
503 						    flow_dv_push_vlan_create_cb,
504 						    flow_dv_push_vlan_match_cb,
505 						    flow_dv_push_vlan_remove_cb,
506 						    flow_dv_push_vlan_clone_cb,
507 					       flow_dv_push_vlan_clone_free_cb);
508 	if (!sh->push_vlan_action_list)
509 		goto error;
510 	/* Init sample action list. */
511 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
512 	sh->sample_action_list = mlx5_list_create(s, sh, true,
513 						  flow_dv_sample_create_cb,
514 						  flow_dv_sample_match_cb,
515 						  flow_dv_sample_remove_cb,
516 						  flow_dv_sample_clone_cb,
517 						  flow_dv_sample_clone_free_cb);
518 	if (!sh->sample_action_list)
519 		goto error;
520 	/* Init dest array action list. */
521 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
522 	sh->dest_array_list = mlx5_list_create(s, sh, true,
523 					       flow_dv_dest_array_create_cb,
524 					       flow_dv_dest_array_match_cb,
525 					       flow_dv_dest_array_remove_cb,
526 					       flow_dv_dest_array_clone_cb,
527 					      flow_dv_dest_array_clone_free_cb);
528 	if (!sh->dest_array_list)
529 		goto error;
530 	/* Init shared flex parsers list, no need lcore_share */
531 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
532 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
533 					       mlx5_flex_parser_create_cb,
534 					       mlx5_flex_parser_match_cb,
535 					       mlx5_flex_parser_remove_cb,
536 					       mlx5_flex_parser_clone_cb,
537 					       mlx5_flex_parser_clone_free_cb);
538 	if (!sh->flex_parsers_dv)
539 		goto error;
540 #endif
541 #ifdef HAVE_MLX5DV_DR
542 	void *domain;
543 
544 	/* Reference counter is zero, we should initialize structures. */
545 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
546 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
547 	if (!domain) {
548 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
549 		err = errno;
550 		goto error;
551 	}
552 	sh->rx_domain = domain;
553 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
554 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
555 	if (!domain) {
556 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
557 		err = errno;
558 		goto error;
559 	}
560 	sh->tx_domain = domain;
561 #ifdef HAVE_MLX5DV_DR_ESWITCH
562 	if (sh->config.dv_esw_en) {
563 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
564 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
565 		if (!domain) {
566 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
567 			err = errno;
568 			goto error;
569 		}
570 		sh->fdb_domain = domain;
571 	}
572 	/*
573 	 * The drop action is just some dummy placeholder in rdma-core. It
574 	 * does not belong to domains and has no any attributes, and, can be
575 	 * shared by the entire device.
576 	 */
577 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
578 	if (!sh->dr_drop_action) {
579 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
580 		err = errno;
581 		goto error;
582 	}
583 #endif
584 	if (!sh->tunnel_hub && sh->config.dv_miss_info)
585 		err = mlx5_alloc_tunnel_hub(sh);
586 	if (err) {
587 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
588 		goto error;
589 	}
590 	if (sh->config.reclaim_mode == MLX5_RCM_AGGR) {
591 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
592 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
593 		if (sh->fdb_domain)
594 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
595 	}
596 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
597 	if (!sh->config.allow_duplicate_pattern) {
598 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
599 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
600 #endif
601 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
602 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
603 		if (sh->fdb_domain)
604 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
605 	}
606 
607 	__mlx5_discovery_misc5_cap(priv);
608 #endif /* HAVE_MLX5DV_DR */
609 	sh->default_miss_action =
610 			mlx5_glue->dr_create_flow_action_default_miss();
611 	if (!sh->default_miss_action)
612 		DRV_LOG(WARNING, "Default miss action is not supported.");
613 	LIST_INIT(&sh->shared_rxqs);
614 	return 0;
615 error:
616 	/* Rollback the created objects. */
617 	if (sh->rx_domain) {
618 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
619 		sh->rx_domain = NULL;
620 	}
621 	if (sh->tx_domain) {
622 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
623 		sh->tx_domain = NULL;
624 	}
625 	if (sh->fdb_domain) {
626 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
627 		sh->fdb_domain = NULL;
628 	}
629 	if (sh->dr_drop_action) {
630 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
631 		sh->dr_drop_action = NULL;
632 	}
633 	if (sh->pop_vlan_action) {
634 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
635 		sh->pop_vlan_action = NULL;
636 	}
637 	if (sh->encaps_decaps) {
638 		mlx5_hlist_destroy(sh->encaps_decaps);
639 		sh->encaps_decaps = NULL;
640 	}
641 	if (sh->modify_cmds) {
642 		mlx5_hlist_destroy(sh->modify_cmds);
643 		sh->modify_cmds = NULL;
644 	}
645 	if (sh->tag_table) {
646 		/* tags should be destroyed with flow before. */
647 		mlx5_hlist_destroy(sh->tag_table);
648 		sh->tag_table = NULL;
649 	}
650 	if (sh->tunnel_hub) {
651 		mlx5_release_tunnel_hub(sh, priv->dev_port);
652 		sh->tunnel_hub = NULL;
653 	}
654 	mlx5_free_table_hash_list(priv);
655 	if (sh->port_id_action_list) {
656 		mlx5_list_destroy(sh->port_id_action_list);
657 		sh->port_id_action_list = NULL;
658 	}
659 	if (sh->push_vlan_action_list) {
660 		mlx5_list_destroy(sh->push_vlan_action_list);
661 		sh->push_vlan_action_list = NULL;
662 	}
663 	if (sh->sample_action_list) {
664 		mlx5_list_destroy(sh->sample_action_list);
665 		sh->sample_action_list = NULL;
666 	}
667 	if (sh->dest_array_list) {
668 		mlx5_list_destroy(sh->dest_array_list);
669 		sh->dest_array_list = NULL;
670 	}
671 	return err;
672 }
673 
674 /**
675  * Destroy DR related data within private structure.
676  *
677  * @param[in] priv
678  *   Pointer to the private device data structure.
679  */
680 void
681 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
682 {
683 	struct mlx5_dev_ctx_shared *sh = priv->sh;
684 
685 	MLX5_ASSERT(sh && sh->refcnt);
686 	if (sh->refcnt > 1)
687 		return;
688 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
689 #ifdef HAVE_MLX5DV_DR
690 	if (sh->rx_domain) {
691 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
692 		sh->rx_domain = NULL;
693 	}
694 	if (sh->tx_domain) {
695 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
696 		sh->tx_domain = NULL;
697 	}
698 #ifdef HAVE_MLX5DV_DR_ESWITCH
699 	if (sh->fdb_domain) {
700 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
701 		sh->fdb_domain = NULL;
702 	}
703 	if (sh->dr_drop_action) {
704 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
705 		sh->dr_drop_action = NULL;
706 	}
707 #endif
708 	if (sh->pop_vlan_action) {
709 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
710 		sh->pop_vlan_action = NULL;
711 	}
712 #endif /* HAVE_MLX5DV_DR */
713 	if (sh->default_miss_action)
714 		mlx5_glue->destroy_flow_action
715 				(sh->default_miss_action);
716 	if (sh->encaps_decaps) {
717 		mlx5_hlist_destroy(sh->encaps_decaps);
718 		sh->encaps_decaps = NULL;
719 	}
720 	if (sh->modify_cmds) {
721 		mlx5_hlist_destroy(sh->modify_cmds);
722 		sh->modify_cmds = NULL;
723 	}
724 	if (sh->tag_table) {
725 		/* tags should be destroyed with flow before. */
726 		mlx5_hlist_destroy(sh->tag_table);
727 		sh->tag_table = NULL;
728 	}
729 	if (sh->tunnel_hub) {
730 		mlx5_release_tunnel_hub(sh, priv->dev_port);
731 		sh->tunnel_hub = NULL;
732 	}
733 	mlx5_free_table_hash_list(priv);
734 	if (sh->port_id_action_list) {
735 		mlx5_list_destroy(sh->port_id_action_list);
736 		sh->port_id_action_list = NULL;
737 	}
738 	if (sh->push_vlan_action_list) {
739 		mlx5_list_destroy(sh->push_vlan_action_list);
740 		sh->push_vlan_action_list = NULL;
741 	}
742 	if (sh->sample_action_list) {
743 		mlx5_list_destroy(sh->sample_action_list);
744 		sh->sample_action_list = NULL;
745 	}
746 	if (sh->dest_array_list) {
747 		mlx5_list_destroy(sh->dest_array_list);
748 		sh->dest_array_list = NULL;
749 	}
750 }
751 
752 /**
753  * Initialize shared data between primary and secondary process.
754  *
755  * A memzone is reserved by primary process and secondary processes attach to
756  * the memzone.
757  *
758  * @return
759  *   0 on success, a negative errno value otherwise and rte_errno is set.
760  */
761 static int
762 mlx5_init_shared_data(void)
763 {
764 	const struct rte_memzone *mz;
765 	int ret = 0;
766 
767 	rte_spinlock_lock(&mlx5_shared_data_lock);
768 	if (mlx5_shared_data == NULL) {
769 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
770 			/* Allocate shared memory. */
771 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
772 						 sizeof(*mlx5_shared_data),
773 						 SOCKET_ID_ANY, 0);
774 			if (mz == NULL) {
775 				DRV_LOG(ERR,
776 					"Cannot allocate mlx5 shared data");
777 				ret = -rte_errno;
778 				goto error;
779 			}
780 			mlx5_shared_data = mz->addr;
781 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
782 			rte_spinlock_init(&mlx5_shared_data->lock);
783 		} else {
784 			/* Lookup allocated shared memory. */
785 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
786 			if (mz == NULL) {
787 				DRV_LOG(ERR,
788 					"Cannot attach mlx5 shared data");
789 				ret = -rte_errno;
790 				goto error;
791 			}
792 			mlx5_shared_data = mz->addr;
793 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
794 		}
795 	}
796 error:
797 	rte_spinlock_unlock(&mlx5_shared_data_lock);
798 	return ret;
799 }
800 
801 /**
802  * PMD global initialization.
803  *
804  * Independent from individual device, this function initializes global
805  * per-PMD data structures distinguishing primary and secondary processes.
806  * Hence, each initialization is called once per a process.
807  *
808  * @return
809  *   0 on success, a negative errno value otherwise and rte_errno is set.
810  */
811 static int
812 mlx5_init_once(void)
813 {
814 	struct mlx5_shared_data *sd;
815 	struct mlx5_local_data *ld = &mlx5_local_data;
816 	int ret = 0;
817 
818 	if (mlx5_init_shared_data())
819 		return -rte_errno;
820 	sd = mlx5_shared_data;
821 	MLX5_ASSERT(sd);
822 	rte_spinlock_lock(&sd->lock);
823 	switch (rte_eal_process_type()) {
824 	case RTE_PROC_PRIMARY:
825 		if (sd->init_done)
826 			break;
827 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
828 					   mlx5_mp_os_primary_handle);
829 		if (ret)
830 			goto out;
831 		sd->init_done = true;
832 		break;
833 	case RTE_PROC_SECONDARY:
834 		if (ld->init_done)
835 			break;
836 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
837 					     mlx5_mp_os_secondary_handle);
838 		if (ret)
839 			goto out;
840 		++sd->secondary_cnt;
841 		ld->init_done = true;
842 		break;
843 	default:
844 		break;
845 	}
846 out:
847 	rte_spinlock_unlock(&sd->lock);
848 	return ret;
849 }
850 
851 /**
852  * DR flow drop action support detect.
853  *
854  * @param dev
855  *   Pointer to rte_eth_dev structure.
856  *
857  */
858 static void
859 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
860 {
861 #ifdef HAVE_MLX5DV_DR
862 	struct mlx5_priv *priv = dev->data->dev_private;
863 
864 	if (!priv->sh->config.dv_flow_en || !priv->sh->dr_drop_action)
865 		return;
866 	/**
867 	 * DR supports drop action placeholder when it is supported;
868 	 * otherwise, use the queue drop action.
869 	 */
870 	if (!priv->sh->drop_action_check_flag) {
871 		if (!mlx5_flow_discover_dr_action_support(dev))
872 			priv->sh->dr_drop_action_en = 1;
873 		priv->sh->drop_action_check_flag = 1;
874 	}
875 	if (priv->sh->dr_drop_action_en)
876 		priv->root_drop_action = priv->sh->dr_drop_action;
877 	else
878 		priv->root_drop_action = priv->drop_queue.hrxq->action;
879 #endif
880 }
881 
882 static void
883 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
884 {
885 	struct mlx5_priv *priv = dev->data->dev_private;
886 	void *ctx = priv->sh->cdev->ctx;
887 
888 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
889 	if (!priv->q_counters) {
890 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
891 		struct ibv_wq *wq;
892 
893 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
894 			"by DevX - fall-back to use the kernel driver global "
895 			"queue counter.", dev->data->port_id);
896 		/* Create WQ by kernel and query its queue counter ID. */
897 		if (cq) {
898 			wq = mlx5_glue->create_wq(ctx,
899 						  &(struct ibv_wq_init_attr){
900 						    .wq_type = IBV_WQT_RQ,
901 						    .max_wr = 1,
902 						    .max_sge = 1,
903 						    .pd = priv->sh->cdev->pd,
904 						    .cq = cq,
905 						});
906 			if (wq) {
907 				/* Counter is assigned only on RDY state. */
908 				int ret = mlx5_glue->modify_wq(wq,
909 						 &(struct ibv_wq_attr){
910 						 .attr_mask = IBV_WQ_ATTR_STATE,
911 						 .wq_state = IBV_WQS_RDY,
912 						});
913 
914 				if (ret == 0)
915 					mlx5_devx_cmd_wq_query(wq,
916 							 &priv->counter_set_id);
917 				claim_zero(mlx5_glue->destroy_wq(wq));
918 			}
919 			claim_zero(mlx5_glue->destroy_cq(cq));
920 		}
921 	} else {
922 		priv->counter_set_id = priv->q_counters->id;
923 	}
924 	if (priv->counter_set_id == 0)
925 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
926 			"available.", dev->data->port_id);
927 }
928 
929 /**
930  * Check if representor spawn info match devargs.
931  *
932  * @param spawn
933  *   Verbs device parameters (name, port, switch_info) to spawn.
934  * @param eth_da
935  *   Device devargs to probe.
936  *
937  * @return
938  *   Match result.
939  */
940 static bool
941 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
942 		       struct rte_eth_devargs *eth_da)
943 {
944 	struct mlx5_switch_info *switch_info = &spawn->info;
945 	unsigned int p, f;
946 	uint16_t id;
947 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
948 						      eth_da->type);
949 
950 	switch (eth_da->type) {
951 	case RTE_ETH_REPRESENTOR_SF:
952 		if (!(spawn->info.port_name == -1 &&
953 		      switch_info->name_type ==
954 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
955 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
956 			rte_errno = EBUSY;
957 			return false;
958 		}
959 		break;
960 	case RTE_ETH_REPRESENTOR_VF:
961 		/* Allows HPF representor index -1 as exception. */
962 		if (!(spawn->info.port_name == -1 &&
963 		      switch_info->name_type ==
964 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
965 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
966 			rte_errno = EBUSY;
967 			return false;
968 		}
969 		break;
970 	case RTE_ETH_REPRESENTOR_NONE:
971 		rte_errno = EBUSY;
972 		return false;
973 	default:
974 		rte_errno = ENOTSUP;
975 		DRV_LOG(ERR, "unsupported representor type");
976 		return false;
977 	}
978 	/* Check representor ID: */
979 	for (p = 0; p < eth_da->nb_ports; ++p) {
980 		if (spawn->pf_bond < 0) {
981 			/* For non-LAG mode, allow and ignore pf. */
982 			switch_info->pf_num = eth_da->ports[p];
983 			repr_id = mlx5_representor_id_encode(switch_info,
984 							     eth_da->type);
985 		}
986 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
987 			id = MLX5_REPRESENTOR_ID
988 				(eth_da->ports[p], eth_da->type,
989 				 eth_da->representor_ports[f]);
990 			if (repr_id == id)
991 				return true;
992 		}
993 	}
994 	rte_errno = EBUSY;
995 	return false;
996 }
997 
998 /**
999  * Spawn an Ethernet device from Verbs information.
1000  *
1001  * @param dpdk_dev
1002  *   Backing DPDK device.
1003  * @param spawn
1004  *   Verbs device parameters (name, port, switch_info) to spawn.
1005  * @param eth_da
1006  *   Device arguments.
1007  * @param mkvlist
1008  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1009  *
1010  * @return
1011  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1012  *   is set. The following errors are defined:
1013  *
1014  *   EBUSY: device is not supposed to be spawned.
1015  *   EEXIST: device is already spawned
1016  */
1017 static struct rte_eth_dev *
1018 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1019 	       struct mlx5_dev_spawn_data *spawn,
1020 	       struct rte_eth_devargs *eth_da,
1021 	       struct mlx5_kvargs_ctrl *mkvlist)
1022 {
1023 	const struct mlx5_switch_info *switch_info = &spawn->info;
1024 	struct mlx5_dev_ctx_shared *sh = NULL;
1025 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
1026 	struct rte_eth_dev *eth_dev = NULL;
1027 	struct mlx5_priv *priv = NULL;
1028 	int err = 0;
1029 	struct rte_ether_addr mac;
1030 	char name[RTE_ETH_NAME_MAX_LEN];
1031 	int own_domain_id = 0;
1032 	uint16_t port_id;
1033 	struct mlx5_port_info vport_info = { .query_flags = 0 };
1034 	int nl_rdma;
1035 	int i;
1036 
1037 	/* Determine if this port representor is supposed to be spawned. */
1038 	if (switch_info->representor && dpdk_dev->devargs &&
1039 	    !mlx5_representor_match(spawn, eth_da))
1040 		return NULL;
1041 	/* Build device name. */
1042 	if (spawn->pf_bond < 0) {
1043 		/* Single device. */
1044 		if (!switch_info->representor)
1045 			strlcpy(name, dpdk_dev->name, sizeof(name));
1046 		else
1047 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
1048 				 dpdk_dev->name,
1049 				 switch_info->name_type ==
1050 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1051 				 switch_info->port_name);
1052 	} else {
1053 		/* Bonding device. */
1054 		if (!switch_info->representor) {
1055 			err = snprintf(name, sizeof(name), "%s_%s",
1056 				       dpdk_dev->name, spawn->phys_dev_name);
1057 		} else {
1058 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1059 				dpdk_dev->name, spawn->phys_dev_name,
1060 				switch_info->ctrl_num,
1061 				switch_info->pf_num,
1062 				switch_info->name_type ==
1063 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1064 				switch_info->port_name);
1065 		}
1066 	}
1067 	if (err >= (int)sizeof(name))
1068 		DRV_LOG(WARNING, "device name overflow %s", name);
1069 	/* check if the device is already spawned */
1070 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1071 		/*
1072 		 * When device is already spawned, its devargs should be set
1073 		 * as used. otherwise, mlx5_kvargs_validate() will fail.
1074 		 */
1075 		if (mkvlist)
1076 			mlx5_port_args_set_used(name, port_id, mkvlist);
1077 		rte_errno = EEXIST;
1078 		return NULL;
1079 	}
1080 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1081 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1082 		struct mlx5_mp_id mp_id;
1083 
1084 		eth_dev = rte_eth_dev_attach_secondary(name);
1085 		if (eth_dev == NULL) {
1086 			DRV_LOG(ERR, "can not attach rte ethdev");
1087 			rte_errno = ENOMEM;
1088 			return NULL;
1089 		}
1090 		eth_dev->device = dpdk_dev;
1091 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1092 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1093 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1094 		err = mlx5_proc_priv_init(eth_dev);
1095 		if (err)
1096 			return NULL;
1097 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1098 		/* Receive command fd from primary process */
1099 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1100 		if (err < 0)
1101 			goto err_secondary;
1102 		/* Remap UAR for Tx queues. */
1103 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1104 		if (err)
1105 			goto err_secondary;
1106 		/*
1107 		 * Ethdev pointer is still required as input since
1108 		 * the primary device is not accessible from the
1109 		 * secondary process.
1110 		 */
1111 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1112 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1113 		return eth_dev;
1114 err_secondary:
1115 		mlx5_dev_close(eth_dev);
1116 		return NULL;
1117 	}
1118 	sh = mlx5_alloc_shared_dev_ctx(spawn, mkvlist);
1119 	if (!sh)
1120 		return NULL;
1121 	nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
1122 	/* Check port status. */
1123 	if (spawn->phys_port <= UINT8_MAX) {
1124 		/* Legacy Verbs api only support u8 port number. */
1125 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1126 					    &port_attr);
1127 		if (err) {
1128 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1129 			goto error;
1130 		}
1131 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1132 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1133 			err = EINVAL;
1134 			goto error;
1135 		}
1136 	} else if (nl_rdma >= 0) {
1137 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1138 		err = mlx5_nl_port_state(nl_rdma,
1139 			spawn->phys_dev_name,
1140 			spawn->phys_port);
1141 		if (err < 0) {
1142 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1143 				strerror(rte_errno));
1144 			err = -rte_errno;
1145 			goto error;
1146 		}
1147 		port_attr.state = (enum ibv_port_state)err;
1148 	}
1149 	if (port_attr.state != IBV_PORT_ACTIVE)
1150 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1151 			mlx5_glue->port_state_str(port_attr.state),
1152 			port_attr.state);
1153 	/* Allocate private eth device data. */
1154 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1155 			   sizeof(*priv),
1156 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1157 	if (priv == NULL) {
1158 		DRV_LOG(ERR, "priv allocation failure");
1159 		err = ENOMEM;
1160 		goto error;
1161 	}
1162 	/*
1163 	 * When user configures remote PD and CTX and device creates RxQ by
1164 	 * DevX, external RxQ is both supported and requested.
1165 	 */
1166 	if (mlx5_imported_pd_and_ctx(sh->cdev) && mlx5_devx_obj_ops_en(sh)) {
1167 		priv->ext_rxqs = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1168 					     sizeof(struct mlx5_external_rxq) *
1169 					     MLX5_MAX_EXT_RX_QUEUES, 0,
1170 					     SOCKET_ID_ANY);
1171 		if (priv->ext_rxqs == NULL) {
1172 			DRV_LOG(ERR, "Fail to allocate external RxQ array.");
1173 			err = ENOMEM;
1174 			goto error;
1175 		}
1176 		DRV_LOG(DEBUG, "External RxQ is supported.");
1177 	}
1178 	priv->sh = sh;
1179 	priv->dev_port = spawn->phys_port;
1180 	priv->pci_dev = spawn->pci_dev;
1181 	priv->mtu = RTE_ETHER_MTU;
1182 	/* Some internal functions rely on Netlink sockets, open them now. */
1183 	priv->nl_socket_rdma = nl_rdma;
1184 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE, 0);
1185 	priv->representor = !!switch_info->representor;
1186 	priv->master = !!switch_info->master;
1187 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1188 	priv->vport_meta_tag = 0;
1189 	priv->vport_meta_mask = 0;
1190 	priv->pf_bond = spawn->pf_bond;
1191 
1192 	DRV_LOG(DEBUG,
1193 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1194 		priv->dev_port, dpdk_dev->bus->name,
1195 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1196 		priv->master, priv->representor, priv->pf_bond);
1197 
1198 	/*
1199 	 * If we have E-Switch we should determine the vport attributes.
1200 	 * E-Switch may use either source vport field or reg_c[0] metadata
1201 	 * register to match on vport index. The engaged part of metadata
1202 	 * register is defined by mask.
1203 	 */
1204 	if (sh->esw_mode) {
1205 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1206 						 spawn->phys_port,
1207 						 &vport_info);
1208 		if (err) {
1209 			DRV_LOG(WARNING,
1210 				"Cannot query devx port %d on device %s",
1211 				spawn->phys_port, spawn->phys_dev_name);
1212 			vport_info.query_flags = 0;
1213 		}
1214 	}
1215 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1216 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1217 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1218 		if (!priv->vport_meta_mask) {
1219 			DRV_LOG(ERR,
1220 				"vport zero mask for port %d on bonding device %s",
1221 				spawn->phys_port, spawn->phys_dev_name);
1222 			err = ENOTSUP;
1223 			goto error;
1224 		}
1225 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1226 			DRV_LOG(ERR,
1227 				"Invalid vport tag for port %d on bonding device %s",
1228 				spawn->phys_port, spawn->phys_dev_name);
1229 			err = ENOTSUP;
1230 			goto error;
1231 		}
1232 	}
1233 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1234 		priv->vport_id = vport_info.vport_id;
1235 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1236 		DRV_LOG(ERR,
1237 			"Cannot deduce vport index for port %d on bonding device %s",
1238 			spawn->phys_port, spawn->phys_dev_name);
1239 		err = ENOTSUP;
1240 		goto error;
1241 	} else {
1242 		/*
1243 		 * Suppose vport index in compatible way. Kernel/rdma_core
1244 		 * support single E-Switch per PF configurations only and
1245 		 * vport_id field contains the vport index for associated VF,
1246 		 * which is deduced from representor port name.
1247 		 * For example, let's have the IB device port 10, it has
1248 		 * attached network device eth0, which has port name attribute
1249 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1250 		 * as 3 (2+1). This assigning schema should be changed if the
1251 		 * multiple E-Switch instances per PF configurations or/and PCI
1252 		 * subfunctions are added.
1253 		 */
1254 		priv->vport_id = switch_info->representor ?
1255 				 switch_info->port_name + 1 : -1;
1256 	}
1257 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1258 							  eth_da->type);
1259 	/*
1260 	 * Look for sibling devices in order to reuse their switch domain
1261 	 * if any, otherwise allocate one.
1262 	 */
1263 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1264 		const struct mlx5_priv *opriv =
1265 			rte_eth_devices[port_id].data->dev_private;
1266 
1267 		if (!opriv ||
1268 		    opriv->sh != priv->sh ||
1269 			opriv->domain_id ==
1270 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1271 			continue;
1272 		priv->domain_id = opriv->domain_id;
1273 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1274 			priv->dev_port, priv->domain_id);
1275 		break;
1276 	}
1277 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1278 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1279 		if (err) {
1280 			err = rte_errno;
1281 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1282 				strerror(rte_errno));
1283 			goto error;
1284 		}
1285 		own_domain_id = 1;
1286 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1287 			priv->dev_port, priv->domain_id);
1288 	}
1289 	if (sh->cdev->config.devx) {
1290 		struct mlx5_hca_attr *hca_attr = &sh->cdev->config.hca_attr;
1291 
1292 		sh->steering_format_version = hca_attr->steering_format_version;
1293 #if defined(HAVE_MLX5DV_DR) && \
1294 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1295 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1296 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1297 		    sh->config.dv_flow_en) {
1298 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1299 			/*
1300 			 * Meter needs two REG_C's for color match and pre-sfx
1301 			 * flow match. Here get the REG_C for color match.
1302 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1303 			 */
1304 			reg_c_mask &= 0xfc;
1305 			if (__builtin_popcount(reg_c_mask) < 1) {
1306 				priv->mtr_en = 0;
1307 				DRV_LOG(WARNING, "No available register for"
1308 					" meter.");
1309 			} else {
1310 				/*
1311 				 * The meter color register is used by the
1312 				 * flow-hit feature as well.
1313 				 * The flow-hit feature must use REG_C_3
1314 				 * Prefer REG_C_3 if it is available.
1315 				 */
1316 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1317 					priv->mtr_color_reg = REG_C_3;
1318 				else
1319 					priv->mtr_color_reg = ffs(reg_c_mask)
1320 							      - 1 + REG_C_0;
1321 				priv->mtr_en = 1;
1322 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1323 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1324 					priv->mtr_color_reg);
1325 			}
1326 		}
1327 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1328 			uint32_t log_obj_size =
1329 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1330 			if (log_obj_size >=
1331 			    hca_attr->qos.log_meter_aso_granularity &&
1332 			    log_obj_size <=
1333 			    hca_attr->qos.log_meter_aso_max_alloc)
1334 				sh->meter_aso_en = 1;
1335 		}
1336 		if (priv->mtr_en) {
1337 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1338 			if (err) {
1339 				err = -err;
1340 				goto error;
1341 			}
1342 		}
1343 		if (hca_attr->flow.tunnel_header_0_1)
1344 			sh->tunnel_header_0_1 = 1;
1345 		if (hca_attr->flow.tunnel_header_2_3)
1346 			sh->tunnel_header_2_3 = 1;
1347 #endif
1348 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1349 		if (hca_attr->flow_hit_aso && priv->mtr_color_reg == REG_C_3) {
1350 			sh->flow_hit_aso_en = 1;
1351 			err = mlx5_flow_aso_age_mng_init(sh);
1352 			if (err) {
1353 				err = -err;
1354 				goto error;
1355 			}
1356 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1357 		}
1358 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1359 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1360 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1361 		if (hca_attr->ct_offload && priv->mtr_color_reg == REG_C_3) {
1362 			err = mlx5_flow_aso_ct_mng_init(sh);
1363 			if (err) {
1364 				err = -err;
1365 				goto error;
1366 			}
1367 			DRV_LOG(DEBUG, "CT ASO is supported.");
1368 			sh->ct_aso_en = 1;
1369 		}
1370 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1371 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1372 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1373 		    sh->config.dv_flow_en) {
1374 			priv->sampler_en = 1;
1375 			DRV_LOG(DEBUG, "Sampler enabled!");
1376 		} else {
1377 			priv->sampler_en = 0;
1378 			if (!hca_attr->log_max_ft_sampler_num)
1379 				DRV_LOG(WARNING,
1380 					"No available register for sampler.");
1381 			else
1382 				DRV_LOG(DEBUG, "DV flow is not supported!");
1383 		}
1384 #endif
1385 	}
1386 	/* Process parameters and store port configuration on priv structure. */
1387 	err = mlx5_port_args_config(priv, mkvlist, &priv->config);
1388 	if (err) {
1389 		err = rte_errno;
1390 		DRV_LOG(ERR, "Failed to process port configure: %s",
1391 			strerror(rte_errno));
1392 		goto error;
1393 	}
1394 	eth_dev = rte_eth_dev_allocate(name);
1395 	if (eth_dev == NULL) {
1396 		DRV_LOG(ERR, "can not allocate rte ethdev");
1397 		err = ENOMEM;
1398 		goto error;
1399 	}
1400 	if (priv->representor) {
1401 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1402 		eth_dev->data->representor_id = priv->representor_id;
1403 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1404 			struct mlx5_priv *opriv =
1405 				rte_eth_devices[port_id].data->dev_private;
1406 			if (opriv &&
1407 			    opriv->master &&
1408 			    opriv->domain_id == priv->domain_id &&
1409 			    opriv->sh == priv->sh) {
1410 				eth_dev->data->backer_port_id = port_id;
1411 				break;
1412 			}
1413 		}
1414 		if (port_id >= RTE_MAX_ETHPORTS)
1415 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1416 	}
1417 	priv->mp_id.port_id = eth_dev->data->port_id;
1418 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1419 	/*
1420 	 * Store associated network device interface index. This index
1421 	 * is permanent throughout the lifetime of device. So, we may store
1422 	 * the ifindex here and use the cached value further.
1423 	 */
1424 	MLX5_ASSERT(spawn->ifindex);
1425 	priv->if_index = spawn->ifindex;
1426 	priv->lag_affinity_idx = sh->refcnt - 1;
1427 	eth_dev->data->dev_private = priv;
1428 	priv->dev_data = eth_dev->data;
1429 	eth_dev->data->mac_addrs = priv->mac;
1430 	eth_dev->device = dpdk_dev;
1431 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1432 	/* Configure the first MAC address by default. */
1433 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1434 		DRV_LOG(ERR,
1435 			"port %u cannot get MAC address, is mlx5_en"
1436 			" loaded? (errno: %s)",
1437 			eth_dev->data->port_id, strerror(rte_errno));
1438 		err = ENODEV;
1439 		goto error;
1440 	}
1441 	DRV_LOG(INFO,
1442 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1443 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1444 #ifdef RTE_LIBRTE_MLX5_DEBUG
1445 	{
1446 		char ifname[MLX5_NAMESIZE];
1447 
1448 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1449 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1450 				eth_dev->data->port_id, ifname);
1451 		else
1452 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1453 				eth_dev->data->port_id);
1454 	}
1455 #endif
1456 	/* Get actual MTU if possible. */
1457 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1458 	if (err) {
1459 		err = rte_errno;
1460 		goto error;
1461 	}
1462 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1463 		priv->mtu);
1464 	/* Initialize burst functions to prevent crashes before link-up. */
1465 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1466 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1467 	eth_dev->dev_ops = &mlx5_dev_ops;
1468 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1469 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1470 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1471 	/* Register MAC address. */
1472 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1473 	if (sh->dev_cap.vf && sh->config.vf_nl_en)
1474 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1475 				      mlx5_ifindex(eth_dev),
1476 				      eth_dev->data->mac_addrs,
1477 				      MLX5_MAX_MAC_ADDRESSES);
1478 	priv->ctrl_flows = 0;
1479 	rte_spinlock_init(&priv->flow_list_lock);
1480 	TAILQ_INIT(&priv->flow_meters);
1481 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1482 	if (!priv->mtr_profile_tbl)
1483 		goto error;
1484 	/* Bring Ethernet device up. */
1485 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1486 		eth_dev->data->port_id);
1487 	/* Read link status in case it is up and there will be no event. */
1488 	mlx5_link_update(eth_dev, 0);
1489 	/* Watch LSC interrupts between port probe and port start. */
1490 	priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1491 							eth_dev->data->port_id;
1492 	mlx5_set_link_up(eth_dev);
1493 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1494 		icfg[i].release_mem_en = !!sh->config.reclaim_mode;
1495 		if (sh->config.reclaim_mode)
1496 			icfg[i].per_core_cache = 0;
1497 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1498 		if (!priv->flows[i])
1499 			goto error;
1500 	}
1501 	/* Create context for virtual machine VLAN workaround. */
1502 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1503 	if (sh->config.dv_flow_en) {
1504 		err = mlx5_alloc_shared_dr(priv);
1505 		if (err)
1506 			goto error;
1507 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1508 			goto error;
1509 	}
1510 	if (mlx5_devx_obj_ops_en(sh)) {
1511 		priv->obj_ops = devx_obj_ops;
1512 		mlx5_queue_counter_id_prepare(eth_dev);
1513 		priv->obj_ops.lb_dummy_queue_create =
1514 					mlx5_rxq_ibv_obj_dummy_lb_create;
1515 		priv->obj_ops.lb_dummy_queue_release =
1516 					mlx5_rxq_ibv_obj_dummy_lb_release;
1517 	} else if (spawn->max_port > UINT8_MAX) {
1518 		/* Verbs can't support ports larger than 255 by design. */
1519 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1520 		err = ENOTSUP;
1521 		goto error;
1522 	} else {
1523 		priv->obj_ops = ibv_obj_ops;
1524 	}
1525 	if (sh->config.tx_pp &&
1526 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1527 		/*
1528 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1529 		 * packet pacing and already checked above.
1530 		 * Hence, we should only make sure the SQs will be created
1531 		 * with DevX, not with Verbs.
1532 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1533 		 * with Clock Queue UAR as required for Tx scheduling.
1534 		 */
1535 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1536 		err = ENODEV;
1537 		goto error;
1538 	}
1539 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1540 	if (!priv->drop_queue.hrxq)
1541 		goto error;
1542 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1543 				       mlx5_hrxq_create_cb,
1544 				       mlx5_hrxq_match_cb,
1545 				       mlx5_hrxq_remove_cb,
1546 				       mlx5_hrxq_clone_cb,
1547 				       mlx5_hrxq_clone_free_cb);
1548 	if (!priv->hrxqs)
1549 		goto error;
1550 	rte_rwlock_init(&priv->ind_tbls_lock);
1551 	if (priv->sh->config.dv_flow_en == 2)
1552 		return eth_dev;
1553 	/* Port representor shares the same max priority with pf port. */
1554 	if (!priv->sh->flow_priority_check_flag) {
1555 		/* Supported Verbs flow priority number detection. */
1556 		err = mlx5_flow_discover_priorities(eth_dev);
1557 		priv->sh->flow_max_priority = err;
1558 		priv->sh->flow_priority_check_flag = 1;
1559 	} else {
1560 		err = priv->sh->flow_max_priority;
1561 	}
1562 	if (err < 0) {
1563 		err = -err;
1564 		goto error;
1565 	}
1566 	mlx5_set_metadata_mask(eth_dev);
1567 	if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1568 	    !priv->sh->dv_regc0_mask) {
1569 		DRV_LOG(ERR, "metadata mode %u is not supported "
1570 			     "(no metadata reg_c[0] is available)",
1571 			     sh->config.dv_xmeta_en);
1572 			err = ENOTSUP;
1573 			goto error;
1574 	}
1575 	/* Query availability of metadata reg_c's. */
1576 	if (!priv->sh->metadata_regc_check_flag) {
1577 		err = mlx5_flow_discover_mreg_c(eth_dev);
1578 		if (err < 0) {
1579 			err = -err;
1580 			goto error;
1581 		}
1582 	}
1583 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1584 		DRV_LOG(DEBUG,
1585 			"port %u extensive metadata register is not supported",
1586 			eth_dev->data->port_id);
1587 		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1588 			DRV_LOG(ERR, "metadata mode %u is not supported "
1589 				     "(no metadata registers available)",
1590 				     sh->config.dv_xmeta_en);
1591 			err = ENOTSUP;
1592 			goto error;
1593 		}
1594 	}
1595 	if (sh->config.dv_flow_en &&
1596 	    sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1597 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1598 	    priv->sh->dv_regc0_mask) {
1599 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1600 						      MLX5_FLOW_MREG_HTABLE_SZ,
1601 						      false, true, eth_dev,
1602 						      flow_dv_mreg_create_cb,
1603 						      flow_dv_mreg_match_cb,
1604 						      flow_dv_mreg_remove_cb,
1605 						      flow_dv_mreg_clone_cb,
1606 						    flow_dv_mreg_clone_free_cb);
1607 		if (!priv->mreg_cp_tbl) {
1608 			err = ENOMEM;
1609 			goto error;
1610 		}
1611 	}
1612 	rte_spinlock_init(&priv->shared_act_sl);
1613 	mlx5_flow_counter_mode_config(eth_dev);
1614 	mlx5_flow_drop_action_config(eth_dev);
1615 	if (sh->config.dv_flow_en)
1616 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1617 	return eth_dev;
1618 error:
1619 	if (priv) {
1620 		if (priv->mreg_cp_tbl)
1621 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1622 		if (priv->sh)
1623 			mlx5_os_free_shared_dr(priv);
1624 		if (priv->nl_socket_route >= 0)
1625 			close(priv->nl_socket_route);
1626 		if (priv->vmwa_context)
1627 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1628 		if (eth_dev && priv->drop_queue.hrxq)
1629 			mlx5_drop_action_destroy(eth_dev);
1630 		if (priv->mtr_profile_tbl)
1631 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1632 		if (own_domain_id)
1633 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1634 		if (priv->hrxqs)
1635 			mlx5_list_destroy(priv->hrxqs);
1636 		if (eth_dev && priv->flex_item_map)
1637 			mlx5_flex_item_port_cleanup(eth_dev);
1638 		mlx5_free(priv->ext_rxqs);
1639 		mlx5_free(priv);
1640 		if (eth_dev != NULL)
1641 			eth_dev->data->dev_private = NULL;
1642 	}
1643 	if (eth_dev != NULL) {
1644 		/* mac_addrs must not be freed alone because part of
1645 		 * dev_private
1646 		 **/
1647 		eth_dev->data->mac_addrs = NULL;
1648 		rte_eth_dev_release_port(eth_dev);
1649 	}
1650 	if (sh)
1651 		mlx5_free_shared_dev_ctx(sh);
1652 	if (nl_rdma >= 0)
1653 		close(nl_rdma);
1654 	MLX5_ASSERT(err > 0);
1655 	rte_errno = err;
1656 	return NULL;
1657 }
1658 
1659 /**
1660  * Comparison callback to sort device data.
1661  *
1662  * This is meant to be used with qsort().
1663  *
1664  * @param a[in]
1665  *   Pointer to pointer to first data object.
1666  * @param b[in]
1667  *   Pointer to pointer to second data object.
1668  *
1669  * @return
1670  *   0 if both objects are equal, less than 0 if the first argument is less
1671  *   than the second, greater than 0 otherwise.
1672  */
1673 static int
1674 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1675 {
1676 	const struct mlx5_switch_info *si_a =
1677 		&((const struct mlx5_dev_spawn_data *)a)->info;
1678 	const struct mlx5_switch_info *si_b =
1679 		&((const struct mlx5_dev_spawn_data *)b)->info;
1680 	int ret;
1681 
1682 	/* Master device first. */
1683 	ret = si_b->master - si_a->master;
1684 	if (ret)
1685 		return ret;
1686 	/* Then representor devices. */
1687 	ret = si_b->representor - si_a->representor;
1688 	if (ret)
1689 		return ret;
1690 	/* Unidentified devices come last in no specific order. */
1691 	if (!si_a->representor)
1692 		return 0;
1693 	/* Order representors by name. */
1694 	return si_a->port_name - si_b->port_name;
1695 }
1696 
1697 /**
1698  * Match PCI information for possible slaves of bonding device.
1699  *
1700  * @param[in] ibdev_name
1701  *   Name of Infiniband device.
1702  * @param[in] pci_dev
1703  *   Pointer to primary PCI address structure to match.
1704  * @param[in] nl_rdma
1705  *   Netlink RDMA group socket handle.
1706  * @param[in] owner
1707  *   Representor owner PF index.
1708  * @param[out] bond_info
1709  *   Pointer to bonding information.
1710  *
1711  * @return
1712  *   negative value if no bonding device found, otherwise
1713  *   positive index of slave PF in bonding.
1714  */
1715 static int
1716 mlx5_device_bond_pci_match(const char *ibdev_name,
1717 			   const struct rte_pci_addr *pci_dev,
1718 			   int nl_rdma, uint16_t owner,
1719 			   struct mlx5_bond_info *bond_info)
1720 {
1721 	char ifname[IF_NAMESIZE + 1];
1722 	unsigned int ifindex;
1723 	unsigned int np, i;
1724 	FILE *bond_file = NULL, *file;
1725 	int pf = -1;
1726 	int ret;
1727 	uint8_t cur_guid[32] = {0};
1728 	uint8_t guid[32] = {0};
1729 
1730 	/*
1731 	 * Try to get master device name. If something goes wrong suppose
1732 	 * the lack of kernel support and no bonding devices.
1733 	 */
1734 	memset(bond_info, 0, sizeof(*bond_info));
1735 	if (nl_rdma < 0)
1736 		return -1;
1737 	if (!strstr(ibdev_name, "bond"))
1738 		return -1;
1739 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1740 	if (!np)
1741 		return -1;
1742 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1743 		return -1;
1744 	/*
1745 	 * The master device might not be on the predefined port(not on port
1746 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1747 	 * device ports and find master.
1748 	 */
1749 	for (i = 1; i <= np; ++i) {
1750 		/* Check whether Infiniband port is populated. */
1751 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1752 		if (!ifindex)
1753 			continue;
1754 		if (!if_indextoname(ifindex, ifname))
1755 			continue;
1756 		/* Try to read bonding slave names from sysfs. */
1757 		MKSTR(slaves,
1758 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1759 		bond_file = fopen(slaves, "r");
1760 		if (bond_file)
1761 			break;
1762 	}
1763 	if (!bond_file)
1764 		return -1;
1765 	/* Use safe format to check maximal buffer length. */
1766 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1767 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1768 		char tmp_str[IF_NAMESIZE + 32];
1769 		struct rte_pci_addr pci_addr;
1770 		struct mlx5_switch_info	info;
1771 		int ret;
1772 
1773 		/* Process slave interface names in the loop. */
1774 		snprintf(tmp_str, sizeof(tmp_str),
1775 			 "/sys/class/net/%s", ifname);
1776 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1777 			DRV_LOG(WARNING,
1778 				"Cannot get PCI address for netdev \"%s\".",
1779 				ifname);
1780 			continue;
1781 		}
1782 		/* Slave interface PCI address match found. */
1783 		snprintf(tmp_str, sizeof(tmp_str),
1784 			 "/sys/class/net/%s/phys_port_name", ifname);
1785 		file = fopen(tmp_str, "rb");
1786 		if (!file)
1787 			break;
1788 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1789 		if (fscanf(file, "%32s", tmp_str) == 1)
1790 			mlx5_translate_port_name(tmp_str, &info);
1791 		fclose(file);
1792 		/* Only process PF ports. */
1793 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1794 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1795 			continue;
1796 		/* Check max bonding member. */
1797 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1798 			DRV_LOG(WARNING, "bonding index out of range, "
1799 				"please increase MLX5_BOND_MAX_PORTS: %s",
1800 				tmp_str);
1801 			break;
1802 		}
1803 		/* Get ifindex. */
1804 		snprintf(tmp_str, sizeof(tmp_str),
1805 			 "/sys/class/net/%s/ifindex", ifname);
1806 		file = fopen(tmp_str, "rb");
1807 		if (!file)
1808 			break;
1809 		ret = fscanf(file, "%u", &ifindex);
1810 		fclose(file);
1811 		if (ret != 1)
1812 			break;
1813 		/* Save bonding info. */
1814 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1815 			sizeof(bond_info->ports[0].ifname));
1816 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1817 		bond_info->ports[info.port_name].ifindex = ifindex;
1818 		bond_info->n_port++;
1819 		/*
1820 		 * Under socket direct mode, bonding will use
1821 		 * system_image_guid as identification.
1822 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1823 		 * All bonding members should have the same guid even if driver
1824 		 * is using PCIe BDF.
1825 		 */
1826 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1827 		if (ret < 0)
1828 			break;
1829 		else if (ret > 0) {
1830 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1831 			    owner == info.port_name &&
1832 			    (owner != 0 || (owner == 0 &&
1833 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1834 				pf = info.port_name;
1835 		} else if (pci_dev->domain == pci_addr.domain &&
1836 		    pci_dev->bus == pci_addr.bus &&
1837 		    pci_dev->devid == pci_addr.devid &&
1838 		    ((pci_dev->function == 0 &&
1839 		      pci_dev->function + owner == pci_addr.function) ||
1840 		     (pci_dev->function == owner &&
1841 		      pci_addr.function == owner)))
1842 			pf = info.port_name;
1843 	}
1844 	if (pf >= 0) {
1845 		/* Get bond interface info */
1846 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1847 					   bond_info->ifname);
1848 		if (ret)
1849 			DRV_LOG(ERR, "unable to get bond info: %s",
1850 				strerror(rte_errno));
1851 		else
1852 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1853 				ifindex, bond_info->ifindex, bond_info->ifname);
1854 	}
1855 	if (owner == 0 && pf != 0) {
1856 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
1857 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1858 				pci_dev->function);
1859 	}
1860 	return pf;
1861 }
1862 
1863 /**
1864  * Register a PCI device within bonding.
1865  *
1866  * This function spawns Ethernet devices out of a given PCI device and
1867  * bonding owner PF index.
1868  *
1869  * @param[in] cdev
1870  *   Pointer to common mlx5 device structure.
1871  * @param[in] req_eth_da
1872  *   Requested ethdev device argument.
1873  * @param[in] owner_id
1874  *   Requested owner PF port ID within bonding device, default to 0.
1875  * @param[in, out] mkvlist
1876  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1877  *
1878  * @return
1879  *   0 on success, a negative errno value otherwise and rte_errno is set.
1880  */
1881 static int
1882 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
1883 		     struct rte_eth_devargs *req_eth_da,
1884 		     uint16_t owner_id, struct mlx5_kvargs_ctrl *mkvlist)
1885 {
1886 	struct ibv_device **ibv_list;
1887 	/*
1888 	 * Number of found IB Devices matching with requested PCI BDF.
1889 	 * nd != 1 means there are multiple IB devices over the same
1890 	 * PCI device and we have representors and master.
1891 	 */
1892 	unsigned int nd = 0;
1893 	/*
1894 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1895 	 * we have the single multiport IB device, and there may be
1896 	 * representors attached to some of found ports.
1897 	 */
1898 	unsigned int np = 0;
1899 	/*
1900 	 * Number of DPDK ethernet devices to Spawn - either over
1901 	 * multiple IB devices or multiple ports of single IB device.
1902 	 * Actually this is the number of iterations to spawn.
1903 	 */
1904 	unsigned int ns = 0;
1905 	/*
1906 	 * Bonding device
1907 	 *   < 0 - no bonding device (single one)
1908 	 *  >= 0 - bonding device (value is slave PF index)
1909 	 */
1910 	int bd = -1;
1911 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
1912 	struct mlx5_dev_spawn_data *list = NULL;
1913 	struct rte_eth_devargs eth_da = *req_eth_da;
1914 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1915 	struct mlx5_bond_info bond_info;
1916 	int ret = -1;
1917 
1918 	errno = 0;
1919 	ibv_list = mlx5_glue->get_device_list(&ret);
1920 	if (!ibv_list) {
1921 		rte_errno = errno ? errno : ENOSYS;
1922 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
1923 		return -rte_errno;
1924 	}
1925 	/*
1926 	 * First scan the list of all Infiniband devices to find
1927 	 * matching ones, gathering into the list.
1928 	 */
1929 	struct ibv_device *ibv_match[ret + 1];
1930 	int nl_route = mlx5_nl_init(NETLINK_ROUTE, 0);
1931 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
1932 	unsigned int i;
1933 
1934 	while (ret-- > 0) {
1935 		struct rte_pci_addr pci_addr;
1936 
1937 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
1938 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
1939 						nl_rdma, owner_id, &bond_info);
1940 		if (bd >= 0) {
1941 			/*
1942 			 * Bonding device detected. Only one match is allowed,
1943 			 * the bonding is supported over multi-port IB device,
1944 			 * there should be no matches on representor PCI
1945 			 * functions or non VF LAG bonding devices with
1946 			 * specified address.
1947 			 */
1948 			if (nd) {
1949 				DRV_LOG(ERR,
1950 					"multiple PCI match on bonding device"
1951 					"\"%s\" found", ibv_list[ret]->name);
1952 				rte_errno = ENOENT;
1953 				ret = -rte_errno;
1954 				goto exit;
1955 			}
1956 			/* Amend owner pci address if owner PF ID specified. */
1957 			if (eth_da.nb_representor_ports)
1958 				owner_pci.function += owner_id;
1959 			DRV_LOG(INFO,
1960 				"PCI information matches for slave %d bonding device \"%s\"",
1961 				bd, ibv_list[ret]->name);
1962 			ibv_match[nd++] = ibv_list[ret];
1963 			break;
1964 		} else {
1965 			/* Bonding device not found. */
1966 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
1967 					      &pci_addr))
1968 				continue;
1969 			if (owner_pci.domain != pci_addr.domain ||
1970 			    owner_pci.bus != pci_addr.bus ||
1971 			    owner_pci.devid != pci_addr.devid ||
1972 			    owner_pci.function != pci_addr.function)
1973 				continue;
1974 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1975 				ibv_list[ret]->name);
1976 			ibv_match[nd++] = ibv_list[ret];
1977 		}
1978 	}
1979 	ibv_match[nd] = NULL;
1980 	if (!nd) {
1981 		/* No device matches, just complain and bail out. */
1982 		DRV_LOG(WARNING,
1983 			"PF %u doesn't have Verbs device matches PCI device " PCI_PRI_FMT ","
1984 			" are kernel drivers loaded?",
1985 			owner_id, owner_pci.domain, owner_pci.bus,
1986 			owner_pci.devid, owner_pci.function);
1987 		rte_errno = ENOENT;
1988 		ret = -rte_errno;
1989 		goto exit;
1990 	}
1991 	if (nd == 1) {
1992 		/*
1993 		 * Found single matching device may have multiple ports.
1994 		 * Each port may be representor, we have to check the port
1995 		 * number and check the representors existence.
1996 		 */
1997 		if (nl_rdma >= 0)
1998 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1999 		if (!np)
2000 			DRV_LOG(WARNING,
2001 				"Cannot get IB device \"%s\" ports number.",
2002 				ibv_match[0]->name);
2003 		if (bd >= 0 && !np) {
2004 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2005 			rte_errno = ENOENT;
2006 			ret = -rte_errno;
2007 			goto exit;
2008 		}
2009 	}
2010 	/* Now we can determine the maximal amount of devices to be spawned. */
2011 	list = mlx5_malloc(MLX5_MEM_ZERO,
2012 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2013 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2014 	if (!list) {
2015 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2016 		rte_errno = ENOMEM;
2017 		ret = -rte_errno;
2018 		goto exit;
2019 	}
2020 	if (bd >= 0 || np > 1) {
2021 		/*
2022 		 * Single IB device with multiple ports found,
2023 		 * it may be E-Switch master device and representors.
2024 		 * We have to perform identification through the ports.
2025 		 */
2026 		MLX5_ASSERT(nl_rdma >= 0);
2027 		MLX5_ASSERT(ns == 0);
2028 		MLX5_ASSERT(nd == 1);
2029 		MLX5_ASSERT(np);
2030 		for (i = 1; i <= np; ++i) {
2031 			list[ns].bond_info = &bond_info;
2032 			list[ns].max_port = np;
2033 			list[ns].phys_port = i;
2034 			list[ns].phys_dev_name = ibv_match[0]->name;
2035 			list[ns].eth_dev = NULL;
2036 			list[ns].pci_dev = pci_dev;
2037 			list[ns].cdev = cdev;
2038 			list[ns].pf_bond = bd;
2039 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2040 							   ibv_match[0]->name,
2041 							   i);
2042 			if (!list[ns].ifindex) {
2043 				/*
2044 				 * No network interface index found for the
2045 				 * specified port, it means there is no
2046 				 * representor on this port. It's OK,
2047 				 * there can be disabled ports, for example
2048 				 * if sriov_numvfs < sriov_totalvfs.
2049 				 */
2050 				continue;
2051 			}
2052 			ret = -1;
2053 			if (nl_route >= 0)
2054 				ret = mlx5_nl_switch_info(nl_route,
2055 							  list[ns].ifindex,
2056 							  &list[ns].info);
2057 			if (ret || (!list[ns].info.representor &&
2058 				    !list[ns].info.master)) {
2059 				/*
2060 				 * We failed to recognize representors with
2061 				 * Netlink, let's try to perform the task
2062 				 * with sysfs.
2063 				 */
2064 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2065 							     &list[ns].info);
2066 			}
2067 			if (!ret && bd >= 0) {
2068 				switch (list[ns].info.name_type) {
2069 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2070 					if (np == 1) {
2071 						/*
2072 						 * Force standalone bonding
2073 						 * device for ROCE LAG
2074 						 * configurations.
2075 						 */
2076 						list[ns].info.master = 0;
2077 						list[ns].info.representor = 0;
2078 					}
2079 					if (list[ns].info.port_name == bd)
2080 						ns++;
2081 					break;
2082 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2083 					/* Fallthrough */
2084 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2085 					/* Fallthrough */
2086 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2087 					if (list[ns].info.pf_num == bd)
2088 						ns++;
2089 					break;
2090 				default:
2091 					break;
2092 				}
2093 				continue;
2094 			}
2095 			if (!ret && (list[ns].info.representor ^
2096 				     list[ns].info.master))
2097 				ns++;
2098 		}
2099 		if (!ns) {
2100 			DRV_LOG(ERR,
2101 				"Unable to recognize master/representors on the IB device with multiple ports.");
2102 			rte_errno = ENOENT;
2103 			ret = -rte_errno;
2104 			goto exit;
2105 		}
2106 	} else {
2107 		/*
2108 		 * The existence of several matching entries (nd > 1) means
2109 		 * port representors have been instantiated. No existing Verbs
2110 		 * call nor sysfs entries can tell them apart, this can only
2111 		 * be done through Netlink calls assuming kernel drivers are
2112 		 * recent enough to support them.
2113 		 *
2114 		 * In the event of identification failure through Netlink,
2115 		 * try again through sysfs, then:
2116 		 *
2117 		 * 1. A single IB device matches (nd == 1) with single
2118 		 *    port (np=0/1) and is not a representor, assume
2119 		 *    no switch support.
2120 		 *
2121 		 * 2. Otherwise no safe assumptions can be made;
2122 		 *    complain louder and bail out.
2123 		 */
2124 		for (i = 0; i != nd; ++i) {
2125 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2126 			list[ns].bond_info = NULL;
2127 			list[ns].max_port = 1;
2128 			list[ns].phys_port = 1;
2129 			list[ns].phys_dev_name = ibv_match[i]->name;
2130 			list[ns].eth_dev = NULL;
2131 			list[ns].pci_dev = pci_dev;
2132 			list[ns].cdev = cdev;
2133 			list[ns].pf_bond = -1;
2134 			list[ns].ifindex = 0;
2135 			if (nl_rdma >= 0)
2136 				list[ns].ifindex = mlx5_nl_ifindex
2137 							    (nl_rdma,
2138 							     ibv_match[i]->name,
2139 							     1);
2140 			if (!list[ns].ifindex) {
2141 				char ifname[IF_NAMESIZE];
2142 
2143 				/*
2144 				 * Netlink failed, it may happen with old
2145 				 * ib_core kernel driver (before 4.16).
2146 				 * We can assume there is old driver because
2147 				 * here we are processing single ports IB
2148 				 * devices. Let's try sysfs to retrieve
2149 				 * the ifindex. The method works for
2150 				 * master device only.
2151 				 */
2152 				if (nd > 1) {
2153 					/*
2154 					 * Multiple devices found, assume
2155 					 * representors, can not distinguish
2156 					 * master/representor and retrieve
2157 					 * ifindex via sysfs.
2158 					 */
2159 					continue;
2160 				}
2161 				ret = mlx5_get_ifname_sysfs
2162 					(ibv_match[i]->ibdev_path, ifname);
2163 				if (!ret)
2164 					list[ns].ifindex =
2165 						if_nametoindex(ifname);
2166 				if (!list[ns].ifindex) {
2167 					/*
2168 					 * No network interface index found
2169 					 * for the specified device, it means
2170 					 * there it is neither representor
2171 					 * nor master.
2172 					 */
2173 					continue;
2174 				}
2175 			}
2176 			ret = -1;
2177 			if (nl_route >= 0)
2178 				ret = mlx5_nl_switch_info(nl_route,
2179 							  list[ns].ifindex,
2180 							  &list[ns].info);
2181 			if (ret || (!list[ns].info.representor &&
2182 				    !list[ns].info.master)) {
2183 				/*
2184 				 * We failed to recognize representors with
2185 				 * Netlink, let's try to perform the task
2186 				 * with sysfs.
2187 				 */
2188 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2189 							     &list[ns].info);
2190 			}
2191 			if (!ret && (list[ns].info.representor ^
2192 				     list[ns].info.master)) {
2193 				ns++;
2194 			} else if ((nd == 1) &&
2195 				   !list[ns].info.representor &&
2196 				   !list[ns].info.master) {
2197 				/*
2198 				 * Single IB device with one physical port and
2199 				 * attached network device.
2200 				 * May be SRIOV is not enabled or there is no
2201 				 * representors.
2202 				 */
2203 				DRV_LOG(INFO, "No E-Switch support detected.");
2204 				ns++;
2205 				break;
2206 			}
2207 		}
2208 		if (!ns) {
2209 			DRV_LOG(ERR,
2210 				"Unable to recognize master/representors on the multiple IB devices.");
2211 			rte_errno = ENOENT;
2212 			ret = -rte_errno;
2213 			goto exit;
2214 		}
2215 		/*
2216 		 * New kernels may add the switch_id attribute for the case
2217 		 * there is no E-Switch and we wrongly recognized the only
2218 		 * device as master. Override this if there is the single
2219 		 * device with single port and new device name format present.
2220 		 */
2221 		if (nd == 1 &&
2222 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2223 			list[0].info.master = 0;
2224 			list[0].info.representor = 0;
2225 		}
2226 	}
2227 	MLX5_ASSERT(ns);
2228 	/*
2229 	 * Sort list to probe devices in natural order for users convenience
2230 	 * (i.e. master first, then representors from lowest to highest ID).
2231 	 */
2232 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2233 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2234 		/* Set devargs default values. */
2235 		if (eth_da.nb_mh_controllers == 0) {
2236 			eth_da.nb_mh_controllers = 1;
2237 			eth_da.mh_controllers[0] = 0;
2238 		}
2239 		if (eth_da.nb_ports == 0 && ns > 0) {
2240 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2241 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2242 					pci_dev->device.devargs->args);
2243 			eth_da.nb_ports = 1;
2244 			eth_da.ports[0] = list[0].info.pf_num;
2245 		}
2246 		if (eth_da.nb_representor_ports == 0) {
2247 			eth_da.nb_representor_ports = 1;
2248 			eth_da.representor_ports[0] = 0;
2249 		}
2250 	}
2251 	for (i = 0; i != ns; ++i) {
2252 		uint32_t restore;
2253 
2254 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i], &eth_da,
2255 						 mkvlist);
2256 		if (!list[i].eth_dev) {
2257 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2258 				break;
2259 			/* Device is disabled or already spawned. Ignore it. */
2260 			continue;
2261 		}
2262 		restore = list[i].eth_dev->data->dev_flags;
2263 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2264 		/**
2265 		 * Each representor has a dedicated interrupts vector.
2266 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2267 		 * representor eth_dev object because representor and PF
2268 		 * share the same PCI address.
2269 		 * Override representor device with a dedicated
2270 		 * interrupts handle here.
2271 		 * Representor interrupts handle is released in mlx5_dev_stop().
2272 		 */
2273 		if (list[i].info.representor) {
2274 			struct rte_intr_handle *intr_handle =
2275 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2276 			if (intr_handle == NULL) {
2277 				DRV_LOG(ERR,
2278 					"port %u failed to allocate memory for interrupt handler "
2279 					"Rx interrupts will not be supported",
2280 					i);
2281 				rte_errno = ENOMEM;
2282 				ret = -rte_errno;
2283 				goto exit;
2284 			}
2285 			list[i].eth_dev->intr_handle = intr_handle;
2286 		}
2287 		/* Restore non-PCI flags cleared by the above call. */
2288 		list[i].eth_dev->data->dev_flags |= restore;
2289 		rte_eth_dev_probing_finish(list[i].eth_dev);
2290 	}
2291 	if (i != ns) {
2292 		DRV_LOG(ERR,
2293 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2294 			" encountering an error: %s",
2295 			owner_pci.domain, owner_pci.bus,
2296 			owner_pci.devid, owner_pci.function,
2297 			strerror(rte_errno));
2298 		ret = -rte_errno;
2299 		/* Roll back. */
2300 		while (i--) {
2301 			if (!list[i].eth_dev)
2302 				continue;
2303 			mlx5_dev_close(list[i].eth_dev);
2304 			/* mac_addrs must not be freed because in dev_private */
2305 			list[i].eth_dev->data->mac_addrs = NULL;
2306 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2307 		}
2308 		/* Restore original error. */
2309 		rte_errno = -ret;
2310 	} else {
2311 		ret = 0;
2312 	}
2313 exit:
2314 	/*
2315 	 * Do the routine cleanup:
2316 	 * - close opened Netlink sockets
2317 	 * - free allocated spawn data array
2318 	 * - free the Infiniband device list
2319 	 */
2320 	if (nl_rdma >= 0)
2321 		close(nl_rdma);
2322 	if (nl_route >= 0)
2323 		close(nl_route);
2324 	if (list)
2325 		mlx5_free(list);
2326 	MLX5_ASSERT(ibv_list);
2327 	mlx5_glue->free_device_list(ibv_list);
2328 	return ret;
2329 }
2330 
2331 static int
2332 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2333 			  struct rte_eth_devargs *eth_da)
2334 {
2335 	int ret = 0;
2336 
2337 	if (dev->devargs == NULL)
2338 		return 0;
2339 	memset(eth_da, 0, sizeof(*eth_da));
2340 	/* Parse representor information first from class argument. */
2341 	if (dev->devargs->cls_str)
2342 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2343 	if (ret != 0) {
2344 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2345 			dev->devargs->cls_str);
2346 		return -rte_errno;
2347 	}
2348 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2349 		/* Parse legacy device argument */
2350 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2351 		if (ret) {
2352 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2353 				dev->devargs->args);
2354 			return -rte_errno;
2355 		}
2356 	}
2357 	return 0;
2358 }
2359 
2360 /**
2361  * Callback to register a PCI device.
2362  *
2363  * This function spawns Ethernet devices out of a given PCI device.
2364  *
2365  * @param[in] cdev
2366  *   Pointer to common mlx5 device structure.
2367  * @param[in, out] mkvlist
2368  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2369  *
2370  * @return
2371  *   0 on success, a negative errno value otherwise and rte_errno is set.
2372  */
2373 static int
2374 mlx5_os_pci_probe(struct mlx5_common_device *cdev,
2375 		  struct mlx5_kvargs_ctrl *mkvlist)
2376 {
2377 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2378 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2379 	int ret = 0;
2380 	uint16_t p;
2381 
2382 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2383 	if (ret != 0)
2384 		return ret;
2385 
2386 	if (eth_da.nb_ports > 0) {
2387 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2388 		for (p = 0; p < eth_da.nb_ports; p++) {
2389 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2390 						   eth_da.ports[p], mkvlist);
2391 			if (ret) {
2392 				DRV_LOG(INFO, "Probe of PCI device " PCI_PRI_FMT " "
2393 					"aborted due to proding failure of PF %u",
2394 					pci_dev->addr.domain, pci_dev->addr.bus,
2395 					pci_dev->addr.devid, pci_dev->addr.function,
2396 					eth_da.ports[p]);
2397 				mlx5_net_remove(cdev);
2398 				if (p != 0)
2399 					break;
2400 			}
2401 		}
2402 	} else {
2403 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0, mkvlist);
2404 	}
2405 	return ret;
2406 }
2407 
2408 /* Probe a single SF device on auxiliary bus, no representor support. */
2409 static int
2410 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev,
2411 			struct mlx5_kvargs_ctrl *mkvlist)
2412 {
2413 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2414 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2415 	struct rte_device *dev = cdev->dev;
2416 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2417 	struct rte_eth_dev *eth_dev;
2418 	int ret = 0;
2419 
2420 	/* Parse ethdev devargs. */
2421 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2422 	if (ret != 0)
2423 		return ret;
2424 	/* Init spawn data. */
2425 	spawn.max_port = 1;
2426 	spawn.phys_port = 1;
2427 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2428 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2429 	if (ret < 0) {
2430 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2431 		return ret;
2432 	}
2433 	spawn.ifindex = ret;
2434 	spawn.cdev = cdev;
2435 	/* Spawn device. */
2436 	eth_dev = mlx5_dev_spawn(dev, &spawn, &eth_da, mkvlist);
2437 	if (eth_dev == NULL)
2438 		return -rte_errno;
2439 	/* Post create. */
2440 	eth_dev->intr_handle = adev->intr_handle;
2441 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2442 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2443 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2444 		eth_dev->data->numa_node = dev->numa_node;
2445 	}
2446 	rte_eth_dev_probing_finish(eth_dev);
2447 	return 0;
2448 }
2449 
2450 /**
2451  * Net class driver callback to probe a device.
2452  *
2453  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2454  *
2455  * @param[in] cdev
2456  *   Pointer to the common mlx5 device.
2457  * @param[in, out] mkvlist
2458  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2459  *
2460  * @return
2461  *   0 on success, a negative errno value otherwise and rte_errno is set.
2462  */
2463 int
2464 mlx5_os_net_probe(struct mlx5_common_device *cdev,
2465 		  struct mlx5_kvargs_ctrl *mkvlist)
2466 {
2467 	int ret;
2468 
2469 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2470 		mlx5_pmd_socket_init();
2471 	ret = mlx5_init_once();
2472 	if (ret) {
2473 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2474 			strerror(rte_errno));
2475 		return -rte_errno;
2476 	}
2477 	ret = mlx5_probe_again_args_validate(cdev, mkvlist);
2478 	if (ret) {
2479 		DRV_LOG(ERR, "Probe again parameters are not compatible : %s",
2480 			strerror(rte_errno));
2481 		return -rte_errno;
2482 	}
2483 	if (mlx5_dev_is_pci(cdev->dev))
2484 		return mlx5_os_pci_probe(cdev, mkvlist);
2485 	else
2486 		return mlx5_os_auxiliary_probe(cdev, mkvlist);
2487 }
2488 
2489 /**
2490  * Cleanup resources when the last device is closed.
2491  */
2492 void
2493 mlx5_os_net_cleanup(void)
2494 {
2495 	mlx5_pmd_socket_uninit();
2496 }
2497 
2498 /**
2499  * Install shared asynchronous device events handler.
2500  * This function is implemented to support event sharing
2501  * between multiple ports of single IB device.
2502  *
2503  * @param sh
2504  *   Pointer to mlx5_dev_ctx_shared object.
2505  */
2506 void
2507 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2508 {
2509 	struct ibv_context *ctx = sh->cdev->ctx;
2510 	int nlsk_fd;
2511 
2512 	sh->intr_handle = mlx5_os_interrupt_handler_create
2513 		(RTE_INTR_INSTANCE_F_SHARED, true,
2514 		 ctx->async_fd, mlx5_dev_interrupt_handler, sh);
2515 	if (!sh->intr_handle) {
2516 		DRV_LOG(ERR, "Failed to allocate intr_handle.");
2517 		return;
2518 	}
2519 	nlsk_fd = mlx5_nl_init(NETLINK_ROUTE, RTMGRP_LINK);
2520 	if (nlsk_fd < 0) {
2521 		DRV_LOG(ERR, "Failed to create a socket for Netlink events: %s",
2522 			rte_strerror(rte_errno));
2523 		return;
2524 	}
2525 	sh->intr_handle_nl = mlx5_os_interrupt_handler_create
2526 		(RTE_INTR_INSTANCE_F_SHARED, true,
2527 		 nlsk_fd, mlx5_dev_interrupt_handler_nl, sh);
2528 	if (sh->intr_handle_nl == NULL) {
2529 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2530 		return;
2531 	}
2532 	if (sh->cdev->config.devx) {
2533 #ifdef HAVE_IBV_DEVX_ASYNC
2534 		struct mlx5dv_devx_cmd_comp *devx_comp;
2535 
2536 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2537 		devx_comp = sh->devx_comp;
2538 		if (!devx_comp) {
2539 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2540 			return;
2541 		}
2542 		sh->intr_handle_devx = mlx5_os_interrupt_handler_create
2543 			(RTE_INTR_INSTANCE_F_SHARED, true,
2544 			 devx_comp->fd,
2545 			 mlx5_dev_interrupt_handler_devx, sh);
2546 		if (!sh->intr_handle_devx) {
2547 			DRV_LOG(ERR, "Failed to allocate intr_handle.");
2548 			return;
2549 		}
2550 #endif /* HAVE_IBV_DEVX_ASYNC */
2551 	}
2552 }
2553 
2554 /**
2555  * Uninstall shared asynchronous device events handler.
2556  * This function is implemented to support event sharing
2557  * between multiple ports of single IB device.
2558  *
2559  * @param dev
2560  *   Pointer to mlx5_dev_ctx_shared object.
2561  */
2562 void
2563 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2564 {
2565 	mlx5_os_interrupt_handler_destroy(sh->intr_handle,
2566 					  mlx5_dev_interrupt_handler, sh);
2567 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_nl,
2568 					  mlx5_dev_interrupt_handler_nl, sh);
2569 #ifdef HAVE_IBV_DEVX_ASYNC
2570 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_devx,
2571 					  mlx5_dev_interrupt_handler_devx, sh);
2572 	if (sh->devx_comp)
2573 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2574 #endif
2575 }
2576 
2577 /**
2578  * Read statistics by a named counter.
2579  *
2580  * @param[in] priv
2581  *   Pointer to the private device data structure.
2582  * @param[in] ctr_name
2583  *   Pointer to the name of the statistic counter to read
2584  * @param[out] stat
2585  *   Pointer to read statistic value.
2586  * @return
2587  *   0 on success and stat is valud, 1 if failed to read the value
2588  *   rte_errno is set.
2589  *
2590  */
2591 int
2592 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2593 		      uint64_t *stat)
2594 {
2595 	int fd;
2596 
2597 	if (priv->sh) {
2598 		if (priv->q_counters != NULL &&
2599 		    strcmp(ctr_name, "out_of_buffer") == 0)
2600 			return mlx5_devx_cmd_queue_counter_query
2601 					(priv->q_counters, 0, (uint32_t *)stat);
2602 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2603 		      priv->sh->ibdev_path,
2604 		      priv->dev_port,
2605 		      ctr_name);
2606 		fd = open(path, O_RDONLY);
2607 		/*
2608 		 * in switchdev the file location is not per port
2609 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2610 		 */
2611 		if (fd == -1) {
2612 			MKSTR(path1, "%s/hw_counters/%s",
2613 			      priv->sh->ibdev_path,
2614 			      ctr_name);
2615 			fd = open(path1, O_RDONLY);
2616 		}
2617 		if (fd != -1) {
2618 			char buf[21] = {'\0'};
2619 			ssize_t n = read(fd, buf, sizeof(buf));
2620 
2621 			close(fd);
2622 			if (n != -1) {
2623 				*stat = strtoull(buf, NULL, 10);
2624 				return 0;
2625 			}
2626 		}
2627 	}
2628 	*stat = 0;
2629 	return 1;
2630 }
2631 
2632 /**
2633  * Remove a MAC address from device
2634  *
2635  * @param dev
2636  *   Pointer to Ethernet device structure.
2637  * @param index
2638  *   MAC address index.
2639  */
2640 void
2641 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2642 {
2643 	struct mlx5_priv *priv = dev->data->dev_private;
2644 	const int vf = priv->sh->dev_cap.vf;
2645 
2646 	if (vf)
2647 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2648 					mlx5_ifindex(dev), priv->mac_own,
2649 					&dev->data->mac_addrs[index], index);
2650 }
2651 
2652 /**
2653  * Adds a MAC address to the device
2654  *
2655  * @param dev
2656  *   Pointer to Ethernet device structure.
2657  * @param mac_addr
2658  *   MAC address to register.
2659  * @param index
2660  *   MAC address index.
2661  *
2662  * @return
2663  *   0 on success, a negative errno value otherwise
2664  */
2665 int
2666 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2667 		     uint32_t index)
2668 {
2669 	struct mlx5_priv *priv = dev->data->dev_private;
2670 	const int vf = priv->sh->dev_cap.vf;
2671 	int ret = 0;
2672 
2673 	if (vf)
2674 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2675 					   mlx5_ifindex(dev), priv->mac_own,
2676 					   mac, index);
2677 	return ret;
2678 }
2679 
2680 /**
2681  * Modify a VF MAC address
2682  *
2683  * @param priv
2684  *   Pointer to device private data.
2685  * @param mac_addr
2686  *   MAC address to modify into.
2687  * @param iface_idx
2688  *   Net device interface index
2689  * @param vf_index
2690  *   VF index
2691  *
2692  * @return
2693  *   0 on success, a negative errno value otherwise
2694  */
2695 int
2696 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2697 			   unsigned int iface_idx,
2698 			   struct rte_ether_addr *mac_addr,
2699 			   int vf_index)
2700 {
2701 	return mlx5_nl_vf_mac_addr_modify
2702 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2703 }
2704 
2705 /**
2706  * Set device promiscuous mode
2707  *
2708  * @param dev
2709  *   Pointer to Ethernet device structure.
2710  * @param enable
2711  *   0 - promiscuous is disabled, otherwise - enabled
2712  *
2713  * @return
2714  *   0 on success, a negative error value otherwise
2715  */
2716 int
2717 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2718 {
2719 	struct mlx5_priv *priv = dev->data->dev_private;
2720 
2721 	return mlx5_nl_promisc(priv->nl_socket_route,
2722 			       mlx5_ifindex(dev), !!enable);
2723 }
2724 
2725 /**
2726  * Set device promiscuous mode
2727  *
2728  * @param dev
2729  *   Pointer to Ethernet device structure.
2730  * @param enable
2731  *   0 - all multicase is disabled, otherwise - enabled
2732  *
2733  * @return
2734  *   0 on success, a negative error value otherwise
2735  */
2736 int
2737 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2738 {
2739 	struct mlx5_priv *priv = dev->data->dev_private;
2740 
2741 	return mlx5_nl_allmulti(priv->nl_socket_route,
2742 				mlx5_ifindex(dev), !!enable);
2743 }
2744 
2745 /**
2746  * Flush device MAC addresses
2747  *
2748  * @param dev
2749  *   Pointer to Ethernet device structure.
2750  *
2751  */
2752 void
2753 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2754 {
2755 	struct mlx5_priv *priv = dev->data->dev_private;
2756 
2757 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2758 			       dev->data->mac_addrs,
2759 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2760 }
2761