xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision c56185fc183fc0532d2f03aaf04bbf0989ea91a5)
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 	sh->dev_cap.scatter_fcs_w_decap_disable =
374 					hca_attr->scatter_fcs_w_decap_disable;
375 	sh->dev_cap.rq_delay_drop_en = hca_attr->rq_delay_drop;
376 	mlx5_rt_timestamp_config(sh, hca_attr);
377 	return 0;
378 }
379 
380 /**
381  * Detect misc5 support or not
382  *
383  * @param[in] priv
384  *   Device private data pointer
385  */
386 #ifdef HAVE_MLX5DV_DR
387 static void
388 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
389 {
390 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
391 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
392 	 * Case: IPv4--->UDP--->VxLAN--->vni
393 	 */
394 	void *tbl;
395 	struct mlx5_flow_dv_match_params matcher_mask;
396 	void *match_m;
397 	void *matcher;
398 	void *headers_m;
399 	void *misc5_m;
400 	uint32_t *tunnel_header_m;
401 	struct mlx5dv_flow_matcher_attr dv_attr;
402 
403 	memset(&matcher_mask, 0, sizeof(matcher_mask));
404 	matcher_mask.size = sizeof(matcher_mask.buf);
405 	match_m = matcher_mask.buf;
406 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
407 	misc5_m = MLX5_ADDR_OF(fte_match_param,
408 			       match_m, misc_parameters_5);
409 	tunnel_header_m = (uint32_t *)
410 				MLX5_ADDR_OF(fte_match_set_misc5,
411 				misc5_m, tunnel_header_1);
412 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
413 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
414 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
415 	*tunnel_header_m = 0xffffff;
416 
417 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
418 	if (!tbl) {
419 		DRV_LOG(INFO, "No SW steering support");
420 		return;
421 	}
422 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
423 	dv_attr.match_mask = (void *)&matcher_mask,
424 	dv_attr.match_criteria_enable =
425 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
426 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
427 	dv_attr.priority = 3;
428 #ifdef HAVE_MLX5DV_DR_ESWITCH
429 	void *misc2_m;
430 	if (priv->sh->config.dv_esw_en) {
431 		/* FDB enabled reg_c_0 */
432 		dv_attr.match_criteria_enable |=
433 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
434 		misc2_m = MLX5_ADDR_OF(fte_match_param,
435 				       match_m, misc_parameters_2);
436 		MLX5_SET(fte_match_set_misc2, misc2_m,
437 			 metadata_reg_c_0, 0xffff);
438 	}
439 #endif
440 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
441 						    &dv_attr, tbl);
442 	if (matcher) {
443 		priv->sh->misc5_cap = 1;
444 		mlx5_glue->dv_destroy_flow_matcher(matcher);
445 	}
446 	mlx5_glue->dr_destroy_flow_tbl(tbl);
447 #else
448 	RTE_SET_USED(priv);
449 #endif
450 }
451 #endif
452 
453 /**
454  * Initialize DR related data within private structure.
455  * Routine checks the reference counter and does actual
456  * resources creation/initialization only if counter is zero.
457  *
458  * @param[in] priv
459  *   Pointer to the private device data structure.
460  *
461  * @return
462  *   Zero on success, positive error code otherwise.
463  */
464 static int
465 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
466 {
467 	struct mlx5_dev_ctx_shared *sh = priv->sh;
468 	char s[MLX5_NAME_SIZE] __rte_unused;
469 	int err;
470 
471 	MLX5_ASSERT(sh && sh->refcnt);
472 	if (sh->refcnt > 1)
473 		return 0;
474 	err = mlx5_alloc_table_hash_list(priv);
475 	if (err)
476 		goto error;
477 	/* The resources below are only valid with DV support. */
478 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
479 	/* Init shared flex parsers list, no need lcore_share */
480 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
481 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
482 					       mlx5_flex_parser_create_cb,
483 					       mlx5_flex_parser_match_cb,
484 					       mlx5_flex_parser_remove_cb,
485 					       mlx5_flex_parser_clone_cb,
486 					       mlx5_flex_parser_clone_free_cb);
487 	if (!sh->flex_parsers_dv)
488 		goto error;
489 	if (priv->sh->config.dv_flow_en == 2)
490 		return 0;
491 	/* Init port id action list. */
492 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
493 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
494 						   flow_dv_port_id_create_cb,
495 						   flow_dv_port_id_match_cb,
496 						   flow_dv_port_id_remove_cb,
497 						   flow_dv_port_id_clone_cb,
498 						 flow_dv_port_id_clone_free_cb);
499 	if (!sh->port_id_action_list)
500 		goto error;
501 	/* Init push vlan action list. */
502 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
503 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
504 						    flow_dv_push_vlan_create_cb,
505 						    flow_dv_push_vlan_match_cb,
506 						    flow_dv_push_vlan_remove_cb,
507 						    flow_dv_push_vlan_clone_cb,
508 					       flow_dv_push_vlan_clone_free_cb);
509 	if (!sh->push_vlan_action_list)
510 		goto error;
511 	/* Init sample action list. */
512 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
513 	sh->sample_action_list = mlx5_list_create(s, sh, true,
514 						  flow_dv_sample_create_cb,
515 						  flow_dv_sample_match_cb,
516 						  flow_dv_sample_remove_cb,
517 						  flow_dv_sample_clone_cb,
518 						  flow_dv_sample_clone_free_cb);
519 	if (!sh->sample_action_list)
520 		goto error;
521 	/* Init dest array action list. */
522 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
523 	sh->dest_array_list = mlx5_list_create(s, sh, true,
524 					       flow_dv_dest_array_create_cb,
525 					       flow_dv_dest_array_match_cb,
526 					       flow_dv_dest_array_remove_cb,
527 					       flow_dv_dest_array_clone_cb,
528 					      flow_dv_dest_array_clone_free_cb);
529 	if (!sh->dest_array_list)
530 		goto error;
531 #else
532 	if (priv->sh->config.dv_flow_en == 2)
533 		return 0;
534 #endif
535 #ifdef HAVE_MLX5DV_DR
536 	void *domain;
537 
538 	/* Reference counter is zero, we should initialize structures. */
539 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
540 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
541 	if (!domain) {
542 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
543 		err = errno;
544 		goto error;
545 	}
546 	sh->rx_domain = domain;
547 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
548 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
549 	if (!domain) {
550 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
551 		err = errno;
552 		goto error;
553 	}
554 	sh->tx_domain = domain;
555 #ifdef HAVE_MLX5DV_DR_ESWITCH
556 	if (sh->config.dv_esw_en) {
557 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
558 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
559 		if (!domain) {
560 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
561 			err = errno;
562 			goto error;
563 		}
564 		sh->fdb_domain = domain;
565 	}
566 	/*
567 	 * The drop action is just some dummy placeholder in rdma-core. It
568 	 * does not belong to domains and has no any attributes, and, can be
569 	 * shared by the entire device.
570 	 */
571 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
572 	if (!sh->dr_drop_action) {
573 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
574 		err = errno;
575 		goto error;
576 	}
577 #endif
578 	if (!sh->tunnel_hub && sh->config.dv_miss_info)
579 		err = mlx5_alloc_tunnel_hub(sh);
580 	if (err) {
581 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
582 		goto error;
583 	}
584 	if (sh->config.reclaim_mode == MLX5_RCM_AGGR) {
585 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
586 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
587 		if (sh->fdb_domain)
588 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
589 	}
590 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
591 	if (!sh->config.allow_duplicate_pattern) {
592 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
593 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
594 #endif
595 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
596 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
597 		if (sh->fdb_domain)
598 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
599 	}
600 
601 	__mlx5_discovery_misc5_cap(priv);
602 #endif /* HAVE_MLX5DV_DR */
603 	sh->default_miss_action =
604 			mlx5_glue->dr_create_flow_action_default_miss();
605 	if (!sh->default_miss_action)
606 		DRV_LOG(WARNING, "Default miss action is not supported.");
607 	LIST_INIT(&sh->shared_rxqs);
608 	return 0;
609 error:
610 	/* Rollback the created objects. */
611 	if (sh->rx_domain) {
612 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
613 		sh->rx_domain = NULL;
614 	}
615 	if (sh->tx_domain) {
616 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
617 		sh->tx_domain = NULL;
618 	}
619 	if (sh->fdb_domain) {
620 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
621 		sh->fdb_domain = NULL;
622 	}
623 	if (sh->dr_drop_action) {
624 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
625 		sh->dr_drop_action = NULL;
626 	}
627 	if (sh->pop_vlan_action) {
628 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
629 		sh->pop_vlan_action = NULL;
630 	}
631 	if (sh->encaps_decaps) {
632 		mlx5_hlist_destroy(sh->encaps_decaps);
633 		sh->encaps_decaps = NULL;
634 	}
635 	if (sh->modify_cmds) {
636 		mlx5_hlist_destroy(sh->modify_cmds);
637 		sh->modify_cmds = NULL;
638 	}
639 	if (sh->tag_table) {
640 		/* tags should be destroyed with flow before. */
641 		mlx5_hlist_destroy(sh->tag_table);
642 		sh->tag_table = NULL;
643 	}
644 	if (sh->tunnel_hub) {
645 		mlx5_release_tunnel_hub(sh, priv->dev_port);
646 		sh->tunnel_hub = NULL;
647 	}
648 	mlx5_free_table_hash_list(priv);
649 	if (sh->port_id_action_list) {
650 		mlx5_list_destroy(sh->port_id_action_list);
651 		sh->port_id_action_list = NULL;
652 	}
653 	if (sh->push_vlan_action_list) {
654 		mlx5_list_destroy(sh->push_vlan_action_list);
655 		sh->push_vlan_action_list = NULL;
656 	}
657 	if (sh->sample_action_list) {
658 		mlx5_list_destroy(sh->sample_action_list);
659 		sh->sample_action_list = NULL;
660 	}
661 	if (sh->dest_array_list) {
662 		mlx5_list_destroy(sh->dest_array_list);
663 		sh->dest_array_list = NULL;
664 	}
665 	return err;
666 }
667 
668 /**
669  * Destroy DR related data within private structure.
670  *
671  * @param[in] priv
672  *   Pointer to the private device data structure.
673  */
674 void
675 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
676 {
677 	struct mlx5_dev_ctx_shared *sh = priv->sh;
678 
679 	MLX5_ASSERT(sh && sh->refcnt);
680 	if (sh->refcnt > 1)
681 		return;
682 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
683 #ifdef HAVE_MLX5DV_DR
684 	if (sh->rx_domain) {
685 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
686 		sh->rx_domain = NULL;
687 	}
688 	if (sh->tx_domain) {
689 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
690 		sh->tx_domain = NULL;
691 	}
692 #ifdef HAVE_MLX5DV_DR_ESWITCH
693 	if (sh->fdb_domain) {
694 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
695 		sh->fdb_domain = NULL;
696 	}
697 	if (sh->dr_drop_action) {
698 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
699 		sh->dr_drop_action = NULL;
700 	}
701 #endif
702 	if (sh->pop_vlan_action) {
703 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
704 		sh->pop_vlan_action = NULL;
705 	}
706 	if (sh->send_to_kernel_action.action) {
707 		void *action = sh->send_to_kernel_action.action;
708 
709 		mlx5_glue->destroy_flow_action(action);
710 		sh->send_to_kernel_action.action = NULL;
711 	}
712 	if (sh->send_to_kernel_action.tbl) {
713 		struct mlx5_flow_tbl_resource *tbl =
714 				sh->send_to_kernel_action.tbl;
715 
716 		flow_dv_tbl_resource_release(sh, tbl);
717 		sh->send_to_kernel_action.tbl = NULL;
718 	}
719 #endif /* HAVE_MLX5DV_DR */
720 	if (sh->default_miss_action)
721 		mlx5_glue->destroy_flow_action
722 				(sh->default_miss_action);
723 	if (sh->encaps_decaps) {
724 		mlx5_hlist_destroy(sh->encaps_decaps);
725 		sh->encaps_decaps = NULL;
726 	}
727 	if (sh->modify_cmds) {
728 		mlx5_hlist_destroy(sh->modify_cmds);
729 		sh->modify_cmds = NULL;
730 	}
731 	if (sh->tag_table) {
732 		/* tags should be destroyed with flow before. */
733 		mlx5_hlist_destroy(sh->tag_table);
734 		sh->tag_table = NULL;
735 	}
736 	if (sh->tunnel_hub) {
737 		mlx5_release_tunnel_hub(sh, priv->dev_port);
738 		sh->tunnel_hub = NULL;
739 	}
740 	mlx5_free_table_hash_list(priv);
741 	if (sh->port_id_action_list) {
742 		mlx5_list_destroy(sh->port_id_action_list);
743 		sh->port_id_action_list = NULL;
744 	}
745 	if (sh->push_vlan_action_list) {
746 		mlx5_list_destroy(sh->push_vlan_action_list);
747 		sh->push_vlan_action_list = NULL;
748 	}
749 	if (sh->sample_action_list) {
750 		mlx5_list_destroy(sh->sample_action_list);
751 		sh->sample_action_list = NULL;
752 	}
753 	if (sh->dest_array_list) {
754 		mlx5_list_destroy(sh->dest_array_list);
755 		sh->dest_array_list = NULL;
756 	}
757 }
758 
759 /**
760  * Initialize shared data between primary and secondary process.
761  *
762  * A memzone is reserved by primary process and secondary processes attach to
763  * the memzone.
764  *
765  * @return
766  *   0 on success, a negative errno value otherwise and rte_errno is set.
767  */
768 static int
769 mlx5_init_shared_data(void)
770 {
771 	const struct rte_memzone *mz;
772 	int ret = 0;
773 
774 	rte_spinlock_lock(&mlx5_shared_data_lock);
775 	if (mlx5_shared_data == NULL) {
776 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
777 			/* Allocate shared memory. */
778 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
779 						 sizeof(*mlx5_shared_data),
780 						 SOCKET_ID_ANY, 0);
781 			if (mz == NULL) {
782 				DRV_LOG(ERR,
783 					"Cannot allocate mlx5 shared data");
784 				ret = -rte_errno;
785 				goto error;
786 			}
787 			mlx5_shared_data = mz->addr;
788 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
789 			rte_spinlock_init(&mlx5_shared_data->lock);
790 		} else {
791 			/* Lookup allocated shared memory. */
792 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
793 			if (mz == NULL) {
794 				DRV_LOG(ERR,
795 					"Cannot attach mlx5 shared data");
796 				ret = -rte_errno;
797 				goto error;
798 			}
799 			mlx5_shared_data = mz->addr;
800 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
801 		}
802 	}
803 error:
804 	rte_spinlock_unlock(&mlx5_shared_data_lock);
805 	return ret;
806 }
807 
808 /**
809  * PMD global initialization.
810  *
811  * Independent from individual device, this function initializes global
812  * per-PMD data structures distinguishing primary and secondary processes.
813  * Hence, each initialization is called once per a process.
814  *
815  * @return
816  *   0 on success, a negative errno value otherwise and rte_errno is set.
817  */
818 static int
819 mlx5_init_once(void)
820 {
821 	struct mlx5_shared_data *sd;
822 	struct mlx5_local_data *ld = &mlx5_local_data;
823 	int ret = 0;
824 
825 	if (mlx5_init_shared_data())
826 		return -rte_errno;
827 	sd = mlx5_shared_data;
828 	MLX5_ASSERT(sd);
829 	rte_spinlock_lock(&sd->lock);
830 	switch (rte_eal_process_type()) {
831 	case RTE_PROC_PRIMARY:
832 		if (sd->init_done)
833 			break;
834 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
835 					   mlx5_mp_os_primary_handle);
836 		if (ret)
837 			goto out;
838 		sd->init_done = true;
839 		break;
840 	case RTE_PROC_SECONDARY:
841 		if (ld->init_done)
842 			break;
843 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
844 					     mlx5_mp_os_secondary_handle);
845 		if (ret)
846 			goto out;
847 		++sd->secondary_cnt;
848 		ld->init_done = true;
849 		break;
850 	default:
851 		break;
852 	}
853 out:
854 	rte_spinlock_unlock(&sd->lock);
855 	return ret;
856 }
857 
858 /**
859  * DR flow drop action support detect.
860  *
861  * @param dev
862  *   Pointer to rte_eth_dev structure.
863  *
864  */
865 static void
866 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
867 {
868 #ifdef HAVE_MLX5DV_DR
869 	struct mlx5_priv *priv = dev->data->dev_private;
870 
871 	if (!priv->sh->config.dv_flow_en || !priv->sh->dr_drop_action)
872 		return;
873 	/**
874 	 * DR supports drop action placeholder when it is supported;
875 	 * otherwise, use the queue drop action.
876 	 */
877 	if (!priv->sh->drop_action_check_flag) {
878 		if (!mlx5_flow_discover_dr_action_support(dev))
879 			priv->sh->dr_root_drop_action_en = 1;
880 		priv->sh->drop_action_check_flag = 1;
881 	}
882 	if (priv->sh->dr_root_drop_action_en)
883 		priv->root_drop_action = priv->sh->dr_drop_action;
884 	else
885 		priv->root_drop_action = priv->drop_queue.hrxq->action;
886 #endif
887 }
888 
889 static void
890 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
891 {
892 	struct mlx5_priv *priv = dev->data->dev_private;
893 	void *ctx = priv->sh->cdev->ctx;
894 
895 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
896 	if (!priv->q_counters) {
897 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
898 		struct ibv_wq *wq;
899 
900 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
901 			"by DevX - fall-back to use the kernel driver global "
902 			"queue counter.", dev->data->port_id);
903 		/* Create WQ by kernel and query its queue counter ID. */
904 		if (cq) {
905 			wq = mlx5_glue->create_wq(ctx,
906 						  &(struct ibv_wq_init_attr){
907 						    .wq_type = IBV_WQT_RQ,
908 						    .max_wr = 1,
909 						    .max_sge = 1,
910 						    .pd = priv->sh->cdev->pd,
911 						    .cq = cq,
912 						});
913 			if (wq) {
914 				/* Counter is assigned only on RDY state. */
915 				int ret = mlx5_glue->modify_wq(wq,
916 						 &(struct ibv_wq_attr){
917 						 .attr_mask = IBV_WQ_ATTR_STATE,
918 						 .wq_state = IBV_WQS_RDY,
919 						});
920 
921 				if (ret == 0)
922 					mlx5_devx_cmd_wq_query(wq,
923 							 &priv->counter_set_id);
924 				claim_zero(mlx5_glue->destroy_wq(wq));
925 			}
926 			claim_zero(mlx5_glue->destroy_cq(cq));
927 		}
928 	} else {
929 		priv->counter_set_id = priv->q_counters->id;
930 	}
931 	if (priv->counter_set_id == 0)
932 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
933 			"available.", dev->data->port_id);
934 }
935 
936 /**
937  * Check if representor spawn info match devargs.
938  *
939  * @param spawn
940  *   Verbs device parameters (name, port, switch_info) to spawn.
941  * @param eth_da
942  *   Device devargs to probe.
943  *
944  * @return
945  *   Match result.
946  */
947 static bool
948 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
949 		       struct rte_eth_devargs *eth_da)
950 {
951 	struct mlx5_switch_info *switch_info = &spawn->info;
952 	unsigned int p, f;
953 	uint16_t id;
954 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
955 						      eth_da->type);
956 
957 	switch (eth_da->type) {
958 	case RTE_ETH_REPRESENTOR_SF:
959 		if (!(spawn->info.port_name == -1 &&
960 		      switch_info->name_type ==
961 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
962 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
963 			rte_errno = EBUSY;
964 			return false;
965 		}
966 		break;
967 	case RTE_ETH_REPRESENTOR_VF:
968 		/* Allows HPF representor index -1 as exception. */
969 		if (!(spawn->info.port_name == -1 &&
970 		      switch_info->name_type ==
971 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
972 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
973 			rte_errno = EBUSY;
974 			return false;
975 		}
976 		break;
977 	case RTE_ETH_REPRESENTOR_NONE:
978 		rte_errno = EBUSY;
979 		return false;
980 	default:
981 		rte_errno = ENOTSUP;
982 		DRV_LOG(ERR, "unsupported representor type");
983 		return false;
984 	}
985 	/* Check representor ID: */
986 	for (p = 0; p < eth_da->nb_ports; ++p) {
987 		if (spawn->pf_bond < 0) {
988 			/* For non-LAG mode, allow and ignore pf. */
989 			switch_info->pf_num = eth_da->ports[p];
990 			repr_id = mlx5_representor_id_encode(switch_info,
991 							     eth_da->type);
992 		}
993 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
994 			id = MLX5_REPRESENTOR_ID
995 				(eth_da->ports[p], eth_da->type,
996 				 eth_da->representor_ports[f]);
997 			if (repr_id == id)
998 				return true;
999 		}
1000 	}
1001 	rte_errno = EBUSY;
1002 	return false;
1003 }
1004 
1005 /**
1006  * Spawn an Ethernet device from Verbs information.
1007  *
1008  * @param dpdk_dev
1009  *   Backing DPDK device.
1010  * @param spawn
1011  *   Verbs device parameters (name, port, switch_info) to spawn.
1012  * @param eth_da
1013  *   Device arguments.
1014  * @param mkvlist
1015  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1016  *
1017  * @return
1018  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1019  *   is set. The following errors are defined:
1020  *
1021  *   EBUSY: device is not supposed to be spawned.
1022  *   EEXIST: device is already spawned
1023  */
1024 static struct rte_eth_dev *
1025 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1026 	       struct mlx5_dev_spawn_data *spawn,
1027 	       struct rte_eth_devargs *eth_da,
1028 	       struct mlx5_kvargs_ctrl *mkvlist)
1029 {
1030 	const struct mlx5_switch_info *switch_info = &spawn->info;
1031 	struct mlx5_dev_ctx_shared *sh = NULL;
1032 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
1033 	struct rte_eth_dev *eth_dev = NULL;
1034 	struct mlx5_priv *priv = NULL;
1035 	int err = 0;
1036 	struct rte_ether_addr mac;
1037 	char name[RTE_ETH_NAME_MAX_LEN];
1038 	int own_domain_id = 0;
1039 	uint16_t port_id;
1040 	struct mlx5_port_info vport_info = { .query_flags = 0 };
1041 	int nl_rdma;
1042 	int i;
1043 
1044 	/* Determine if this port representor is supposed to be spawned. */
1045 	if (switch_info->representor && dpdk_dev->devargs &&
1046 	    !mlx5_representor_match(spawn, eth_da))
1047 		return NULL;
1048 	/* Build device name. */
1049 	if (spawn->pf_bond < 0) {
1050 		/* Single device. */
1051 		if (!switch_info->representor)
1052 			strlcpy(name, dpdk_dev->name, sizeof(name));
1053 		else
1054 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
1055 				 dpdk_dev->name,
1056 				 switch_info->name_type ==
1057 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1058 				 switch_info->port_name);
1059 	} else {
1060 		/* Bonding device. */
1061 		if (!switch_info->representor) {
1062 			err = snprintf(name, sizeof(name), "%s_%s",
1063 				       dpdk_dev->name, spawn->phys_dev_name);
1064 		} else {
1065 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1066 				dpdk_dev->name, spawn->phys_dev_name,
1067 				switch_info->ctrl_num,
1068 				switch_info->pf_num,
1069 				switch_info->name_type ==
1070 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1071 				switch_info->port_name);
1072 		}
1073 	}
1074 	if (err >= (int)sizeof(name))
1075 		DRV_LOG(WARNING, "device name overflow %s", name);
1076 	/* check if the device is already spawned */
1077 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1078 		/*
1079 		 * When device is already spawned, its devargs should be set
1080 		 * as used. otherwise, mlx5_kvargs_validate() will fail.
1081 		 */
1082 		if (mkvlist)
1083 			mlx5_port_args_set_used(name, port_id, mkvlist);
1084 		rte_errno = EEXIST;
1085 		return NULL;
1086 	}
1087 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1088 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1089 		struct mlx5_mp_id mp_id;
1090 		int fd;
1091 
1092 		eth_dev = rte_eth_dev_attach_secondary(name);
1093 		if (eth_dev == NULL) {
1094 			DRV_LOG(ERR, "can not attach rte ethdev");
1095 			rte_errno = ENOMEM;
1096 			return NULL;
1097 		}
1098 		eth_dev->device = dpdk_dev;
1099 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1100 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1101 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1102 		err = mlx5_proc_priv_init(eth_dev);
1103 		if (err)
1104 			return NULL;
1105 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1106 		/* Receive command fd from primary process */
1107 		fd = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1108 		if (fd < 0)
1109 			goto err_secondary;
1110 		/* Remap UAR for Tx queues. */
1111 		err = mlx5_tx_uar_init_secondary(eth_dev, fd);
1112 		close(fd);
1113 		if (err)
1114 			goto err_secondary;
1115 		/*
1116 		 * Ethdev pointer is still required as input since
1117 		 * the primary device is not accessible from the
1118 		 * secondary process.
1119 		 */
1120 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1121 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1122 		return eth_dev;
1123 err_secondary:
1124 		mlx5_dev_close(eth_dev);
1125 		return NULL;
1126 	}
1127 	sh = mlx5_alloc_shared_dev_ctx(spawn, mkvlist);
1128 	if (!sh)
1129 		return NULL;
1130 	nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
1131 	/* Check port status. */
1132 	if (spawn->phys_port <= UINT8_MAX) {
1133 		/* Legacy Verbs api only support u8 port number. */
1134 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1135 					    &port_attr);
1136 		if (err) {
1137 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1138 			goto error;
1139 		}
1140 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1141 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1142 			err = EINVAL;
1143 			goto error;
1144 		}
1145 	} else if (nl_rdma >= 0) {
1146 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1147 		err = mlx5_nl_port_state(nl_rdma,
1148 			spawn->phys_dev_name,
1149 			spawn->phys_port);
1150 		if (err < 0) {
1151 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1152 				strerror(rte_errno));
1153 			err = -rte_errno;
1154 			goto error;
1155 		}
1156 		port_attr.state = (enum ibv_port_state)err;
1157 	}
1158 	if (port_attr.state != IBV_PORT_ACTIVE)
1159 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1160 			mlx5_glue->port_state_str(port_attr.state),
1161 			port_attr.state);
1162 	/* Allocate private eth device data. */
1163 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1164 			   sizeof(*priv),
1165 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1166 	if (priv == NULL) {
1167 		DRV_LOG(ERR, "priv allocation failure");
1168 		err = ENOMEM;
1169 		goto error;
1170 	}
1171 	/*
1172 	 * When user configures remote PD and CTX and device creates RxQ by
1173 	 * DevX, external RxQ is both supported and requested.
1174 	 */
1175 	if (mlx5_imported_pd_and_ctx(sh->cdev) && mlx5_devx_obj_ops_en(sh)) {
1176 		priv->ext_rxqs = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1177 					     sizeof(struct mlx5_external_rxq) *
1178 					     MLX5_MAX_EXT_RX_QUEUES, 0,
1179 					     SOCKET_ID_ANY);
1180 		if (priv->ext_rxqs == NULL) {
1181 			DRV_LOG(ERR, "Fail to allocate external RxQ array.");
1182 			err = ENOMEM;
1183 			goto error;
1184 		}
1185 		DRV_LOG(DEBUG, "External RxQ is supported.");
1186 	}
1187 	priv->sh = sh;
1188 	priv->dev_port = spawn->phys_port;
1189 	priv->pci_dev = spawn->pci_dev;
1190 	priv->mtu = RTE_ETHER_MTU;
1191 	/* Some internal functions rely on Netlink sockets, open them now. */
1192 	priv->nl_socket_rdma = nl_rdma;
1193 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE, 0);
1194 	priv->representor = !!switch_info->representor;
1195 	priv->master = !!switch_info->master;
1196 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1197 	priv->vport_meta_tag = 0;
1198 	priv->vport_meta_mask = 0;
1199 	priv->pf_bond = spawn->pf_bond;
1200 
1201 	DRV_LOG(DEBUG,
1202 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1203 		priv->dev_port, dpdk_dev->bus->name,
1204 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1205 		priv->master, priv->representor, priv->pf_bond);
1206 
1207 	/*
1208 	 * If we have E-Switch we should determine the vport attributes.
1209 	 * E-Switch may use either source vport field or reg_c[0] metadata
1210 	 * register to match on vport index. The engaged part of metadata
1211 	 * register is defined by mask.
1212 	 */
1213 	if (sh->esw_mode) {
1214 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1215 						 spawn->phys_port,
1216 						 &vport_info);
1217 		if (err) {
1218 			DRV_LOG(WARNING,
1219 				"Cannot query devx port %d on device %s",
1220 				spawn->phys_port, spawn->phys_dev_name);
1221 			vport_info.query_flags = 0;
1222 		}
1223 	}
1224 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1225 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1226 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1227 		if (!priv->vport_meta_mask) {
1228 			DRV_LOG(ERR,
1229 				"vport zero mask for port %d on bonding device %s",
1230 				spawn->phys_port, spawn->phys_dev_name);
1231 			err = ENOTSUP;
1232 			goto error;
1233 		}
1234 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1235 			DRV_LOG(ERR,
1236 				"Invalid vport tag for port %d on bonding device %s",
1237 				spawn->phys_port, spawn->phys_dev_name);
1238 			err = ENOTSUP;
1239 			goto error;
1240 		}
1241 	}
1242 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1243 		priv->vport_id = vport_info.vport_id;
1244 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1245 		DRV_LOG(ERR,
1246 			"Cannot deduce vport index for port %d on bonding device %s",
1247 			spawn->phys_port, spawn->phys_dev_name);
1248 		err = ENOTSUP;
1249 		goto error;
1250 	} else {
1251 		/*
1252 		 * Suppose vport index in compatible way. Kernel/rdma_core
1253 		 * support single E-Switch per PF configurations only and
1254 		 * vport_id field contains the vport index for associated VF,
1255 		 * which is deduced from representor port name.
1256 		 * For example, let's have the IB device port 10, it has
1257 		 * attached network device eth0, which has port name attribute
1258 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1259 		 * as 3 (2+1). This assigning schema should be changed if the
1260 		 * multiple E-Switch instances per PF configurations or/and PCI
1261 		 * subfunctions are added.
1262 		 */
1263 		priv->vport_id = switch_info->representor ?
1264 				 switch_info->port_name + 1 : -1;
1265 	}
1266 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1267 							  eth_da->type);
1268 	/*
1269 	 * Look for sibling devices in order to reuse their switch domain
1270 	 * if any, otherwise allocate one.
1271 	 */
1272 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1273 		const struct mlx5_priv *opriv =
1274 			rte_eth_devices[port_id].data->dev_private;
1275 
1276 		if (!opriv ||
1277 		    opriv->sh != priv->sh ||
1278 			opriv->domain_id ==
1279 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1280 			continue;
1281 		priv->domain_id = opriv->domain_id;
1282 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1283 			priv->dev_port, priv->domain_id);
1284 		break;
1285 	}
1286 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1287 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1288 		if (err) {
1289 			err = rte_errno;
1290 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1291 				strerror(rte_errno));
1292 			goto error;
1293 		}
1294 		own_domain_id = 1;
1295 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1296 			priv->dev_port, priv->domain_id);
1297 	}
1298 	if (sh->cdev->config.devx) {
1299 		struct mlx5_hca_attr *hca_attr = &sh->cdev->config.hca_attr;
1300 
1301 		sh->steering_format_version = hca_attr->steering_format_version;
1302 #if defined(HAVE_MLX5DV_DR) && \
1303 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1304 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1305 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1306 		    sh->config.dv_flow_en) {
1307 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1308 			/*
1309 			 * Meter needs two REG_C's for color match and pre-sfx
1310 			 * flow match. Here get the REG_C for color match.
1311 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1312 			 */
1313 			reg_c_mask &= 0xfc;
1314 			if (rte_popcount32(reg_c_mask) < 1) {
1315 				priv->mtr_en = 0;
1316 				DRV_LOG(WARNING, "No available register for"
1317 					" meter.");
1318 			} else {
1319 				/*
1320 				 * The meter color register is used by the
1321 				 * flow-hit feature as well.
1322 				 * The flow-hit feature must use REG_C_3
1323 				 * Prefer REG_C_3 if it is available.
1324 				 */
1325 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1326 					priv->mtr_color_reg = REG_C_3;
1327 				else
1328 					priv->mtr_color_reg = ffs(reg_c_mask)
1329 							      - 1 + REG_C_0;
1330 				priv->mtr_en = 1;
1331 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1332 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1333 					priv->mtr_color_reg);
1334 			}
1335 		}
1336 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1337 			uint32_t log_obj_size =
1338 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1339 			if (log_obj_size >=
1340 			    hca_attr->qos.log_meter_aso_granularity &&
1341 			    log_obj_size <=
1342 			    hca_attr->qos.log_meter_aso_max_alloc)
1343 				sh->meter_aso_en = 1;
1344 		}
1345 		if (priv->mtr_en) {
1346 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1347 			if (err) {
1348 				err = -err;
1349 				goto error;
1350 			}
1351 		}
1352 		if (hca_attr->flow.tunnel_header_0_1)
1353 			sh->tunnel_header_0_1 = 1;
1354 		if (hca_attr->flow.tunnel_header_2_3)
1355 			sh->tunnel_header_2_3 = 1;
1356 #endif
1357 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1358 		if (hca_attr->flow_hit_aso && priv->mtr_color_reg == REG_C_3) {
1359 			sh->flow_hit_aso_en = 1;
1360 			err = mlx5_flow_aso_age_mng_init(sh);
1361 			if (err) {
1362 				err = -err;
1363 				goto error;
1364 			}
1365 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1366 		}
1367 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1368 #if defined (HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1369     defined (HAVE_MLX5_DR_ACTION_ASO_CT)
1370 		/* HWS create CT ASO SQ based on HWS configure queue number. */
1371 		if (sh->config.dv_flow_en != 2 &&
1372 		    hca_attr->ct_offload && priv->mtr_color_reg == REG_C_3) {
1373 			err = mlx5_flow_aso_ct_mng_init(sh);
1374 			if (err) {
1375 				err = -err;
1376 				goto error;
1377 			}
1378 			DRV_LOG(DEBUG, "CT ASO is supported.");
1379 			sh->ct_aso_en = 1;
1380 		}
1381 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1382 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1383 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1384 		    sh->config.dv_flow_en) {
1385 			priv->sampler_en = 1;
1386 			DRV_LOG(DEBUG, "Sampler enabled!");
1387 		} else {
1388 			priv->sampler_en = 0;
1389 			if (!hca_attr->log_max_ft_sampler_num)
1390 				DRV_LOG(WARNING,
1391 					"No available register for sampler.");
1392 			else
1393 				DRV_LOG(DEBUG, "DV flow is not supported!");
1394 		}
1395 #endif
1396 		if (hca_attr->lag_rx_port_affinity) {
1397 			sh->lag_rx_port_affinity_en = 1;
1398 			DRV_LOG(DEBUG, "LAG Rx Port Affinity enabled");
1399 		}
1400 		priv->num_lag_ports = hca_attr->num_lag_ports;
1401 		DRV_LOG(DEBUG, "The number of lag ports is %d", priv->num_lag_ports);
1402 	}
1403 	/* Process parameters and store port configuration on priv structure. */
1404 	err = mlx5_port_args_config(priv, mkvlist, &priv->config);
1405 	if (err) {
1406 		err = rte_errno;
1407 		DRV_LOG(ERR, "Failed to process port configure: %s",
1408 			strerror(rte_errno));
1409 		goto error;
1410 	}
1411 	eth_dev = rte_eth_dev_allocate(name);
1412 	if (eth_dev == NULL) {
1413 		DRV_LOG(ERR, "can not allocate rte ethdev");
1414 		err = ENOMEM;
1415 		goto error;
1416 	}
1417 	if (priv->representor) {
1418 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1419 		eth_dev->data->representor_id = priv->representor_id;
1420 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1421 			struct mlx5_priv *opriv =
1422 				rte_eth_devices[port_id].data->dev_private;
1423 			if (opriv &&
1424 			    opriv->master &&
1425 			    opriv->domain_id == priv->domain_id &&
1426 			    opriv->sh == priv->sh) {
1427 				eth_dev->data->backer_port_id = port_id;
1428 				break;
1429 			}
1430 		}
1431 		if (port_id >= RTE_MAX_ETHPORTS)
1432 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1433 	}
1434 	priv->mp_id.port_id = eth_dev->data->port_id;
1435 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1436 	/*
1437 	 * Store associated network device interface index. This index
1438 	 * is permanent throughout the lifetime of device. So, we may store
1439 	 * the ifindex here and use the cached value further.
1440 	 */
1441 	MLX5_ASSERT(spawn->ifindex);
1442 	priv->if_index = spawn->ifindex;
1443 	priv->lag_affinity_idx = sh->refcnt - 1;
1444 	eth_dev->data->dev_private = priv;
1445 	priv->dev_data = eth_dev->data;
1446 	eth_dev->data->mac_addrs = priv->mac;
1447 	eth_dev->device = dpdk_dev;
1448 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1449 	/* Configure the first MAC address by default. */
1450 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1451 		DRV_LOG(ERR,
1452 			"port %u cannot get MAC address, is mlx5_en"
1453 			" loaded? (errno: %s)",
1454 			eth_dev->data->port_id, strerror(rte_errno));
1455 		err = ENODEV;
1456 		goto error;
1457 	}
1458 	DRV_LOG(INFO,
1459 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1460 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1461 #ifdef RTE_LIBRTE_MLX5_DEBUG
1462 	{
1463 		char ifname[MLX5_NAMESIZE];
1464 
1465 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1466 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1467 				eth_dev->data->port_id, ifname);
1468 		else
1469 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1470 				eth_dev->data->port_id);
1471 	}
1472 #endif
1473 	/* Get actual MTU if possible. */
1474 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1475 	if (err) {
1476 		err = rte_errno;
1477 		goto error;
1478 	}
1479 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1480 		priv->mtu);
1481 	/* Initialize burst functions to prevent crashes before link-up. */
1482 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1483 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1484 	eth_dev->dev_ops = &mlx5_dev_ops;
1485 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1486 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1487 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1488 	/* Register MAC address. */
1489 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1490 	if (sh->dev_cap.vf && sh->config.vf_nl_en)
1491 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1492 				      mlx5_ifindex(eth_dev),
1493 				      eth_dev->data->mac_addrs,
1494 				      MLX5_MAX_MAC_ADDRESSES);
1495 	priv->ctrl_flows = 0;
1496 	rte_spinlock_init(&priv->flow_list_lock);
1497 	TAILQ_INIT(&priv->flow_meters);
1498 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1499 	if (!priv->mtr_profile_tbl)
1500 		goto error;
1501 	/* Bring Ethernet device up. */
1502 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1503 		eth_dev->data->port_id);
1504 	/* Read link status in case it is up and there will be no event. */
1505 	mlx5_link_update(eth_dev, 0);
1506 	/* Watch LSC interrupts between port probe and port start. */
1507 	priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1508 							eth_dev->data->port_id;
1509 	mlx5_set_link_up(eth_dev);
1510 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1511 		icfg[i].release_mem_en = !!sh->config.reclaim_mode;
1512 		if (sh->config.reclaim_mode)
1513 			icfg[i].per_core_cache = 0;
1514 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1515 		if (!priv->flows[i])
1516 			goto error;
1517 	}
1518 	/* Create context for virtual machine VLAN workaround. */
1519 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1520 	if (sh->config.dv_flow_en) {
1521 		err = mlx5_alloc_shared_dr(priv);
1522 		if (err)
1523 			goto error;
1524 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1525 			goto error;
1526 	}
1527 	if (mlx5_devx_obj_ops_en(sh)) {
1528 		priv->obj_ops = devx_obj_ops;
1529 		mlx5_queue_counter_id_prepare(eth_dev);
1530 		priv->obj_ops.lb_dummy_queue_create =
1531 					mlx5_rxq_ibv_obj_dummy_lb_create;
1532 		priv->obj_ops.lb_dummy_queue_release =
1533 					mlx5_rxq_ibv_obj_dummy_lb_release;
1534 	} else if (spawn->max_port > UINT8_MAX) {
1535 		/* Verbs can't support ports larger than 255 by design. */
1536 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1537 		err = ENOTSUP;
1538 		goto error;
1539 	} else {
1540 		priv->obj_ops = ibv_obj_ops;
1541 	}
1542 	if (sh->config.tx_pp &&
1543 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1544 		/*
1545 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1546 		 * packet pacing and already checked above.
1547 		 * Hence, we should only make sure the SQs will be created
1548 		 * with DevX, not with Verbs.
1549 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1550 		 * with Clock Queue UAR as required for Tx scheduling.
1551 		 */
1552 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1553 		err = ENODEV;
1554 		goto error;
1555 	}
1556 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1557 	if (!priv->drop_queue.hrxq)
1558 		goto error;
1559 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1560 				       mlx5_hrxq_create_cb,
1561 				       mlx5_hrxq_match_cb,
1562 				       mlx5_hrxq_remove_cb,
1563 				       mlx5_hrxq_clone_cb,
1564 				       mlx5_hrxq_clone_free_cb);
1565 	if (!priv->hrxqs)
1566 		goto error;
1567 	mlx5_set_metadata_mask(eth_dev);
1568 	if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1569 	    !priv->sh->dv_regc0_mask) {
1570 		DRV_LOG(ERR, "metadata mode %u is not supported "
1571 			     "(no metadata reg_c[0] is available)",
1572 			     sh->config.dv_xmeta_en);
1573 			err = ENOTSUP;
1574 			goto error;
1575 	}
1576 	rte_rwlock_init(&priv->ind_tbls_lock);
1577 	if (priv->sh->config.dv_flow_en == 2) {
1578 #ifdef HAVE_MLX5_HWS_SUPPORT
1579 		if (priv->sh->config.dv_esw_en) {
1580 			uint32_t usable_bits;
1581 			uint32_t required_bits;
1582 
1583 			if (priv->sh->dv_regc0_mask == UINT32_MAX) {
1584 				DRV_LOG(ERR, "E-Switch port metadata is required when using HWS "
1585 					     "but it is disabled (configure it through devlink)");
1586 				err = ENOTSUP;
1587 				goto error;
1588 			}
1589 			if (priv->sh->dv_regc0_mask == 0) {
1590 				DRV_LOG(ERR, "E-Switch with HWS is not supported "
1591 					     "(no available bits in reg_c[0])");
1592 				err = ENOTSUP;
1593 				goto error;
1594 			}
1595 			usable_bits = rte_popcount32(priv->sh->dv_regc0_mask);
1596 			required_bits = rte_popcount32(priv->vport_meta_mask);
1597 			if (usable_bits < required_bits) {
1598 				DRV_LOG(ERR, "Not enough bits available in reg_c[0] to provide "
1599 					     "representor matching.");
1600 				err = ENOTSUP;
1601 				goto error;
1602 			}
1603 		}
1604 		if (priv->vport_meta_mask)
1605 			flow_hw_set_port_info(eth_dev);
1606 		if (priv->sh->config.dv_esw_en &&
1607 		    priv->sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1608 		    priv->sh->config.dv_xmeta_en != MLX5_XMETA_MODE_META32_HWS) {
1609 			DRV_LOG(ERR,
1610 				"metadata mode %u is not supported in HWS eswitch mode",
1611 				priv->sh->config.dv_xmeta_en);
1612 				err = ENOTSUP;
1613 				goto error;
1614 		}
1615 		/* Only HWS requires this information. */
1616 		flow_hw_init_tags_set(eth_dev);
1617 		flow_hw_init_flow_metadata_config(eth_dev);
1618 		if (priv->sh->config.dv_esw_en &&
1619 		    flow_hw_create_vport_action(eth_dev)) {
1620 			DRV_LOG(ERR, "port %u failed to create vport action",
1621 				eth_dev->data->port_id);
1622 			err = EINVAL;
1623 			goto error;
1624 		}
1625 		/*
1626 		 * If representor matching is disabled, PMD cannot create default flow rules
1627 		 * to receive traffic for all ports, since implicit source port match is not added.
1628 		 * Isolated mode is forced.
1629 		 */
1630 		if (priv->sh->config.dv_esw_en && !priv->sh->config.repr_matching) {
1631 			err = mlx5_flow_isolate(eth_dev, 1, NULL);
1632 			if (err < 0) {
1633 				err = -err;
1634 				goto error;
1635 			}
1636 			DRV_LOG(WARNING, "port %u ingress traffic is restricted to defined "
1637 					 "flow rules (isolated mode) since representor "
1638 					 "matching is disabled",
1639 				eth_dev->data->port_id);
1640 		}
1641 		return eth_dev;
1642 #else
1643 		DRV_LOG(ERR, "DV support is missing for HWS.");
1644 		goto error;
1645 #endif
1646 	}
1647 	if (!priv->sh->flow_priority_check_flag) {
1648 		/* Supported Verbs flow priority number detection. */
1649 		err = mlx5_flow_discover_priorities(eth_dev);
1650 		priv->sh->flow_max_priority = err;
1651 		priv->sh->flow_priority_check_flag = 1;
1652 	} else {
1653 		err = priv->sh->flow_max_priority;
1654 	}
1655 	if (err < 0) {
1656 		err = -err;
1657 		goto error;
1658 	}
1659 	/* Query availability of metadata reg_c's. */
1660 	if (!priv->sh->metadata_regc_check_flag) {
1661 		err = mlx5_flow_discover_mreg_c(eth_dev);
1662 		if (err < 0) {
1663 			err = -err;
1664 			goto error;
1665 		}
1666 	}
1667 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1668 		DRV_LOG(DEBUG,
1669 			"port %u extensive metadata register is not supported",
1670 			eth_dev->data->port_id);
1671 		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1672 			DRV_LOG(ERR, "metadata mode %u is not supported "
1673 				     "(no metadata registers available)",
1674 				     sh->config.dv_xmeta_en);
1675 			err = ENOTSUP;
1676 			goto error;
1677 		}
1678 	}
1679 	if (sh->config.dv_flow_en &&
1680 	    sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1681 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1682 	    priv->sh->dv_regc0_mask) {
1683 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1684 						      MLX5_FLOW_MREG_HTABLE_SZ,
1685 						      false, true, eth_dev,
1686 						      flow_dv_mreg_create_cb,
1687 						      flow_dv_mreg_match_cb,
1688 						      flow_dv_mreg_remove_cb,
1689 						      flow_dv_mreg_clone_cb,
1690 						    flow_dv_mreg_clone_free_cb);
1691 		if (!priv->mreg_cp_tbl) {
1692 			err = ENOMEM;
1693 			goto error;
1694 		}
1695 	}
1696 	rte_spinlock_init(&priv->shared_act_sl);
1697 	mlx5_flow_counter_mode_config(eth_dev);
1698 	mlx5_flow_drop_action_config(eth_dev);
1699 	if (sh->config.dv_flow_en)
1700 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1701 	return eth_dev;
1702 error:
1703 	if (priv) {
1704 		priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1705 							       RTE_MAX_ETHPORTS;
1706 		rte_io_wmb();
1707 #ifdef HAVE_MLX5_HWS_SUPPORT
1708 		if (eth_dev &&
1709 		    priv->sh &&
1710 		    priv->sh->config.dv_flow_en == 2 &&
1711 		    priv->sh->config.dv_esw_en)
1712 			flow_hw_destroy_vport_action(eth_dev);
1713 #endif
1714 		if (priv->mreg_cp_tbl)
1715 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1716 		if (priv->sh)
1717 			mlx5_os_free_shared_dr(priv);
1718 		if (priv->nl_socket_route >= 0)
1719 			close(priv->nl_socket_route);
1720 		if (priv->vmwa_context)
1721 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1722 		if (eth_dev && priv->drop_queue.hrxq)
1723 			mlx5_drop_action_destroy(eth_dev);
1724 		if (priv->mtr_profile_tbl)
1725 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1726 		if (own_domain_id)
1727 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1728 		if (priv->hrxqs)
1729 			mlx5_list_destroy(priv->hrxqs);
1730 		if (eth_dev && priv->flex_item_map)
1731 			mlx5_flex_item_port_cleanup(eth_dev);
1732 		mlx5_free(priv->ext_rxqs);
1733 		mlx5_free(priv);
1734 		if (eth_dev != NULL)
1735 			eth_dev->data->dev_private = NULL;
1736 	}
1737 	if (eth_dev != NULL) {
1738 		/* mac_addrs must not be freed alone because part of
1739 		 * dev_private
1740 		 **/
1741 		eth_dev->data->mac_addrs = NULL;
1742 		rte_eth_dev_release_port(eth_dev);
1743 	}
1744 	if (sh)
1745 		mlx5_free_shared_dev_ctx(sh);
1746 	if (nl_rdma >= 0)
1747 		close(nl_rdma);
1748 	MLX5_ASSERT(err > 0);
1749 	rte_errno = err;
1750 	return NULL;
1751 }
1752 
1753 /**
1754  * Comparison callback to sort device data.
1755  *
1756  * This is meant to be used with qsort().
1757  *
1758  * @param a[in]
1759  *   Pointer to pointer to first data object.
1760  * @param b[in]
1761  *   Pointer to pointer to second data object.
1762  *
1763  * @return
1764  *   0 if both objects are equal, less than 0 if the first argument is less
1765  *   than the second, greater than 0 otherwise.
1766  */
1767 static int
1768 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1769 {
1770 	const struct mlx5_switch_info *si_a =
1771 		&((const struct mlx5_dev_spawn_data *)a)->info;
1772 	const struct mlx5_switch_info *si_b =
1773 		&((const struct mlx5_dev_spawn_data *)b)->info;
1774 	int ret;
1775 
1776 	/* Master device first. */
1777 	ret = si_b->master - si_a->master;
1778 	if (ret)
1779 		return ret;
1780 	/* Then representor devices. */
1781 	ret = si_b->representor - si_a->representor;
1782 	if (ret)
1783 		return ret;
1784 	/* Unidentified devices come last in no specific order. */
1785 	if (!si_a->representor)
1786 		return 0;
1787 	/* Order representors by name. */
1788 	return si_a->port_name - si_b->port_name;
1789 }
1790 
1791 /**
1792  * Match PCI information for possible slaves of bonding device.
1793  *
1794  * @param[in] ibdev_name
1795  *   Name of Infiniband device.
1796  * @param[in] pci_dev
1797  *   Pointer to primary PCI address structure to match.
1798  * @param[in] nl_rdma
1799  *   Netlink RDMA group socket handle.
1800  * @param[in] owner
1801  *   Representor owner PF index.
1802  * @param[out] bond_info
1803  *   Pointer to bonding information.
1804  *
1805  * @return
1806  *   negative value if no bonding device found, otherwise
1807  *   positive index of slave PF in bonding.
1808  */
1809 static int
1810 mlx5_device_bond_pci_match(const char *ibdev_name,
1811 			   const struct rte_pci_addr *pci_dev,
1812 			   int nl_rdma, uint16_t owner,
1813 			   struct mlx5_bond_info *bond_info)
1814 {
1815 	char ifname[IF_NAMESIZE + 1];
1816 	unsigned int ifindex;
1817 	unsigned int np, i;
1818 	FILE *bond_file = NULL, *file;
1819 	int pf = -1;
1820 	int ret;
1821 	uint8_t cur_guid[32] = {0};
1822 	uint8_t guid[32] = {0};
1823 
1824 	/*
1825 	 * Try to get master device name. If something goes wrong suppose
1826 	 * the lack of kernel support and no bonding devices.
1827 	 */
1828 	memset(bond_info, 0, sizeof(*bond_info));
1829 	if (nl_rdma < 0)
1830 		return -1;
1831 	if (!strstr(ibdev_name, "bond"))
1832 		return -1;
1833 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1834 	if (!np)
1835 		return -1;
1836 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1837 		return -1;
1838 	/*
1839 	 * The master device might not be on the predefined port(not on port
1840 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1841 	 * device ports and find master.
1842 	 */
1843 	for (i = 1; i <= np; ++i) {
1844 		/* Check whether Infiniband port is populated. */
1845 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1846 		if (!ifindex)
1847 			continue;
1848 		if (!if_indextoname(ifindex, ifname))
1849 			continue;
1850 		/* Try to read bonding slave names from sysfs. */
1851 		MKSTR(slaves,
1852 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1853 		bond_file = fopen(slaves, "r");
1854 		if (bond_file)
1855 			break;
1856 	}
1857 	if (!bond_file)
1858 		return -1;
1859 	/* Use safe format to check maximal buffer length. */
1860 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1861 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1862 		char tmp_str[IF_NAMESIZE + 32];
1863 		struct rte_pci_addr pci_addr;
1864 		struct mlx5_switch_info	info;
1865 		int ret;
1866 
1867 		/* Process slave interface names in the loop. */
1868 		snprintf(tmp_str, sizeof(tmp_str),
1869 			 "/sys/class/net/%s", ifname);
1870 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1871 			DRV_LOG(WARNING,
1872 				"Cannot get PCI address for netdev \"%s\".",
1873 				ifname);
1874 			continue;
1875 		}
1876 		/* Slave interface PCI address match found. */
1877 		snprintf(tmp_str, sizeof(tmp_str),
1878 			 "/sys/class/net/%s/phys_port_name", ifname);
1879 		file = fopen(tmp_str, "rb");
1880 		if (!file)
1881 			break;
1882 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1883 		if (fscanf(file, "%32s", tmp_str) == 1)
1884 			mlx5_translate_port_name(tmp_str, &info);
1885 		fclose(file);
1886 		/* Only process PF ports. */
1887 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1888 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1889 			continue;
1890 		/* Check max bonding member. */
1891 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1892 			DRV_LOG(WARNING, "bonding index out of range, "
1893 				"please increase MLX5_BOND_MAX_PORTS: %s",
1894 				tmp_str);
1895 			break;
1896 		}
1897 		/* Get ifindex. */
1898 		snprintf(tmp_str, sizeof(tmp_str),
1899 			 "/sys/class/net/%s/ifindex", ifname);
1900 		file = fopen(tmp_str, "rb");
1901 		if (!file)
1902 			break;
1903 		ret = fscanf(file, "%u", &ifindex);
1904 		fclose(file);
1905 		if (ret != 1)
1906 			break;
1907 		/* Save bonding info. */
1908 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1909 			sizeof(bond_info->ports[0].ifname));
1910 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1911 		bond_info->ports[info.port_name].ifindex = ifindex;
1912 		bond_info->n_port++;
1913 		/*
1914 		 * Under socket direct mode, bonding will use
1915 		 * system_image_guid as identification.
1916 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1917 		 * All bonding members should have the same guid even if driver
1918 		 * is using PCIe BDF.
1919 		 */
1920 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1921 		if (ret < 0)
1922 			break;
1923 		else if (ret > 0) {
1924 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1925 			    owner == info.port_name &&
1926 			    (owner != 0 || (owner == 0 &&
1927 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1928 				pf = info.port_name;
1929 		} else if (pci_dev->domain == pci_addr.domain &&
1930 		    pci_dev->bus == pci_addr.bus &&
1931 		    pci_dev->devid == pci_addr.devid &&
1932 		    ((pci_dev->function == 0 &&
1933 		      pci_dev->function + owner == pci_addr.function) ||
1934 		     (pci_dev->function == owner &&
1935 		      pci_addr.function == owner)))
1936 			pf = info.port_name;
1937 	}
1938 	if (pf >= 0) {
1939 		/* Get bond interface info */
1940 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1941 					   bond_info->ifname);
1942 		if (ret)
1943 			DRV_LOG(ERR, "unable to get bond info: %s",
1944 				strerror(rte_errno));
1945 		else
1946 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1947 				ifindex, bond_info->ifindex, bond_info->ifname);
1948 	}
1949 	if (owner == 0 && pf != 0) {
1950 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
1951 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1952 				pci_dev->function);
1953 	}
1954 	return pf;
1955 }
1956 
1957 /**
1958  * Register a PCI device within bonding.
1959  *
1960  * This function spawns Ethernet devices out of a given PCI device and
1961  * bonding owner PF index.
1962  *
1963  * @param[in] cdev
1964  *   Pointer to common mlx5 device structure.
1965  * @param[in] req_eth_da
1966  *   Requested ethdev device argument.
1967  * @param[in] owner_id
1968  *   Requested owner PF port ID within bonding device, default to 0.
1969  * @param[in, out] mkvlist
1970  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1971  *
1972  * @return
1973  *   0 on success, a negative errno value otherwise and rte_errno is set.
1974  */
1975 static int
1976 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
1977 		     struct rte_eth_devargs *req_eth_da,
1978 		     uint16_t owner_id, struct mlx5_kvargs_ctrl *mkvlist)
1979 {
1980 	struct ibv_device **ibv_list;
1981 	/*
1982 	 * Number of found IB Devices matching with requested PCI BDF.
1983 	 * nd != 1 means there are multiple IB devices over the same
1984 	 * PCI device and we have representors and master.
1985 	 */
1986 	unsigned int nd = 0;
1987 	/*
1988 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1989 	 * we have the single multiport IB device, and there may be
1990 	 * representors attached to some of found ports.
1991 	 */
1992 	unsigned int np = 0;
1993 	/*
1994 	 * Number of DPDK ethernet devices to Spawn - either over
1995 	 * multiple IB devices or multiple ports of single IB device.
1996 	 * Actually this is the number of iterations to spawn.
1997 	 */
1998 	unsigned int ns = 0;
1999 	/*
2000 	 * Bonding device
2001 	 *   < 0 - no bonding device (single one)
2002 	 *  >= 0 - bonding device (value is slave PF index)
2003 	 */
2004 	int bd = -1;
2005 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2006 	struct mlx5_dev_spawn_data *list = NULL;
2007 	struct rte_eth_devargs eth_da = *req_eth_da;
2008 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2009 	struct mlx5_bond_info bond_info;
2010 	int ret = -1;
2011 
2012 	errno = 0;
2013 	ibv_list = mlx5_glue->get_device_list(&ret);
2014 	if (!ibv_list) {
2015 		rte_errno = errno ? errno : ENOSYS;
2016 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2017 		return -rte_errno;
2018 	}
2019 	/*
2020 	 * First scan the list of all Infiniband devices to find
2021 	 * matching ones, gathering into the list.
2022 	 */
2023 	struct ibv_device *ibv_match[ret + 1];
2024 	int nl_route = mlx5_nl_init(NETLINK_ROUTE, 0);
2025 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
2026 	unsigned int i;
2027 
2028 	while (ret-- > 0) {
2029 		struct rte_pci_addr pci_addr;
2030 
2031 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2032 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2033 						nl_rdma, owner_id, &bond_info);
2034 		if (bd >= 0) {
2035 			/*
2036 			 * Bonding device detected. Only one match is allowed,
2037 			 * the bonding is supported over multi-port IB device,
2038 			 * there should be no matches on representor PCI
2039 			 * functions or non VF LAG bonding devices with
2040 			 * specified address.
2041 			 */
2042 			if (nd) {
2043 				DRV_LOG(ERR,
2044 					"multiple PCI match on bonding device"
2045 					"\"%s\" found", ibv_list[ret]->name);
2046 				rte_errno = ENOENT;
2047 				ret = -rte_errno;
2048 				goto exit;
2049 			}
2050 			/* Amend owner pci address if owner PF ID specified. */
2051 			if (eth_da.nb_representor_ports)
2052 				owner_pci.function += owner_id;
2053 			DRV_LOG(INFO,
2054 				"PCI information matches for slave %d bonding device \"%s\"",
2055 				bd, ibv_list[ret]->name);
2056 			ibv_match[nd++] = ibv_list[ret];
2057 			break;
2058 		} else {
2059 			/* Bonding device not found. */
2060 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2061 					      &pci_addr))
2062 				continue;
2063 			if (owner_pci.domain != pci_addr.domain ||
2064 			    owner_pci.bus != pci_addr.bus ||
2065 			    owner_pci.devid != pci_addr.devid ||
2066 			    owner_pci.function != pci_addr.function)
2067 				continue;
2068 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2069 				ibv_list[ret]->name);
2070 			ibv_match[nd++] = ibv_list[ret];
2071 		}
2072 	}
2073 	ibv_match[nd] = NULL;
2074 	if (!nd) {
2075 		/* No device matches, just complain and bail out. */
2076 		DRV_LOG(WARNING,
2077 			"PF %u doesn't have Verbs device matches PCI device " PCI_PRI_FMT ","
2078 			" are kernel drivers loaded?",
2079 			owner_id, owner_pci.domain, owner_pci.bus,
2080 			owner_pci.devid, owner_pci.function);
2081 		rte_errno = ENOENT;
2082 		ret = -rte_errno;
2083 		goto exit;
2084 	}
2085 	if (nd == 1) {
2086 		/*
2087 		 * Found single matching device may have multiple ports.
2088 		 * Each port may be representor, we have to check the port
2089 		 * number and check the representors existence.
2090 		 */
2091 		if (nl_rdma >= 0)
2092 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2093 		if (!np)
2094 			DRV_LOG(WARNING,
2095 				"Cannot get IB device \"%s\" ports number.",
2096 				ibv_match[0]->name);
2097 		if (bd >= 0 && !np) {
2098 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2099 			rte_errno = ENOENT;
2100 			ret = -rte_errno;
2101 			goto exit;
2102 		}
2103 	}
2104 	/* Now we can determine the maximal amount of devices to be spawned. */
2105 	list = mlx5_malloc(MLX5_MEM_ZERO,
2106 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2107 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2108 	if (!list) {
2109 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2110 		rte_errno = ENOMEM;
2111 		ret = -rte_errno;
2112 		goto exit;
2113 	}
2114 	if (bd >= 0 || np > 1) {
2115 		/*
2116 		 * Single IB device with multiple ports found,
2117 		 * it may be E-Switch master device and representors.
2118 		 * We have to perform identification through the ports.
2119 		 */
2120 		MLX5_ASSERT(nl_rdma >= 0);
2121 		MLX5_ASSERT(ns == 0);
2122 		MLX5_ASSERT(nd == 1);
2123 		MLX5_ASSERT(np);
2124 		for (i = 1; i <= np; ++i) {
2125 			list[ns].bond_info = &bond_info;
2126 			list[ns].max_port = np;
2127 			list[ns].phys_port = i;
2128 			list[ns].phys_dev_name = ibv_match[0]->name;
2129 			list[ns].eth_dev = NULL;
2130 			list[ns].pci_dev = pci_dev;
2131 			list[ns].cdev = cdev;
2132 			list[ns].pf_bond = bd;
2133 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2134 							   ibv_match[0]->name,
2135 							   i);
2136 			if (!list[ns].ifindex) {
2137 				/*
2138 				 * No network interface index found for the
2139 				 * specified port, it means there is no
2140 				 * representor on this port. It's OK,
2141 				 * there can be disabled ports, for example
2142 				 * if sriov_numvfs < sriov_totalvfs.
2143 				 */
2144 				continue;
2145 			}
2146 			ret = -1;
2147 			if (nl_route >= 0)
2148 				ret = mlx5_nl_switch_info(nl_route,
2149 							  list[ns].ifindex,
2150 							  &list[ns].info);
2151 			if (ret || (!list[ns].info.representor &&
2152 				    !list[ns].info.master)) {
2153 				/*
2154 				 * We failed to recognize representors with
2155 				 * Netlink, let's try to perform the task
2156 				 * with sysfs.
2157 				 */
2158 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2159 							     &list[ns].info);
2160 			}
2161 			if (!ret && bd >= 0) {
2162 				switch (list[ns].info.name_type) {
2163 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2164 					if (np == 1) {
2165 						/*
2166 						 * Force standalone bonding
2167 						 * device for ROCE LAG
2168 						 * configurations.
2169 						 */
2170 						list[ns].info.master = 0;
2171 						list[ns].info.representor = 0;
2172 					}
2173 					if (list[ns].info.port_name == bd)
2174 						ns++;
2175 					break;
2176 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2177 					/* Fallthrough */
2178 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2179 					/* Fallthrough */
2180 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2181 					if (list[ns].info.pf_num == bd)
2182 						ns++;
2183 					break;
2184 				default:
2185 					break;
2186 				}
2187 				continue;
2188 			}
2189 			if (!ret && (list[ns].info.representor ^
2190 				     list[ns].info.master))
2191 				ns++;
2192 		}
2193 		if (!ns) {
2194 			DRV_LOG(ERR,
2195 				"Unable to recognize master/representors on the IB device with multiple ports.");
2196 			rte_errno = ENOENT;
2197 			ret = -rte_errno;
2198 			goto exit;
2199 		}
2200 	} else {
2201 		/*
2202 		 * The existence of several matching entries (nd > 1) means
2203 		 * port representors have been instantiated. No existing Verbs
2204 		 * call nor sysfs entries can tell them apart, this can only
2205 		 * be done through Netlink calls assuming kernel drivers are
2206 		 * recent enough to support them.
2207 		 *
2208 		 * In the event of identification failure through Netlink,
2209 		 * try again through sysfs, then:
2210 		 *
2211 		 * 1. A single IB device matches (nd == 1) with single
2212 		 *    port (np=0/1) and is not a representor, assume
2213 		 *    no switch support.
2214 		 *
2215 		 * 2. Otherwise no safe assumptions can be made;
2216 		 *    complain louder and bail out.
2217 		 */
2218 		for (i = 0; i != nd; ++i) {
2219 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2220 			list[ns].bond_info = NULL;
2221 			list[ns].max_port = 1;
2222 			list[ns].phys_port = 1;
2223 			list[ns].phys_dev_name = ibv_match[i]->name;
2224 			list[ns].eth_dev = NULL;
2225 			list[ns].pci_dev = pci_dev;
2226 			list[ns].cdev = cdev;
2227 			list[ns].pf_bond = -1;
2228 			list[ns].ifindex = 0;
2229 			if (nl_rdma >= 0)
2230 				list[ns].ifindex = mlx5_nl_ifindex
2231 							    (nl_rdma,
2232 							     ibv_match[i]->name,
2233 							     1);
2234 			if (!list[ns].ifindex) {
2235 				char ifname[IF_NAMESIZE];
2236 
2237 				/*
2238 				 * Netlink failed, it may happen with old
2239 				 * ib_core kernel driver (before 4.16).
2240 				 * We can assume there is old driver because
2241 				 * here we are processing single ports IB
2242 				 * devices. Let's try sysfs to retrieve
2243 				 * the ifindex. The method works for
2244 				 * master device only.
2245 				 */
2246 				if (nd > 1) {
2247 					/*
2248 					 * Multiple devices found, assume
2249 					 * representors, can not distinguish
2250 					 * master/representor and retrieve
2251 					 * ifindex via sysfs.
2252 					 */
2253 					continue;
2254 				}
2255 				ret = mlx5_get_ifname_sysfs
2256 					(ibv_match[i]->ibdev_path, ifname);
2257 				if (!ret)
2258 					list[ns].ifindex =
2259 						if_nametoindex(ifname);
2260 				if (!list[ns].ifindex) {
2261 					/*
2262 					 * No network interface index found
2263 					 * for the specified device, it means
2264 					 * there it is neither representor
2265 					 * nor master.
2266 					 */
2267 					continue;
2268 				}
2269 			}
2270 			ret = -1;
2271 			if (nl_route >= 0)
2272 				ret = mlx5_nl_switch_info(nl_route,
2273 							  list[ns].ifindex,
2274 							  &list[ns].info);
2275 			if (ret || (!list[ns].info.representor &&
2276 				    !list[ns].info.master)) {
2277 				/*
2278 				 * We failed to recognize representors with
2279 				 * Netlink, let's try to perform the task
2280 				 * with sysfs.
2281 				 */
2282 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2283 							     &list[ns].info);
2284 			}
2285 			if (!ret && (list[ns].info.representor ^
2286 				     list[ns].info.master)) {
2287 				ns++;
2288 			} else if ((nd == 1) &&
2289 				   !list[ns].info.representor &&
2290 				   !list[ns].info.master) {
2291 				/*
2292 				 * Single IB device with one physical port and
2293 				 * attached network device.
2294 				 * May be SRIOV is not enabled or there is no
2295 				 * representors.
2296 				 */
2297 				DRV_LOG(INFO, "No E-Switch support detected.");
2298 				ns++;
2299 				break;
2300 			}
2301 		}
2302 		if (!ns) {
2303 			DRV_LOG(ERR,
2304 				"Unable to recognize master/representors on the multiple IB devices.");
2305 			rte_errno = ENOENT;
2306 			ret = -rte_errno;
2307 			goto exit;
2308 		}
2309 		/*
2310 		 * New kernels may add the switch_id attribute for the case
2311 		 * there is no E-Switch and we wrongly recognized the only
2312 		 * device as master. Override this if there is the single
2313 		 * device with single port and new device name format present.
2314 		 */
2315 		if (nd == 1 &&
2316 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2317 			list[0].info.master = 0;
2318 			list[0].info.representor = 0;
2319 		}
2320 	}
2321 	MLX5_ASSERT(ns);
2322 	/*
2323 	 * Sort list to probe devices in natural order for users convenience
2324 	 * (i.e. master first, then representors from lowest to highest ID).
2325 	 */
2326 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2327 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2328 		/* Set devargs default values. */
2329 		if (eth_da.nb_mh_controllers == 0) {
2330 			eth_da.nb_mh_controllers = 1;
2331 			eth_da.mh_controllers[0] = 0;
2332 		}
2333 		if (eth_da.nb_ports == 0 && ns > 0) {
2334 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2335 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2336 					pci_dev->device.devargs->args);
2337 			eth_da.nb_ports = 1;
2338 			eth_da.ports[0] = list[0].info.pf_num;
2339 		}
2340 		if (eth_da.nb_representor_ports == 0) {
2341 			eth_da.nb_representor_ports = 1;
2342 			eth_da.representor_ports[0] = 0;
2343 		}
2344 	}
2345 	for (i = 0; i != ns; ++i) {
2346 		uint32_t restore;
2347 
2348 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i], &eth_da,
2349 						 mkvlist);
2350 		if (!list[i].eth_dev) {
2351 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2352 				break;
2353 			/* Device is disabled or already spawned. Ignore it. */
2354 			continue;
2355 		}
2356 		restore = list[i].eth_dev->data->dev_flags;
2357 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2358 		/**
2359 		 * Each representor has a dedicated interrupts vector.
2360 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2361 		 * representor eth_dev object because representor and PF
2362 		 * share the same PCI address.
2363 		 * Override representor device with a dedicated
2364 		 * interrupts handle here.
2365 		 * Representor interrupts handle is released in mlx5_dev_stop().
2366 		 */
2367 		if (list[i].info.representor) {
2368 			struct rte_intr_handle *intr_handle =
2369 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2370 			if (intr_handle == NULL) {
2371 				DRV_LOG(ERR,
2372 					"port %u failed to allocate memory for interrupt handler "
2373 					"Rx interrupts will not be supported",
2374 					i);
2375 				rte_errno = ENOMEM;
2376 				ret = -rte_errno;
2377 				goto exit;
2378 			}
2379 			list[i].eth_dev->intr_handle = intr_handle;
2380 		}
2381 		/* Restore non-PCI flags cleared by the above call. */
2382 		list[i].eth_dev->data->dev_flags |= restore;
2383 		rte_eth_dev_probing_finish(list[i].eth_dev);
2384 	}
2385 	if (i != ns) {
2386 		DRV_LOG(ERR,
2387 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2388 			" encountering an error: %s",
2389 			owner_pci.domain, owner_pci.bus,
2390 			owner_pci.devid, owner_pci.function,
2391 			strerror(rte_errno));
2392 		ret = -rte_errno;
2393 		/* Roll back. */
2394 		while (i--) {
2395 			if (!list[i].eth_dev)
2396 				continue;
2397 			mlx5_dev_close(list[i].eth_dev);
2398 			/* mac_addrs must not be freed because in dev_private */
2399 			list[i].eth_dev->data->mac_addrs = NULL;
2400 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2401 		}
2402 		/* Restore original error. */
2403 		rte_errno = -ret;
2404 	} else {
2405 		ret = 0;
2406 	}
2407 exit:
2408 	/*
2409 	 * Do the routine cleanup:
2410 	 * - close opened Netlink sockets
2411 	 * - free allocated spawn data array
2412 	 * - free the Infiniband device list
2413 	 */
2414 	if (nl_rdma >= 0)
2415 		close(nl_rdma);
2416 	if (nl_route >= 0)
2417 		close(nl_route);
2418 	if (list)
2419 		mlx5_free(list);
2420 	MLX5_ASSERT(ibv_list);
2421 	mlx5_glue->free_device_list(ibv_list);
2422 	return ret;
2423 }
2424 
2425 static int
2426 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2427 			  struct rte_eth_devargs *eth_da)
2428 {
2429 	int ret = 0;
2430 
2431 	if (dev->devargs == NULL)
2432 		return 0;
2433 	memset(eth_da, 0, sizeof(*eth_da));
2434 	/* Parse representor information first from class argument. */
2435 	if (dev->devargs->cls_str)
2436 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2437 	if (ret != 0) {
2438 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2439 			dev->devargs->cls_str);
2440 		return -rte_errno;
2441 	}
2442 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE && dev->devargs->args) {
2443 		/* Parse legacy device argument */
2444 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2445 		if (ret) {
2446 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2447 				dev->devargs->args);
2448 			return -rte_errno;
2449 		}
2450 	}
2451 	return 0;
2452 }
2453 
2454 /**
2455  * Callback to register a PCI device.
2456  *
2457  * This function spawns Ethernet devices out of a given PCI device.
2458  *
2459  * @param[in] cdev
2460  *   Pointer to common mlx5 device structure.
2461  * @param[in, out] mkvlist
2462  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2463  *
2464  * @return
2465  *   0 on success, a negative errno value otherwise and rte_errno is set.
2466  */
2467 static int
2468 mlx5_os_pci_probe(struct mlx5_common_device *cdev,
2469 		  struct mlx5_kvargs_ctrl *mkvlist)
2470 {
2471 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2472 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2473 	int ret = 0;
2474 	uint16_t p;
2475 
2476 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2477 	if (ret != 0)
2478 		return ret;
2479 
2480 	if (eth_da.nb_ports > 0) {
2481 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2482 		for (p = 0; p < eth_da.nb_ports; p++) {
2483 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2484 						   eth_da.ports[p], mkvlist);
2485 			if (ret) {
2486 				DRV_LOG(INFO, "Probe of PCI device " PCI_PRI_FMT " "
2487 					"aborted due to proding failure of PF %u",
2488 					pci_dev->addr.domain, pci_dev->addr.bus,
2489 					pci_dev->addr.devid, pci_dev->addr.function,
2490 					eth_da.ports[p]);
2491 				mlx5_net_remove(cdev);
2492 				if (p != 0)
2493 					break;
2494 			}
2495 		}
2496 	} else {
2497 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0, mkvlist);
2498 	}
2499 	return ret;
2500 }
2501 
2502 /* Probe a single SF device on auxiliary bus, no representor support. */
2503 static int
2504 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev,
2505 			struct mlx5_kvargs_ctrl *mkvlist)
2506 {
2507 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2508 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2509 	struct rte_device *dev = cdev->dev;
2510 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2511 	struct rte_eth_dev *eth_dev;
2512 	int ret = 0;
2513 
2514 	/* Parse ethdev devargs. */
2515 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2516 	if (ret != 0)
2517 		return ret;
2518 	/* Init spawn data. */
2519 	spawn.max_port = 1;
2520 	spawn.phys_port = 1;
2521 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2522 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2523 	if (ret < 0) {
2524 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2525 		return ret;
2526 	}
2527 	spawn.ifindex = ret;
2528 	spawn.cdev = cdev;
2529 	/* Spawn device. */
2530 	eth_dev = mlx5_dev_spawn(dev, &spawn, &eth_da, mkvlist);
2531 	if (eth_dev == NULL)
2532 		return -rte_errno;
2533 	/* Post create. */
2534 	eth_dev->intr_handle = adev->intr_handle;
2535 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2536 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2537 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2538 		eth_dev->data->numa_node = dev->numa_node;
2539 	}
2540 	rte_eth_dev_probing_finish(eth_dev);
2541 	return 0;
2542 }
2543 
2544 /**
2545  * Net class driver callback to probe a device.
2546  *
2547  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2548  *
2549  * @param[in] cdev
2550  *   Pointer to the common mlx5 device.
2551  * @param[in, out] mkvlist
2552  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2553  *
2554  * @return
2555  *   0 on success, a negative errno value otherwise and rte_errno is set.
2556  */
2557 int
2558 mlx5_os_net_probe(struct mlx5_common_device *cdev,
2559 		  struct mlx5_kvargs_ctrl *mkvlist)
2560 {
2561 	int ret;
2562 
2563 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2564 		mlx5_pmd_socket_init();
2565 	ret = mlx5_init_once();
2566 	if (ret) {
2567 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2568 			strerror(rte_errno));
2569 		return -rte_errno;
2570 	}
2571 	ret = mlx5_probe_again_args_validate(cdev, mkvlist);
2572 	if (ret) {
2573 		DRV_LOG(ERR, "Probe again parameters are not compatible : %s",
2574 			strerror(rte_errno));
2575 		return -rte_errno;
2576 	}
2577 	if (mlx5_dev_is_pci(cdev->dev))
2578 		return mlx5_os_pci_probe(cdev, mkvlist);
2579 	else
2580 		return mlx5_os_auxiliary_probe(cdev, mkvlist);
2581 }
2582 
2583 /**
2584  * Cleanup resources when the last device is closed.
2585  */
2586 void
2587 mlx5_os_net_cleanup(void)
2588 {
2589 	mlx5_pmd_socket_uninit();
2590 }
2591 
2592 /**
2593  * Install shared asynchronous device events handler.
2594  * This function is implemented to support event sharing
2595  * between multiple ports of single IB device.
2596  *
2597  * @param sh
2598  *   Pointer to mlx5_dev_ctx_shared object.
2599  */
2600 void
2601 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2602 {
2603 	struct ibv_context *ctx = sh->cdev->ctx;
2604 	int nlsk_fd;
2605 
2606 	sh->intr_handle = mlx5_os_interrupt_handler_create
2607 		(RTE_INTR_INSTANCE_F_SHARED, true,
2608 		 ctx->async_fd, mlx5_dev_interrupt_handler, sh);
2609 	if (!sh->intr_handle) {
2610 		DRV_LOG(ERR, "Failed to allocate intr_handle.");
2611 		return;
2612 	}
2613 	nlsk_fd = mlx5_nl_init(NETLINK_ROUTE, RTMGRP_LINK);
2614 	if (nlsk_fd < 0) {
2615 		DRV_LOG(ERR, "Failed to create a socket for Netlink events: %s",
2616 			rte_strerror(rte_errno));
2617 		return;
2618 	}
2619 	sh->intr_handle_nl = mlx5_os_interrupt_handler_create
2620 		(RTE_INTR_INSTANCE_F_SHARED, true,
2621 		 nlsk_fd, mlx5_dev_interrupt_handler_nl, sh);
2622 	if (sh->intr_handle_nl == NULL) {
2623 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2624 		return;
2625 	}
2626 	if (sh->cdev->config.devx) {
2627 #ifdef HAVE_IBV_DEVX_ASYNC
2628 		struct mlx5dv_devx_cmd_comp *devx_comp;
2629 
2630 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2631 		devx_comp = sh->devx_comp;
2632 		if (!devx_comp) {
2633 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2634 			return;
2635 		}
2636 		sh->intr_handle_devx = mlx5_os_interrupt_handler_create
2637 			(RTE_INTR_INSTANCE_F_SHARED, true,
2638 			 devx_comp->fd,
2639 			 mlx5_dev_interrupt_handler_devx, sh);
2640 		if (!sh->intr_handle_devx) {
2641 			DRV_LOG(ERR, "Failed to allocate intr_handle.");
2642 			return;
2643 		}
2644 #endif /* HAVE_IBV_DEVX_ASYNC */
2645 	}
2646 }
2647 
2648 /**
2649  * Uninstall shared asynchronous device events handler.
2650  * This function is implemented to support event sharing
2651  * between multiple ports of single IB device.
2652  *
2653  * @param dev
2654  *   Pointer to mlx5_dev_ctx_shared object.
2655  */
2656 void
2657 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2658 {
2659 	mlx5_os_interrupt_handler_destroy(sh->intr_handle,
2660 					  mlx5_dev_interrupt_handler, sh);
2661 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_nl,
2662 					  mlx5_dev_interrupt_handler_nl, sh);
2663 #ifdef HAVE_IBV_DEVX_ASYNC
2664 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_devx,
2665 					  mlx5_dev_interrupt_handler_devx, sh);
2666 	if (sh->devx_comp)
2667 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2668 #endif
2669 }
2670 
2671 /**
2672  * Read statistics by a named counter.
2673  *
2674  * @param[in] priv
2675  *   Pointer to the private device data structure.
2676  * @param[in] ctr_name
2677  *   Pointer to the name of the statistic counter to read
2678  * @param[out] stat
2679  *   Pointer to read statistic value.
2680  * @return
2681  *   0 on success and stat is valud, 1 if failed to read the value
2682  *   rte_errno is set.
2683  *
2684  */
2685 int
2686 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2687 		      uint64_t *stat)
2688 {
2689 	int fd;
2690 
2691 	if (priv->sh) {
2692 		if (priv->q_counters != NULL &&
2693 		    strcmp(ctr_name, "out_of_buffer") == 0)
2694 			return mlx5_devx_cmd_queue_counter_query
2695 					(priv->q_counters, 0, (uint32_t *)stat);
2696 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2697 		      priv->sh->ibdev_path,
2698 		      priv->dev_port,
2699 		      ctr_name);
2700 		fd = open(path, O_RDONLY);
2701 		/*
2702 		 * in switchdev the file location is not per port
2703 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2704 		 */
2705 		if (fd == -1) {
2706 			MKSTR(path1, "%s/hw_counters/%s",
2707 			      priv->sh->ibdev_path,
2708 			      ctr_name);
2709 			fd = open(path1, O_RDONLY);
2710 		}
2711 		if (fd != -1) {
2712 			char buf[21] = {'\0'};
2713 			ssize_t n = read(fd, buf, sizeof(buf));
2714 
2715 			close(fd);
2716 			if (n != -1) {
2717 				*stat = strtoull(buf, NULL, 10);
2718 				return 0;
2719 			}
2720 		}
2721 	}
2722 	*stat = 0;
2723 	return 1;
2724 }
2725 
2726 /**
2727  * Remove a MAC address from device
2728  *
2729  * @param dev
2730  *   Pointer to Ethernet device structure.
2731  * @param index
2732  *   MAC address index.
2733  */
2734 void
2735 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2736 {
2737 	struct mlx5_priv *priv = dev->data->dev_private;
2738 	const int vf = priv->sh->dev_cap.vf;
2739 
2740 	if (vf)
2741 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2742 					mlx5_ifindex(dev), priv->mac_own,
2743 					&dev->data->mac_addrs[index], index);
2744 }
2745 
2746 /**
2747  * Adds a MAC address to the device
2748  *
2749  * @param dev
2750  *   Pointer to Ethernet device structure.
2751  * @param mac_addr
2752  *   MAC address to register.
2753  * @param index
2754  *   MAC address index.
2755  *
2756  * @return
2757  *   0 on success, a negative errno value otherwise
2758  */
2759 int
2760 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2761 		     uint32_t index)
2762 {
2763 	struct mlx5_priv *priv = dev->data->dev_private;
2764 	const int vf = priv->sh->dev_cap.vf;
2765 	int ret = 0;
2766 
2767 	if (vf)
2768 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2769 					   mlx5_ifindex(dev), priv->mac_own,
2770 					   mac, index);
2771 	return ret;
2772 }
2773 
2774 /**
2775  * Modify a VF MAC address
2776  *
2777  * @param priv
2778  *   Pointer to device private data.
2779  * @param mac_addr
2780  *   MAC address to modify into.
2781  * @param iface_idx
2782  *   Net device interface index
2783  * @param vf_index
2784  *   VF index
2785  *
2786  * @return
2787  *   0 on success, a negative errno value otherwise
2788  */
2789 int
2790 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2791 			   unsigned int iface_idx,
2792 			   struct rte_ether_addr *mac_addr,
2793 			   int vf_index)
2794 {
2795 	return mlx5_nl_vf_mac_addr_modify
2796 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2797 }
2798 
2799 /**
2800  * Set device promiscuous mode
2801  *
2802  * @param dev
2803  *   Pointer to Ethernet device structure.
2804  * @param enable
2805  *   0 - promiscuous is disabled, otherwise - enabled
2806  *
2807  * @return
2808  *   0 on success, a negative error value otherwise
2809  */
2810 int
2811 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2812 {
2813 	struct mlx5_priv *priv = dev->data->dev_private;
2814 
2815 	return mlx5_nl_promisc(priv->nl_socket_route,
2816 			       mlx5_ifindex(dev), !!enable);
2817 }
2818 
2819 /**
2820  * Set device promiscuous mode
2821  *
2822  * @param dev
2823  *   Pointer to Ethernet device structure.
2824  * @param enable
2825  *   0 - all multicase is disabled, otherwise - enabled
2826  *
2827  * @return
2828  *   0 on success, a negative error value otherwise
2829  */
2830 int
2831 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2832 {
2833 	struct mlx5_priv *priv = dev->data->dev_private;
2834 
2835 	return mlx5_nl_allmulti(priv->nl_socket_route,
2836 				mlx5_ifindex(dev), !!enable);
2837 }
2838 
2839 /**
2840  * Flush device MAC addresses
2841  *
2842  * @param dev
2843  *   Pointer to Ethernet device structure.
2844  *
2845  */
2846 void
2847 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2848 {
2849 	struct mlx5_priv *priv = dev->data->dev_private;
2850 
2851 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2852 			       dev->data->mac_addrs,
2853 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2854 }
2855