xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 5e9f9a280640c8db77e5fc57f12aeae11ef09fba)
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 #ifdef HAVE_MLX5DV_DR
679 	int i;
680 #endif
681 
682 	MLX5_ASSERT(sh && sh->refcnt);
683 	if (sh->refcnt > 1)
684 		return;
685 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
686 #ifdef HAVE_MLX5DV_DR
687 	if (sh->rx_domain) {
688 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
689 		sh->rx_domain = NULL;
690 	}
691 	if (sh->tx_domain) {
692 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
693 		sh->tx_domain = NULL;
694 	}
695 #ifdef HAVE_MLX5DV_DR_ESWITCH
696 	if (sh->fdb_domain) {
697 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
698 		sh->fdb_domain = NULL;
699 	}
700 	if (sh->dr_drop_action) {
701 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
702 		sh->dr_drop_action = NULL;
703 	}
704 #endif
705 	if (sh->pop_vlan_action) {
706 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
707 		sh->pop_vlan_action = NULL;
708 	}
709 	for (i = 0; i < MLX5DR_TABLE_TYPE_MAX; i++) {
710 		if (sh->send_to_kernel_action[i].action) {
711 			void *action = sh->send_to_kernel_action[i].action;
712 
713 			mlx5_glue->destroy_flow_action(action);
714 			sh->send_to_kernel_action[i].action = NULL;
715 		}
716 		if (sh->send_to_kernel_action[i].tbl) {
717 			struct mlx5_flow_tbl_resource *tbl =
718 					sh->send_to_kernel_action[i].tbl;
719 
720 			flow_dv_tbl_resource_release(sh, tbl);
721 			sh->send_to_kernel_action[i].tbl = NULL;
722 		}
723 	}
724 #endif /* HAVE_MLX5DV_DR */
725 	if (sh->default_miss_action)
726 		mlx5_glue->destroy_flow_action
727 				(sh->default_miss_action);
728 	if (sh->encaps_decaps) {
729 		mlx5_hlist_destroy(sh->encaps_decaps);
730 		sh->encaps_decaps = NULL;
731 	}
732 	if (sh->modify_cmds) {
733 		mlx5_hlist_destroy(sh->modify_cmds);
734 		sh->modify_cmds = NULL;
735 	}
736 	if (sh->tag_table) {
737 		/* tags should be destroyed with flow before. */
738 		mlx5_hlist_destroy(sh->tag_table);
739 		sh->tag_table = NULL;
740 	}
741 	if (sh->tunnel_hub) {
742 		mlx5_release_tunnel_hub(sh, priv->dev_port);
743 		sh->tunnel_hub = NULL;
744 	}
745 	mlx5_free_table_hash_list(priv);
746 	if (sh->port_id_action_list) {
747 		mlx5_list_destroy(sh->port_id_action_list);
748 		sh->port_id_action_list = NULL;
749 	}
750 	if (sh->push_vlan_action_list) {
751 		mlx5_list_destroy(sh->push_vlan_action_list);
752 		sh->push_vlan_action_list = NULL;
753 	}
754 	if (sh->sample_action_list) {
755 		mlx5_list_destroy(sh->sample_action_list);
756 		sh->sample_action_list = NULL;
757 	}
758 	if (sh->dest_array_list) {
759 		mlx5_list_destroy(sh->dest_array_list);
760 		sh->dest_array_list = NULL;
761 	}
762 }
763 
764 /**
765  * Initialize shared data between primary and secondary process.
766  *
767  * A memzone is reserved by primary process and secondary processes attach to
768  * the memzone.
769  *
770  * @return
771  *   0 on success, a negative errno value otherwise and rte_errno is set.
772  */
773 static int
774 mlx5_init_shared_data(void)
775 {
776 	const struct rte_memzone *mz;
777 	int ret = 0;
778 
779 	rte_spinlock_lock(&mlx5_shared_data_lock);
780 	if (mlx5_shared_data == NULL) {
781 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
782 			/* Allocate shared memory. */
783 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
784 						 sizeof(*mlx5_shared_data),
785 						 SOCKET_ID_ANY, 0);
786 			if (mz == NULL) {
787 				DRV_LOG(ERR,
788 					"Cannot allocate mlx5 shared data");
789 				ret = -rte_errno;
790 				goto error;
791 			}
792 			mlx5_shared_data = mz->addr;
793 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
794 			rte_spinlock_init(&mlx5_shared_data->lock);
795 		} else {
796 			/* Lookup allocated shared memory. */
797 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
798 			if (mz == NULL) {
799 				DRV_LOG(ERR,
800 					"Cannot attach mlx5 shared data");
801 				ret = -rte_errno;
802 				goto error;
803 			}
804 			mlx5_shared_data = mz->addr;
805 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
806 		}
807 	}
808 error:
809 	rte_spinlock_unlock(&mlx5_shared_data_lock);
810 	return ret;
811 }
812 
813 /**
814  * PMD global initialization.
815  *
816  * Independent from individual device, this function initializes global
817  * per-PMD data structures distinguishing primary and secondary processes.
818  * Hence, each initialization is called once per a process.
819  *
820  * @return
821  *   0 on success, a negative errno value otherwise and rte_errno is set.
822  */
823 static int
824 mlx5_init_once(void)
825 {
826 	struct mlx5_shared_data *sd;
827 	struct mlx5_local_data *ld = &mlx5_local_data;
828 	int ret = 0;
829 
830 	if (mlx5_init_shared_data())
831 		return -rte_errno;
832 	sd = mlx5_shared_data;
833 	MLX5_ASSERT(sd);
834 	rte_spinlock_lock(&sd->lock);
835 	switch (rte_eal_process_type()) {
836 	case RTE_PROC_PRIMARY:
837 		if (sd->init_done)
838 			break;
839 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
840 					   mlx5_mp_os_primary_handle);
841 		if (ret)
842 			goto out;
843 		sd->init_done = true;
844 		break;
845 	case RTE_PROC_SECONDARY:
846 		if (ld->init_done)
847 			break;
848 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
849 					     mlx5_mp_os_secondary_handle);
850 		if (ret)
851 			goto out;
852 		++sd->secondary_cnt;
853 		ld->init_done = true;
854 		break;
855 	default:
856 		break;
857 	}
858 out:
859 	rte_spinlock_unlock(&sd->lock);
860 	return ret;
861 }
862 
863 /**
864  * DR flow drop action support detect.
865  *
866  * @param dev
867  *   Pointer to rte_eth_dev structure.
868  *
869  */
870 static void
871 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
872 {
873 #ifdef HAVE_MLX5DV_DR
874 	struct mlx5_priv *priv = dev->data->dev_private;
875 
876 	if (!priv->sh->config.dv_flow_en || !priv->sh->dr_drop_action)
877 		return;
878 	/**
879 	 * DR supports drop action placeholder when it is supported;
880 	 * otherwise, use the queue drop action.
881 	 */
882 	if (!priv->sh->drop_action_check_flag) {
883 		if (!mlx5_flow_discover_dr_action_support(dev))
884 			priv->sh->dr_root_drop_action_en = 1;
885 		priv->sh->drop_action_check_flag = 1;
886 	}
887 	if (priv->sh->dr_root_drop_action_en)
888 		priv->root_drop_action = priv->sh->dr_drop_action;
889 	else
890 		priv->root_drop_action = priv->drop_queue.hrxq->action;
891 #endif
892 }
893 
894 static void
895 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
896 {
897 	struct mlx5_priv *priv = dev->data->dev_private;
898 	void *ctx = priv->sh->cdev->ctx;
899 
900 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
901 	if (!priv->q_counters) {
902 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
903 		struct ibv_wq *wq;
904 
905 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
906 			"by DevX - fall-back to use the kernel driver global "
907 			"queue counter.", dev->data->port_id);
908 		/* Create WQ by kernel and query its queue counter ID. */
909 		if (cq) {
910 			wq = mlx5_glue->create_wq(ctx,
911 						  &(struct ibv_wq_init_attr){
912 						    .wq_type = IBV_WQT_RQ,
913 						    .max_wr = 1,
914 						    .max_sge = 1,
915 						    .pd = priv->sh->cdev->pd,
916 						    .cq = cq,
917 						});
918 			if (wq) {
919 				/* Counter is assigned only on RDY state. */
920 				int ret = mlx5_glue->modify_wq(wq,
921 						 &(struct ibv_wq_attr){
922 						 .attr_mask = IBV_WQ_ATTR_STATE,
923 						 .wq_state = IBV_WQS_RDY,
924 						});
925 
926 				if (ret == 0)
927 					mlx5_devx_cmd_wq_query(wq,
928 							 &priv->counter_set_id);
929 				claim_zero(mlx5_glue->destroy_wq(wq));
930 			}
931 			claim_zero(mlx5_glue->destroy_cq(cq));
932 		}
933 	} else {
934 		priv->counter_set_id = priv->q_counters->id;
935 	}
936 	if (priv->counter_set_id == 0)
937 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
938 			"available.", dev->data->port_id);
939 }
940 
941 /**
942  * Check if representor spawn info match devargs.
943  *
944  * @param spawn
945  *   Verbs device parameters (name, port, switch_info) to spawn.
946  * @param eth_da
947  *   Device devargs to probe.
948  *
949  * @return
950  *   Match result.
951  */
952 static bool
953 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
954 		       struct rte_eth_devargs *eth_da)
955 {
956 	struct mlx5_switch_info *switch_info = &spawn->info;
957 	unsigned int p, f;
958 	uint16_t id;
959 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
960 						      eth_da->type);
961 
962 	switch (eth_da->type) {
963 	case RTE_ETH_REPRESENTOR_SF:
964 		if (!(spawn->info.port_name == -1 &&
965 		      switch_info->name_type ==
966 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
967 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
968 			rte_errno = EBUSY;
969 			return false;
970 		}
971 		break;
972 	case RTE_ETH_REPRESENTOR_VF:
973 		/* Allows HPF representor index -1 as exception. */
974 		if (!(spawn->info.port_name == -1 &&
975 		      switch_info->name_type ==
976 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
977 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
978 			rte_errno = EBUSY;
979 			return false;
980 		}
981 		break;
982 	case RTE_ETH_REPRESENTOR_NONE:
983 		rte_errno = EBUSY;
984 		return false;
985 	default:
986 		rte_errno = ENOTSUP;
987 		DRV_LOG(ERR, "unsupported representor type");
988 		return false;
989 	}
990 	/* Check representor ID: */
991 	for (p = 0; p < eth_da->nb_ports; ++p) {
992 		if (spawn->pf_bond < 0) {
993 			/* For non-LAG mode, allow and ignore pf. */
994 			switch_info->pf_num = eth_da->ports[p];
995 			repr_id = mlx5_representor_id_encode(switch_info,
996 							     eth_da->type);
997 		}
998 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
999 			id = MLX5_REPRESENTOR_ID
1000 				(eth_da->ports[p], eth_da->type,
1001 				 eth_da->representor_ports[f]);
1002 			if (repr_id == id)
1003 				return true;
1004 		}
1005 	}
1006 	rte_errno = EBUSY;
1007 	return false;
1008 }
1009 
1010 /**
1011  * Spawn an Ethernet device from Verbs information.
1012  *
1013  * @param dpdk_dev
1014  *   Backing DPDK device.
1015  * @param spawn
1016  *   Verbs device parameters (name, port, switch_info) to spawn.
1017  * @param eth_da
1018  *   Device arguments.
1019  * @param mkvlist
1020  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1021  *
1022  * @return
1023  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1024  *   is set. The following errors are defined:
1025  *
1026  *   EBUSY: device is not supposed to be spawned.
1027  *   EEXIST: device is already spawned
1028  */
1029 static struct rte_eth_dev *
1030 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1031 	       struct mlx5_dev_spawn_data *spawn,
1032 	       struct rte_eth_devargs *eth_da,
1033 	       struct mlx5_kvargs_ctrl *mkvlist)
1034 {
1035 	const struct mlx5_switch_info *switch_info = &spawn->info;
1036 	struct mlx5_dev_ctx_shared *sh = NULL;
1037 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
1038 	struct rte_eth_dev *eth_dev = NULL;
1039 	struct mlx5_priv *priv = NULL;
1040 	int err = 0;
1041 	struct rte_ether_addr mac;
1042 	char name[RTE_ETH_NAME_MAX_LEN];
1043 	int own_domain_id = 0;
1044 	uint16_t port_id;
1045 	struct mlx5_port_info vport_info = { .query_flags = 0 };
1046 	int nl_rdma;
1047 	int i;
1048 
1049 	/* Determine if this port representor is supposed to be spawned. */
1050 	if (switch_info->representor && dpdk_dev->devargs &&
1051 	    !mlx5_representor_match(spawn, eth_da))
1052 		return NULL;
1053 	/* Build device name. */
1054 	if (spawn->pf_bond < 0) {
1055 		/* Single device. */
1056 		if (!switch_info->representor)
1057 			strlcpy(name, dpdk_dev->name, sizeof(name));
1058 		else
1059 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
1060 				 dpdk_dev->name,
1061 				 switch_info->name_type ==
1062 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1063 				 switch_info->port_name);
1064 	} else {
1065 		/* Bonding device. */
1066 		if (!switch_info->representor) {
1067 			err = snprintf(name, sizeof(name), "%s_%s",
1068 				       dpdk_dev->name, spawn->phys_dev_name);
1069 		} else {
1070 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1071 				dpdk_dev->name, spawn->phys_dev_name,
1072 				switch_info->ctrl_num,
1073 				switch_info->pf_num,
1074 				switch_info->name_type ==
1075 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1076 				switch_info->port_name);
1077 		}
1078 	}
1079 	if (err >= (int)sizeof(name))
1080 		DRV_LOG(WARNING, "device name overflow %s", name);
1081 	/* check if the device is already spawned */
1082 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1083 		/*
1084 		 * When device is already spawned, its devargs should be set
1085 		 * as used. otherwise, mlx5_kvargs_validate() will fail.
1086 		 */
1087 		if (mkvlist)
1088 			mlx5_port_args_set_used(name, port_id, mkvlist);
1089 		rte_errno = EEXIST;
1090 		return NULL;
1091 	}
1092 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1093 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1094 		struct mlx5_mp_id mp_id;
1095 		int fd;
1096 
1097 		eth_dev = rte_eth_dev_attach_secondary(name);
1098 		if (eth_dev == NULL) {
1099 			DRV_LOG(ERR, "can not attach rte ethdev");
1100 			rte_errno = ENOMEM;
1101 			return NULL;
1102 		}
1103 		eth_dev->device = dpdk_dev;
1104 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1105 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1106 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1107 		err = mlx5_proc_priv_init(eth_dev);
1108 		if (err)
1109 			return NULL;
1110 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1111 		/* Receive command fd from primary process */
1112 		fd = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1113 		if (fd < 0)
1114 			goto err_secondary;
1115 		/* Remap UAR for Tx queues. */
1116 		err = mlx5_tx_uar_init_secondary(eth_dev, fd);
1117 		close(fd);
1118 		if (err)
1119 			goto err_secondary;
1120 		/*
1121 		 * Ethdev pointer is still required as input since
1122 		 * the primary device is not accessible from the
1123 		 * secondary process.
1124 		 */
1125 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1126 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1127 		return eth_dev;
1128 err_secondary:
1129 		mlx5_dev_close(eth_dev);
1130 		return NULL;
1131 	}
1132 	sh = mlx5_alloc_shared_dev_ctx(spawn, mkvlist);
1133 	if (!sh)
1134 		return NULL;
1135 	nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
1136 	/* Check port status. */
1137 	if (spawn->phys_port <= UINT8_MAX) {
1138 		/* Legacy Verbs api only support u8 port number. */
1139 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1140 					    &port_attr);
1141 		if (err) {
1142 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1143 			goto error;
1144 		}
1145 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1146 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1147 			err = EINVAL;
1148 			goto error;
1149 		}
1150 	} else if (nl_rdma >= 0) {
1151 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1152 		err = mlx5_nl_port_state(nl_rdma,
1153 			spawn->phys_dev_name,
1154 			spawn->phys_port);
1155 		if (err < 0) {
1156 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1157 				strerror(rte_errno));
1158 			err = -rte_errno;
1159 			goto error;
1160 		}
1161 		port_attr.state = (enum ibv_port_state)err;
1162 	}
1163 	if (port_attr.state != IBV_PORT_ACTIVE)
1164 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1165 			mlx5_glue->port_state_str(port_attr.state),
1166 			port_attr.state);
1167 	/* Allocate private eth device data. */
1168 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1169 			   sizeof(*priv),
1170 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1171 	if (priv == NULL) {
1172 		DRV_LOG(ERR, "priv allocation failure");
1173 		err = ENOMEM;
1174 		goto error;
1175 	}
1176 	/*
1177 	 * When user configures remote PD and CTX and device creates RxQ by
1178 	 * DevX, external RxQ is both supported and requested.
1179 	 */
1180 	if (mlx5_imported_pd_and_ctx(sh->cdev) && mlx5_devx_obj_ops_en(sh)) {
1181 		priv->ext_rxqs = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1182 					     sizeof(struct mlx5_external_rxq) *
1183 					     MLX5_MAX_EXT_RX_QUEUES, 0,
1184 					     SOCKET_ID_ANY);
1185 		if (priv->ext_rxqs == NULL) {
1186 			DRV_LOG(ERR, "Fail to allocate external RxQ array.");
1187 			err = ENOMEM;
1188 			goto error;
1189 		}
1190 		DRV_LOG(DEBUG, "External RxQ is supported.");
1191 	}
1192 	priv->sh = sh;
1193 	priv->dev_port = spawn->phys_port;
1194 	priv->pci_dev = spawn->pci_dev;
1195 	priv->mtu = RTE_ETHER_MTU;
1196 	/* Some internal functions rely on Netlink sockets, open them now. */
1197 	priv->nl_socket_rdma = nl_rdma;
1198 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE, 0);
1199 	priv->representor = !!switch_info->representor;
1200 	priv->master = !!switch_info->master;
1201 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1202 	priv->vport_meta_tag = 0;
1203 	priv->vport_meta_mask = 0;
1204 	priv->pf_bond = spawn->pf_bond;
1205 
1206 	DRV_LOG(DEBUG,
1207 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1208 		priv->dev_port, dpdk_dev->bus->name,
1209 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1210 		priv->master, priv->representor, priv->pf_bond);
1211 
1212 	/*
1213 	 * If we have E-Switch we should determine the vport attributes.
1214 	 * E-Switch may use either source vport field or reg_c[0] metadata
1215 	 * register to match on vport index. The engaged part of metadata
1216 	 * register is defined by mask.
1217 	 */
1218 	if (sh->esw_mode) {
1219 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1220 						 spawn->phys_port,
1221 						 &vport_info);
1222 		if (err) {
1223 			DRV_LOG(WARNING,
1224 				"Cannot query devx port %d on device %s",
1225 				spawn->phys_port, spawn->phys_dev_name);
1226 			vport_info.query_flags = 0;
1227 		}
1228 	}
1229 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1230 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1231 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1232 		if (!priv->vport_meta_mask) {
1233 			DRV_LOG(ERR,
1234 				"vport zero mask for port %d on bonding device %s",
1235 				spawn->phys_port, spawn->phys_dev_name);
1236 			err = ENOTSUP;
1237 			goto error;
1238 		}
1239 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1240 			DRV_LOG(ERR,
1241 				"Invalid vport tag for port %d on bonding device %s",
1242 				spawn->phys_port, spawn->phys_dev_name);
1243 			err = ENOTSUP;
1244 			goto error;
1245 		}
1246 	}
1247 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1248 		priv->vport_id = vport_info.vport_id;
1249 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1250 		DRV_LOG(ERR,
1251 			"Cannot deduce vport index for port %d on bonding device %s",
1252 			spawn->phys_port, spawn->phys_dev_name);
1253 		err = ENOTSUP;
1254 		goto error;
1255 	} else {
1256 		/*
1257 		 * Suppose vport index in compatible way. Kernel/rdma_core
1258 		 * support single E-Switch per PF configurations only and
1259 		 * vport_id field contains the vport index for associated VF,
1260 		 * which is deduced from representor port name.
1261 		 * For example, let's have the IB device port 10, it has
1262 		 * attached network device eth0, which has port name attribute
1263 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1264 		 * as 3 (2+1). This assigning schema should be changed if the
1265 		 * multiple E-Switch instances per PF configurations or/and PCI
1266 		 * subfunctions are added.
1267 		 */
1268 		priv->vport_id = switch_info->representor ?
1269 				 switch_info->port_name + 1 : -1;
1270 	}
1271 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1272 							  eth_da->type);
1273 	/*
1274 	 * Look for sibling devices in order to reuse their switch domain
1275 	 * if any, otherwise allocate one.
1276 	 */
1277 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1278 		const struct mlx5_priv *opriv =
1279 			rte_eth_devices[port_id].data->dev_private;
1280 
1281 		if (!opriv ||
1282 		    opriv->sh != priv->sh ||
1283 			opriv->domain_id ==
1284 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1285 			continue;
1286 		priv->domain_id = opriv->domain_id;
1287 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1288 			priv->dev_port, priv->domain_id);
1289 		break;
1290 	}
1291 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1292 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1293 		if (err) {
1294 			err = rte_errno;
1295 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1296 				strerror(rte_errno));
1297 			goto error;
1298 		}
1299 		own_domain_id = 1;
1300 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1301 			priv->dev_port, priv->domain_id);
1302 	}
1303 	if (sh->cdev->config.devx) {
1304 		struct mlx5_hca_attr *hca_attr = &sh->cdev->config.hca_attr;
1305 
1306 		sh->steering_format_version = hca_attr->steering_format_version;
1307 #if defined(HAVE_MLX5DV_DR) && \
1308 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1309 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1310 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1311 		    sh->config.dv_flow_en) {
1312 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1313 			/*
1314 			 * Meter needs two REG_C's for color match and pre-sfx
1315 			 * flow match. Here get the REG_C for color match.
1316 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1317 			 */
1318 			reg_c_mask &= 0xfc;
1319 			if (rte_popcount32(reg_c_mask) < 1) {
1320 				priv->mtr_en = 0;
1321 				DRV_LOG(WARNING, "No available register for"
1322 					" meter.");
1323 			} else {
1324 				/*
1325 				 * The meter color register is used by the
1326 				 * flow-hit feature as well.
1327 				 * The flow-hit feature must use REG_C_3
1328 				 * Prefer REG_C_3 if it is available.
1329 				 */
1330 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1331 					sh->registers.aso_reg = REG_C_3;
1332 				else
1333 					sh->registers.aso_reg =
1334 						ffs(reg_c_mask) - 1 + REG_C_0;
1335 				priv->mtr_en = 1;
1336 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1337 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1338 					sh->registers.aso_reg);
1339 			}
1340 		}
1341 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1342 			uint32_t log_obj_size =
1343 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1344 			if (log_obj_size >=
1345 			    hca_attr->qos.log_meter_aso_granularity &&
1346 			    log_obj_size <=
1347 			    hca_attr->qos.log_meter_aso_max_alloc)
1348 				sh->meter_aso_en = 1;
1349 		}
1350 		if (priv->mtr_en) {
1351 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1352 			if (err) {
1353 				err = -err;
1354 				goto error;
1355 			}
1356 		}
1357 		if (hca_attr->flow.tunnel_header_0_1)
1358 			sh->tunnel_header_0_1 = 1;
1359 		if (hca_attr->flow.tunnel_header_2_3)
1360 			sh->tunnel_header_2_3 = 1;
1361 #endif
1362 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1363 		if (hca_attr->flow_hit_aso && sh->registers.aso_reg == REG_C_3) {
1364 			sh->flow_hit_aso_en = 1;
1365 			err = mlx5_flow_aso_age_mng_init(sh);
1366 			if (err) {
1367 				err = -err;
1368 				goto error;
1369 			}
1370 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1371 		}
1372 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1373 #if defined (HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1374     defined (HAVE_MLX5_DR_ACTION_ASO_CT)
1375 		/* HWS create CT ASO SQ based on HWS configure queue number. */
1376 		if (sh->config.dv_flow_en != 2 &&
1377 		    hca_attr->ct_offload && sh->registers.aso_reg == REG_C_3) {
1378 			err = mlx5_flow_aso_ct_mng_init(sh);
1379 			if (err) {
1380 				err = -err;
1381 				goto error;
1382 			}
1383 			DRV_LOG(DEBUG, "CT ASO is supported.");
1384 			sh->ct_aso_en = 1;
1385 		}
1386 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1387 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1388 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1389 		    sh->config.dv_flow_en) {
1390 			priv->sampler_en = 1;
1391 			DRV_LOG(DEBUG, "Sampler enabled!");
1392 		} else {
1393 			priv->sampler_en = 0;
1394 			if (!hca_attr->log_max_ft_sampler_num)
1395 				DRV_LOG(WARNING,
1396 					"No available register for sampler.");
1397 			else
1398 				DRV_LOG(DEBUG, "DV flow is not supported!");
1399 		}
1400 #endif
1401 		if (hca_attr->lag_rx_port_affinity) {
1402 			sh->lag_rx_port_affinity_en = 1;
1403 			DRV_LOG(DEBUG, "LAG Rx Port Affinity enabled");
1404 		}
1405 		priv->num_lag_ports = hca_attr->num_lag_ports;
1406 		DRV_LOG(DEBUG, "The number of lag ports is %d", priv->num_lag_ports);
1407 	}
1408 	/* Process parameters and store port configuration on priv structure. */
1409 	err = mlx5_port_args_config(priv, mkvlist, &priv->config);
1410 	if (err) {
1411 		err = rte_errno;
1412 		DRV_LOG(ERR, "Failed to process port configure: %s",
1413 			strerror(rte_errno));
1414 		goto error;
1415 	}
1416 	eth_dev = rte_eth_dev_allocate(name);
1417 	if (eth_dev == NULL) {
1418 		DRV_LOG(ERR, "can not allocate rte ethdev");
1419 		err = ENOMEM;
1420 		goto error;
1421 	}
1422 	if (priv->representor) {
1423 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1424 		eth_dev->data->representor_id = priv->representor_id;
1425 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1426 			struct mlx5_priv *opriv =
1427 				rte_eth_devices[port_id].data->dev_private;
1428 			if (opriv &&
1429 			    opriv->master &&
1430 			    opriv->domain_id == priv->domain_id &&
1431 			    opriv->sh == priv->sh) {
1432 				eth_dev->data->backer_port_id = port_id;
1433 				break;
1434 			}
1435 		}
1436 		if (port_id >= RTE_MAX_ETHPORTS)
1437 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1438 	}
1439 	priv->mp_id.port_id = eth_dev->data->port_id;
1440 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1441 	/*
1442 	 * Store associated network device interface index. This index
1443 	 * is permanent throughout the lifetime of device. So, we may store
1444 	 * the ifindex here and use the cached value further.
1445 	 */
1446 	MLX5_ASSERT(spawn->ifindex);
1447 	priv->if_index = spawn->ifindex;
1448 	priv->lag_affinity_idx = sh->refcnt - 1;
1449 	eth_dev->data->dev_private = priv;
1450 	priv->dev_data = eth_dev->data;
1451 	eth_dev->data->mac_addrs = priv->mac;
1452 	eth_dev->device = dpdk_dev;
1453 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1454 	/* Configure the first MAC address by default. */
1455 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1456 		DRV_LOG(ERR,
1457 			"port %u cannot get MAC address, is mlx5_en"
1458 			" loaded? (errno: %s)",
1459 			eth_dev->data->port_id, strerror(rte_errno));
1460 		err = ENODEV;
1461 		goto error;
1462 	}
1463 	DRV_LOG(INFO,
1464 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1465 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1466 #ifdef RTE_LIBRTE_MLX5_DEBUG
1467 	{
1468 		char ifname[MLX5_NAMESIZE];
1469 
1470 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1471 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1472 				eth_dev->data->port_id, ifname);
1473 		else
1474 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1475 				eth_dev->data->port_id);
1476 	}
1477 #endif
1478 	/* Get actual MTU if possible. */
1479 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1480 	if (err) {
1481 		err = rte_errno;
1482 		goto error;
1483 	}
1484 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1485 		priv->mtu);
1486 	/* Initialize burst functions to prevent crashes before link-up. */
1487 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1488 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1489 	eth_dev->dev_ops = &mlx5_dev_ops;
1490 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1491 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1492 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1493 	/* Register MAC address. */
1494 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1495 	if (sh->dev_cap.vf && sh->config.vf_nl_en)
1496 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1497 				      mlx5_ifindex(eth_dev),
1498 				      eth_dev->data->mac_addrs,
1499 				      MLX5_MAX_MAC_ADDRESSES);
1500 	priv->ctrl_flows = 0;
1501 	rte_spinlock_init(&priv->flow_list_lock);
1502 	TAILQ_INIT(&priv->flow_meters);
1503 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1504 	if (!priv->mtr_profile_tbl)
1505 		goto error;
1506 	/* Bring Ethernet device up. */
1507 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1508 		eth_dev->data->port_id);
1509 	/* Read link status in case it is up and there will be no event. */
1510 	mlx5_link_update(eth_dev, 0);
1511 	/* Watch LSC interrupts between port probe and port start. */
1512 	priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1513 							eth_dev->data->port_id;
1514 	mlx5_set_link_up(eth_dev);
1515 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1516 		icfg[i].release_mem_en = !!sh->config.reclaim_mode;
1517 		if (sh->config.reclaim_mode)
1518 			icfg[i].per_core_cache = 0;
1519 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1520 		if (!priv->flows[i])
1521 			goto error;
1522 	}
1523 	/* Create context for virtual machine VLAN workaround. */
1524 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1525 	if (sh->config.dv_flow_en) {
1526 		err = mlx5_alloc_shared_dr(priv);
1527 		if (err)
1528 			goto error;
1529 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1530 			goto error;
1531 	}
1532 	if (mlx5_devx_obj_ops_en(sh)) {
1533 		priv->obj_ops = devx_obj_ops;
1534 		mlx5_queue_counter_id_prepare(eth_dev);
1535 		priv->obj_ops.lb_dummy_queue_create =
1536 					mlx5_rxq_ibv_obj_dummy_lb_create;
1537 		priv->obj_ops.lb_dummy_queue_release =
1538 					mlx5_rxq_ibv_obj_dummy_lb_release;
1539 	} else if (spawn->max_port > UINT8_MAX) {
1540 		/* Verbs can't support ports larger than 255 by design. */
1541 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1542 		err = ENOTSUP;
1543 		goto error;
1544 	} else {
1545 		priv->obj_ops = ibv_obj_ops;
1546 	}
1547 	if (sh->config.tx_pp &&
1548 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1549 		/*
1550 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1551 		 * packet pacing and already checked above.
1552 		 * Hence, we should only make sure the SQs will be created
1553 		 * with DevX, not with Verbs.
1554 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1555 		 * with Clock Queue UAR as required for Tx scheduling.
1556 		 */
1557 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1558 		err = ENODEV;
1559 		goto error;
1560 	}
1561 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1562 	if (!priv->drop_queue.hrxq)
1563 		goto error;
1564 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1565 				       mlx5_hrxq_create_cb,
1566 				       mlx5_hrxq_match_cb,
1567 				       mlx5_hrxq_remove_cb,
1568 				       mlx5_hrxq_clone_cb,
1569 				       mlx5_hrxq_clone_free_cb);
1570 	if (!priv->hrxqs)
1571 		goto error;
1572 	mlx5_set_metadata_mask(eth_dev);
1573 	if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1574 	    !priv->sh->dv_regc0_mask) {
1575 		DRV_LOG(ERR, "metadata mode %u is not supported "
1576 			     "(no metadata reg_c[0] is available)",
1577 			     sh->config.dv_xmeta_en);
1578 			err = ENOTSUP;
1579 			goto error;
1580 	}
1581 	rte_rwlock_init(&priv->ind_tbls_lock);
1582 	if (priv->sh->config.dv_flow_en == 2) {
1583 #ifdef HAVE_MLX5_HWS_SUPPORT
1584 		if (priv->sh->config.dv_esw_en) {
1585 			uint32_t usable_bits;
1586 			uint32_t required_bits;
1587 
1588 			if (priv->sh->dv_regc0_mask == UINT32_MAX) {
1589 				DRV_LOG(ERR, "E-Switch port metadata is required when using HWS "
1590 					     "but it is disabled (configure it through devlink)");
1591 				err = ENOTSUP;
1592 				goto error;
1593 			}
1594 			if (priv->sh->dv_regc0_mask == 0) {
1595 				DRV_LOG(ERR, "E-Switch with HWS is not supported "
1596 					     "(no available bits in reg_c[0])");
1597 				err = ENOTSUP;
1598 				goto error;
1599 			}
1600 			usable_bits = rte_popcount32(priv->sh->dv_regc0_mask);
1601 			required_bits = rte_popcount32(priv->vport_meta_mask);
1602 			if (usable_bits < required_bits) {
1603 				DRV_LOG(ERR, "Not enough bits available in reg_c[0] to provide "
1604 					     "representor matching.");
1605 				err = ENOTSUP;
1606 				goto error;
1607 			}
1608 		}
1609 		if (priv->vport_meta_mask)
1610 			flow_hw_set_port_info(eth_dev);
1611 		if (priv->sh->config.dv_esw_en &&
1612 		    priv->sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1613 		    priv->sh->config.dv_xmeta_en != MLX5_XMETA_MODE_META32_HWS) {
1614 			DRV_LOG(ERR,
1615 				"metadata mode %u is not supported in HWS eswitch mode",
1616 				priv->sh->config.dv_xmeta_en);
1617 				err = ENOTSUP;
1618 				goto error;
1619 		}
1620 		/* Only HWS requires this information. */
1621 		if (sh->refcnt == 1)
1622 			flow_hw_init_tags_set(eth_dev);
1623 		if (priv->sh->config.dv_esw_en &&
1624 		    flow_hw_create_vport_action(eth_dev)) {
1625 			DRV_LOG(ERR, "port %u failed to create vport action",
1626 				eth_dev->data->port_id);
1627 			err = EINVAL;
1628 			goto error;
1629 		}
1630 		/*
1631 		 * If representor matching is disabled, PMD cannot create default flow rules
1632 		 * to receive traffic for all ports, since implicit source port match is not added.
1633 		 * Isolated mode is forced.
1634 		 */
1635 		if (priv->sh->config.dv_esw_en && !priv->sh->config.repr_matching) {
1636 			err = mlx5_flow_isolate(eth_dev, 1, NULL);
1637 			if (err < 0) {
1638 				err = -err;
1639 				goto error;
1640 			}
1641 			DRV_LOG(WARNING, "port %u ingress traffic is restricted to defined "
1642 					 "flow rules (isolated mode) since representor "
1643 					 "matching is disabled",
1644 				eth_dev->data->port_id);
1645 		}
1646 		return eth_dev;
1647 #else
1648 		DRV_LOG(ERR, "DV support is missing for HWS.");
1649 		goto error;
1650 #endif
1651 	}
1652 	if (!priv->sh->flow_priority_check_flag) {
1653 		/* Supported Verbs flow priority number detection. */
1654 		err = mlx5_flow_discover_priorities(eth_dev);
1655 		priv->sh->flow_max_priority = err;
1656 		priv->sh->flow_priority_check_flag = 1;
1657 	} else {
1658 		err = priv->sh->flow_max_priority;
1659 	}
1660 	if (err < 0) {
1661 		err = -err;
1662 		goto error;
1663 	}
1664 	/* Query availability of metadata reg_c's. */
1665 	if (!priv->sh->metadata_regc_check_flag) {
1666 		err = mlx5_flow_discover_mreg_c(eth_dev);
1667 		if (err < 0) {
1668 			err = -err;
1669 			goto error;
1670 		}
1671 	}
1672 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1673 		DRV_LOG(DEBUG,
1674 			"port %u extensive metadata register is not supported",
1675 			eth_dev->data->port_id);
1676 		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1677 			DRV_LOG(ERR, "metadata mode %u is not supported "
1678 				     "(no metadata registers available)",
1679 				     sh->config.dv_xmeta_en);
1680 			err = ENOTSUP;
1681 			goto error;
1682 		}
1683 	}
1684 	if (sh->config.dv_flow_en &&
1685 	    sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1686 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1687 	    priv->sh->dv_regc0_mask) {
1688 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1689 						      MLX5_FLOW_MREG_HTABLE_SZ,
1690 						      false, true, eth_dev,
1691 						      flow_dv_mreg_create_cb,
1692 						      flow_dv_mreg_match_cb,
1693 						      flow_dv_mreg_remove_cb,
1694 						      flow_dv_mreg_clone_cb,
1695 						    flow_dv_mreg_clone_free_cb);
1696 		if (!priv->mreg_cp_tbl) {
1697 			err = ENOMEM;
1698 			goto error;
1699 		}
1700 	}
1701 	rte_spinlock_init(&priv->shared_act_sl);
1702 	mlx5_flow_counter_mode_config(eth_dev);
1703 	mlx5_flow_drop_action_config(eth_dev);
1704 	if (sh->config.dv_flow_en)
1705 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1706 	return eth_dev;
1707 error:
1708 	if (priv) {
1709 		priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1710 							       RTE_MAX_ETHPORTS;
1711 		rte_io_wmb();
1712 #ifdef HAVE_MLX5_HWS_SUPPORT
1713 		if (eth_dev &&
1714 		    priv->sh &&
1715 		    priv->sh->config.dv_flow_en == 2 &&
1716 		    priv->sh->config.dv_esw_en)
1717 			flow_hw_destroy_vport_action(eth_dev);
1718 #endif
1719 		if (priv->mreg_cp_tbl)
1720 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1721 		if (priv->sh)
1722 			mlx5_os_free_shared_dr(priv);
1723 		if (priv->nl_socket_route >= 0)
1724 			close(priv->nl_socket_route);
1725 		if (priv->vmwa_context)
1726 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1727 		if (eth_dev && priv->drop_queue.hrxq)
1728 			mlx5_drop_action_destroy(eth_dev);
1729 		if (priv->mtr_profile_tbl)
1730 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1731 		if (own_domain_id)
1732 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1733 		if (priv->hrxqs)
1734 			mlx5_list_destroy(priv->hrxqs);
1735 		if (eth_dev && priv->flex_item_map)
1736 			mlx5_flex_item_port_cleanup(eth_dev);
1737 		mlx5_free(priv->ext_rxqs);
1738 		mlx5_free(priv);
1739 		if (eth_dev != NULL)
1740 			eth_dev->data->dev_private = NULL;
1741 	}
1742 	if (eth_dev != NULL) {
1743 		/* mac_addrs must not be freed alone because part of
1744 		 * dev_private
1745 		 **/
1746 		eth_dev->data->mac_addrs = NULL;
1747 		rte_eth_dev_release_port(eth_dev);
1748 	}
1749 	if (sh)
1750 		mlx5_free_shared_dev_ctx(sh);
1751 	if (nl_rdma >= 0)
1752 		close(nl_rdma);
1753 	MLX5_ASSERT(err > 0);
1754 	rte_errno = err;
1755 	return NULL;
1756 }
1757 
1758 /**
1759  * Comparison callback to sort device data.
1760  *
1761  * This is meant to be used with qsort().
1762  *
1763  * @param a[in]
1764  *   Pointer to pointer to first data object.
1765  * @param b[in]
1766  *   Pointer to pointer to second data object.
1767  *
1768  * @return
1769  *   0 if both objects are equal, less than 0 if the first argument is less
1770  *   than the second, greater than 0 otherwise.
1771  */
1772 static int
1773 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1774 {
1775 	const struct mlx5_switch_info *si_a =
1776 		&((const struct mlx5_dev_spawn_data *)a)->info;
1777 	const struct mlx5_switch_info *si_b =
1778 		&((const struct mlx5_dev_spawn_data *)b)->info;
1779 	int ret;
1780 
1781 	/* Master device first. */
1782 	ret = si_b->master - si_a->master;
1783 	if (ret)
1784 		return ret;
1785 	/* Then representor devices. */
1786 	ret = si_b->representor - si_a->representor;
1787 	if (ret)
1788 		return ret;
1789 	/* Unidentified devices come last in no specific order. */
1790 	if (!si_a->representor)
1791 		return 0;
1792 	/* Order representors by name. */
1793 	return si_a->port_name - si_b->port_name;
1794 }
1795 
1796 /**
1797  * Match PCI information for possible slaves of bonding device.
1798  *
1799  * @param[in] ibdev_name
1800  *   Name of Infiniband device.
1801  * @param[in] pci_dev
1802  *   Pointer to primary PCI address structure to match.
1803  * @param[in] nl_rdma
1804  *   Netlink RDMA group socket handle.
1805  * @param[in] owner
1806  *   Representor owner PF index.
1807  * @param[out] bond_info
1808  *   Pointer to bonding information.
1809  *
1810  * @return
1811  *   negative value if no bonding device found, otherwise
1812  *   positive index of slave PF in bonding.
1813  */
1814 static int
1815 mlx5_device_bond_pci_match(const char *ibdev_name,
1816 			   const struct rte_pci_addr *pci_dev,
1817 			   int nl_rdma, uint16_t owner,
1818 			   struct mlx5_bond_info *bond_info)
1819 {
1820 	char ifname[IF_NAMESIZE + 1];
1821 	unsigned int ifindex;
1822 	unsigned int np, i;
1823 	FILE *bond_file = NULL, *file;
1824 	int pf = -1;
1825 	int ret;
1826 	uint8_t cur_guid[32] = {0};
1827 	uint8_t guid[32] = {0};
1828 
1829 	/*
1830 	 * Try to get master device name. If something goes wrong suppose
1831 	 * the lack of kernel support and no bonding devices.
1832 	 */
1833 	memset(bond_info, 0, sizeof(*bond_info));
1834 	if (nl_rdma < 0)
1835 		return -1;
1836 	if (!strstr(ibdev_name, "bond"))
1837 		return -1;
1838 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1839 	if (!np)
1840 		return -1;
1841 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1842 		return -1;
1843 	/*
1844 	 * The master device might not be on the predefined port(not on port
1845 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1846 	 * device ports and find master.
1847 	 */
1848 	for (i = 1; i <= np; ++i) {
1849 		/* Check whether Infiniband port is populated. */
1850 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1851 		if (!ifindex)
1852 			continue;
1853 		if (!if_indextoname(ifindex, ifname))
1854 			continue;
1855 		/* Try to read bonding slave names from sysfs. */
1856 		MKSTR(slaves,
1857 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1858 		bond_file = fopen(slaves, "r");
1859 		if (bond_file)
1860 			break;
1861 	}
1862 	if (!bond_file)
1863 		return -1;
1864 	/* Use safe format to check maximal buffer length. */
1865 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1866 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1867 		char tmp_str[IF_NAMESIZE + 32];
1868 		struct rte_pci_addr pci_addr;
1869 		struct mlx5_switch_info	info;
1870 		int ret;
1871 
1872 		/* Process slave interface names in the loop. */
1873 		snprintf(tmp_str, sizeof(tmp_str),
1874 			 "/sys/class/net/%s", ifname);
1875 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1876 			DRV_LOG(WARNING,
1877 				"Cannot get PCI address for netdev \"%s\".",
1878 				ifname);
1879 			continue;
1880 		}
1881 		/* Slave interface PCI address match found. */
1882 		snprintf(tmp_str, sizeof(tmp_str),
1883 			 "/sys/class/net/%s/phys_port_name", ifname);
1884 		file = fopen(tmp_str, "rb");
1885 		if (!file)
1886 			break;
1887 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1888 		if (fscanf(file, "%32s", tmp_str) == 1)
1889 			mlx5_translate_port_name(tmp_str, &info);
1890 		fclose(file);
1891 		/* Only process PF ports. */
1892 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1893 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1894 			continue;
1895 		/* Check max bonding member. */
1896 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1897 			DRV_LOG(WARNING, "bonding index out of range, "
1898 				"please increase MLX5_BOND_MAX_PORTS: %s",
1899 				tmp_str);
1900 			break;
1901 		}
1902 		/* Get ifindex. */
1903 		snprintf(tmp_str, sizeof(tmp_str),
1904 			 "/sys/class/net/%s/ifindex", ifname);
1905 		file = fopen(tmp_str, "rb");
1906 		if (!file)
1907 			break;
1908 		ret = fscanf(file, "%u", &ifindex);
1909 		fclose(file);
1910 		if (ret != 1)
1911 			break;
1912 		/* Save bonding info. */
1913 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1914 			sizeof(bond_info->ports[0].ifname));
1915 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1916 		bond_info->ports[info.port_name].ifindex = ifindex;
1917 		bond_info->n_port++;
1918 		/*
1919 		 * Under socket direct mode, bonding will use
1920 		 * system_image_guid as identification.
1921 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1922 		 * All bonding members should have the same guid even if driver
1923 		 * is using PCIe BDF.
1924 		 */
1925 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1926 		if (ret < 0)
1927 			break;
1928 		else if (ret > 0) {
1929 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1930 			    owner == info.port_name &&
1931 			    (owner != 0 || (owner == 0 &&
1932 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1933 				pf = info.port_name;
1934 		} else if (pci_dev->domain == pci_addr.domain &&
1935 		    pci_dev->bus == pci_addr.bus &&
1936 		    pci_dev->devid == pci_addr.devid &&
1937 		    ((pci_dev->function == 0 &&
1938 		      pci_dev->function + owner == pci_addr.function) ||
1939 		     (pci_dev->function == owner &&
1940 		      pci_addr.function == owner)))
1941 			pf = info.port_name;
1942 	}
1943 	if (pf >= 0) {
1944 		/* Get bond interface info */
1945 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1946 					   bond_info->ifname);
1947 		if (ret)
1948 			DRV_LOG(ERR, "unable to get bond info: %s",
1949 				strerror(rte_errno));
1950 		else
1951 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1952 				ifindex, bond_info->ifindex, bond_info->ifname);
1953 	}
1954 	if (owner == 0 && pf != 0) {
1955 		DRV_LOG(INFO, "PCIe instance " PCI_PRI_FMT " isn't bonding owner",
1956 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1957 				pci_dev->function);
1958 	}
1959 	return pf;
1960 }
1961 
1962 /**
1963  * Register a PCI device within bonding.
1964  *
1965  * This function spawns Ethernet devices out of a given PCI device and
1966  * bonding owner PF index.
1967  *
1968  * @param[in] cdev
1969  *   Pointer to common mlx5 device structure.
1970  * @param[in] req_eth_da
1971  *   Requested ethdev device argument.
1972  * @param[in] owner_id
1973  *   Requested owner PF port ID within bonding device, default to 0.
1974  * @param[in, out] mkvlist
1975  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1976  *
1977  * @return
1978  *   0 on success, a negative errno value otherwise and rte_errno is set.
1979  */
1980 static int
1981 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
1982 		     struct rte_eth_devargs *req_eth_da,
1983 		     uint16_t owner_id, struct mlx5_kvargs_ctrl *mkvlist)
1984 {
1985 	struct ibv_device **ibv_list;
1986 	/*
1987 	 * Number of found IB Devices matching with requested PCI BDF.
1988 	 * nd != 1 means there are multiple IB devices over the same
1989 	 * PCI device and we have representors and master.
1990 	 */
1991 	unsigned int nd = 0;
1992 	/*
1993 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1994 	 * we have the single multiport IB device, and there may be
1995 	 * representors attached to some of found ports.
1996 	 */
1997 	unsigned int np = 0;
1998 	/*
1999 	 * Number of DPDK ethernet devices to Spawn - either over
2000 	 * multiple IB devices or multiple ports of single IB device.
2001 	 * Actually this is the number of iterations to spawn.
2002 	 */
2003 	unsigned int ns = 0;
2004 	/*
2005 	 * Bonding device
2006 	 *   < 0 - no bonding device (single one)
2007 	 *  >= 0 - bonding device (value is slave PF index)
2008 	 */
2009 	int bd = -1;
2010 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2011 	struct mlx5_dev_spawn_data *list = NULL;
2012 	struct rte_eth_devargs eth_da = *req_eth_da;
2013 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2014 	struct mlx5_bond_info bond_info;
2015 	int ret = -1;
2016 
2017 	errno = 0;
2018 	ibv_list = mlx5_glue->get_device_list(&ret);
2019 	if (!ibv_list) {
2020 		rte_errno = errno ? errno : ENOSYS;
2021 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2022 		return -rte_errno;
2023 	}
2024 	/*
2025 	 * First scan the list of all Infiniband devices to find
2026 	 * matching ones, gathering into the list.
2027 	 */
2028 	struct ibv_device *ibv_match[ret + 1];
2029 	int nl_route = mlx5_nl_init(NETLINK_ROUTE, 0);
2030 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
2031 	unsigned int i;
2032 
2033 	while (ret-- > 0) {
2034 		struct rte_pci_addr pci_addr;
2035 
2036 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2037 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2038 						nl_rdma, owner_id, &bond_info);
2039 		if (bd >= 0) {
2040 			/*
2041 			 * Bonding device detected. Only one match is allowed,
2042 			 * the bonding is supported over multi-port IB device,
2043 			 * there should be no matches on representor PCI
2044 			 * functions or non VF LAG bonding devices with
2045 			 * specified address.
2046 			 */
2047 			if (nd) {
2048 				DRV_LOG(ERR,
2049 					"multiple PCI match on bonding device"
2050 					"\"%s\" found", ibv_list[ret]->name);
2051 				rte_errno = ENOENT;
2052 				ret = -rte_errno;
2053 				goto exit;
2054 			}
2055 			/* Amend owner pci address if owner PF ID specified. */
2056 			if (eth_da.nb_representor_ports)
2057 				owner_pci.function += owner_id;
2058 			DRV_LOG(INFO,
2059 				"PCI information matches for slave %d bonding device \"%s\"",
2060 				bd, ibv_list[ret]->name);
2061 			ibv_match[nd++] = ibv_list[ret];
2062 			break;
2063 		} else {
2064 			/* Bonding device not found. */
2065 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2066 					      &pci_addr))
2067 				continue;
2068 			if (rte_pci_addr_cmp(&owner_pci, &pci_addr) != 0)
2069 				continue;
2070 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2071 				ibv_list[ret]->name);
2072 			ibv_match[nd++] = ibv_list[ret];
2073 		}
2074 	}
2075 	ibv_match[nd] = NULL;
2076 	if (!nd) {
2077 		/* No device matches, just complain and bail out. */
2078 		DRV_LOG(WARNING,
2079 			"PF %u doesn't have Verbs device matches PCI device " PCI_PRI_FMT ","
2080 			" are kernel drivers loaded?",
2081 			owner_id, owner_pci.domain, owner_pci.bus,
2082 			owner_pci.devid, owner_pci.function);
2083 		rte_errno = ENOENT;
2084 		ret = -rte_errno;
2085 		goto exit;
2086 	}
2087 	if (nd == 1) {
2088 		/*
2089 		 * Found single matching device may have multiple ports.
2090 		 * Each port may be representor, we have to check the port
2091 		 * number and check the representors existence.
2092 		 */
2093 		if (nl_rdma >= 0)
2094 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2095 		if (!np)
2096 			DRV_LOG(WARNING,
2097 				"Cannot get IB device \"%s\" ports number.",
2098 				ibv_match[0]->name);
2099 		if (bd >= 0 && !np) {
2100 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2101 			rte_errno = ENOENT;
2102 			ret = -rte_errno;
2103 			goto exit;
2104 		}
2105 	}
2106 	/* Now we can determine the maximal amount of devices to be spawned. */
2107 	list = mlx5_malloc(MLX5_MEM_ZERO,
2108 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2109 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2110 	if (!list) {
2111 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2112 		rte_errno = ENOMEM;
2113 		ret = -rte_errno;
2114 		goto exit;
2115 	}
2116 	if (bd >= 0 || np > 1) {
2117 		/*
2118 		 * Single IB device with multiple ports found,
2119 		 * it may be E-Switch master device and representors.
2120 		 * We have to perform identification through the ports.
2121 		 */
2122 		MLX5_ASSERT(nl_rdma >= 0);
2123 		MLX5_ASSERT(ns == 0);
2124 		MLX5_ASSERT(nd == 1);
2125 		MLX5_ASSERT(np);
2126 		for (i = 1; i <= np; ++i) {
2127 			list[ns].bond_info = &bond_info;
2128 			list[ns].max_port = np;
2129 			list[ns].phys_port = i;
2130 			list[ns].phys_dev_name = ibv_match[0]->name;
2131 			list[ns].eth_dev = NULL;
2132 			list[ns].pci_dev = pci_dev;
2133 			list[ns].cdev = cdev;
2134 			list[ns].pf_bond = bd;
2135 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2136 							   ibv_match[0]->name,
2137 							   i);
2138 			if (!list[ns].ifindex) {
2139 				/*
2140 				 * No network interface index found for the
2141 				 * specified port, it means there is no
2142 				 * representor on this port. It's OK,
2143 				 * there can be disabled ports, for example
2144 				 * if sriov_numvfs < sriov_totalvfs.
2145 				 */
2146 				continue;
2147 			}
2148 			ret = -1;
2149 			if (nl_route >= 0)
2150 				ret = mlx5_nl_switch_info(nl_route,
2151 							  list[ns].ifindex,
2152 							  &list[ns].info);
2153 			if (ret || (!list[ns].info.representor &&
2154 				    !list[ns].info.master)) {
2155 				/*
2156 				 * We failed to recognize representors with
2157 				 * Netlink, let's try to perform the task
2158 				 * with sysfs.
2159 				 */
2160 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2161 							     &list[ns].info);
2162 			}
2163 			if (!ret && bd >= 0) {
2164 				switch (list[ns].info.name_type) {
2165 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2166 					if (np == 1) {
2167 						/*
2168 						 * Force standalone bonding
2169 						 * device for ROCE LAG
2170 						 * configurations.
2171 						 */
2172 						list[ns].info.master = 0;
2173 						list[ns].info.representor = 0;
2174 					}
2175 					if (list[ns].info.port_name == bd)
2176 						ns++;
2177 					break;
2178 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2179 					/* Fallthrough */
2180 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2181 					/* Fallthrough */
2182 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2183 					if (list[ns].info.pf_num == bd)
2184 						ns++;
2185 					break;
2186 				default:
2187 					break;
2188 				}
2189 				continue;
2190 			}
2191 			if (!ret && (list[ns].info.representor ^
2192 				     list[ns].info.master))
2193 				ns++;
2194 		}
2195 		if (!ns) {
2196 			DRV_LOG(ERR,
2197 				"Unable to recognize master/representors on the IB device with multiple ports.");
2198 			rte_errno = ENOENT;
2199 			ret = -rte_errno;
2200 			goto exit;
2201 		}
2202 	} else {
2203 		/*
2204 		 * The existence of several matching entries (nd > 1) means
2205 		 * port representors have been instantiated. No existing Verbs
2206 		 * call nor sysfs entries can tell them apart, this can only
2207 		 * be done through Netlink calls assuming kernel drivers are
2208 		 * recent enough to support them.
2209 		 *
2210 		 * In the event of identification failure through Netlink,
2211 		 * try again through sysfs, then:
2212 		 *
2213 		 * 1. A single IB device matches (nd == 1) with single
2214 		 *    port (np=0/1) and is not a representor, assume
2215 		 *    no switch support.
2216 		 *
2217 		 * 2. Otherwise no safe assumptions can be made;
2218 		 *    complain louder and bail out.
2219 		 */
2220 		for (i = 0; i != nd; ++i) {
2221 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2222 			list[ns].bond_info = NULL;
2223 			list[ns].max_port = 1;
2224 			list[ns].phys_port = 1;
2225 			list[ns].phys_dev_name = ibv_match[i]->name;
2226 			list[ns].eth_dev = NULL;
2227 			list[ns].pci_dev = pci_dev;
2228 			list[ns].cdev = cdev;
2229 			list[ns].pf_bond = -1;
2230 			list[ns].ifindex = 0;
2231 			if (nl_rdma >= 0)
2232 				list[ns].ifindex = mlx5_nl_ifindex
2233 							    (nl_rdma,
2234 							     ibv_match[i]->name,
2235 							     1);
2236 			if (!list[ns].ifindex) {
2237 				char ifname[IF_NAMESIZE];
2238 
2239 				/*
2240 				 * Netlink failed, it may happen with old
2241 				 * ib_core kernel driver (before 4.16).
2242 				 * We can assume there is old driver because
2243 				 * here we are processing single ports IB
2244 				 * devices. Let's try sysfs to retrieve
2245 				 * the ifindex. The method works for
2246 				 * master device only.
2247 				 */
2248 				if (nd > 1) {
2249 					/*
2250 					 * Multiple devices found, assume
2251 					 * representors, can not distinguish
2252 					 * master/representor and retrieve
2253 					 * ifindex via sysfs.
2254 					 */
2255 					continue;
2256 				}
2257 				ret = mlx5_get_ifname_sysfs
2258 					(ibv_match[i]->ibdev_path, ifname);
2259 				if (!ret)
2260 					list[ns].ifindex =
2261 						if_nametoindex(ifname);
2262 				if (!list[ns].ifindex) {
2263 					/*
2264 					 * No network interface index found
2265 					 * for the specified device, it means
2266 					 * there it is neither representor
2267 					 * nor master.
2268 					 */
2269 					continue;
2270 				}
2271 			}
2272 			ret = -1;
2273 			if (nl_route >= 0)
2274 				ret = mlx5_nl_switch_info(nl_route,
2275 							  list[ns].ifindex,
2276 							  &list[ns].info);
2277 			if (ret || (!list[ns].info.representor &&
2278 				    !list[ns].info.master)) {
2279 				/*
2280 				 * We failed to recognize representors with
2281 				 * Netlink, let's try to perform the task
2282 				 * with sysfs.
2283 				 */
2284 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2285 							     &list[ns].info);
2286 			}
2287 			if (!ret && (list[ns].info.representor ^
2288 				     list[ns].info.master)) {
2289 				ns++;
2290 			} else if ((nd == 1) &&
2291 				   !list[ns].info.representor &&
2292 				   !list[ns].info.master) {
2293 				/*
2294 				 * Single IB device with one physical port and
2295 				 * attached network device.
2296 				 * May be SRIOV is not enabled or there is no
2297 				 * representors.
2298 				 */
2299 				DRV_LOG(INFO, "No E-Switch support detected.");
2300 				ns++;
2301 				break;
2302 			}
2303 		}
2304 		if (!ns) {
2305 			DRV_LOG(ERR,
2306 				"Unable to recognize master/representors on the multiple IB devices.");
2307 			rte_errno = ENOENT;
2308 			ret = -rte_errno;
2309 			goto exit;
2310 		}
2311 		/*
2312 		 * New kernels may add the switch_id attribute for the case
2313 		 * there is no E-Switch and we wrongly recognized the only
2314 		 * device as master. Override this if there is the single
2315 		 * device with single port and new device name format present.
2316 		 */
2317 		if (nd == 1 &&
2318 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2319 			list[0].info.master = 0;
2320 			list[0].info.representor = 0;
2321 		}
2322 	}
2323 	MLX5_ASSERT(ns);
2324 	/*
2325 	 * Sort list to probe devices in natural order for users convenience
2326 	 * (i.e. master first, then representors from lowest to highest ID).
2327 	 */
2328 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2329 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2330 		/* Set devargs default values. */
2331 		if (eth_da.nb_mh_controllers == 0) {
2332 			eth_da.nb_mh_controllers = 1;
2333 			eth_da.mh_controllers[0] = 0;
2334 		}
2335 		if (eth_da.nb_ports == 0 && ns > 0) {
2336 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2337 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2338 					pci_dev->device.devargs->args);
2339 			eth_da.nb_ports = 1;
2340 			eth_da.ports[0] = list[0].info.pf_num;
2341 		}
2342 		if (eth_da.nb_representor_ports == 0) {
2343 			eth_da.nb_representor_ports = 1;
2344 			eth_da.representor_ports[0] = 0;
2345 		}
2346 	}
2347 	for (i = 0; i != ns; ++i) {
2348 		uint32_t restore;
2349 
2350 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i], &eth_da,
2351 						 mkvlist);
2352 		if (!list[i].eth_dev) {
2353 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2354 				break;
2355 			/* Device is disabled or already spawned. Ignore it. */
2356 			continue;
2357 		}
2358 		restore = list[i].eth_dev->data->dev_flags;
2359 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2360 		/**
2361 		 * Each representor has a dedicated interrupts vector.
2362 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2363 		 * representor eth_dev object because representor and PF
2364 		 * share the same PCI address.
2365 		 * Override representor device with a dedicated
2366 		 * interrupts handle here.
2367 		 * Representor interrupts handle is released in mlx5_dev_stop().
2368 		 */
2369 		if (list[i].info.representor) {
2370 			struct rte_intr_handle *intr_handle =
2371 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2372 			if (intr_handle == NULL) {
2373 				DRV_LOG(ERR,
2374 					"port %u failed to allocate memory for interrupt handler "
2375 					"Rx interrupts will not be supported",
2376 					i);
2377 				rte_errno = ENOMEM;
2378 				ret = -rte_errno;
2379 				goto exit;
2380 			}
2381 			list[i].eth_dev->intr_handle = intr_handle;
2382 		}
2383 		/* Restore non-PCI flags cleared by the above call. */
2384 		list[i].eth_dev->data->dev_flags |= restore;
2385 		rte_eth_dev_probing_finish(list[i].eth_dev);
2386 	}
2387 	if (i != ns) {
2388 		DRV_LOG(ERR,
2389 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2390 			" encountering an error: %s",
2391 			owner_pci.domain, owner_pci.bus,
2392 			owner_pci.devid, owner_pci.function,
2393 			strerror(rte_errno));
2394 		ret = -rte_errno;
2395 		/* Roll back. */
2396 		while (i--) {
2397 			if (!list[i].eth_dev)
2398 				continue;
2399 			mlx5_dev_close(list[i].eth_dev);
2400 			/* mac_addrs must not be freed because in dev_private */
2401 			list[i].eth_dev->data->mac_addrs = NULL;
2402 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2403 		}
2404 		/* Restore original error. */
2405 		rte_errno = -ret;
2406 	} else {
2407 		ret = 0;
2408 	}
2409 exit:
2410 	/*
2411 	 * Do the routine cleanup:
2412 	 * - close opened Netlink sockets
2413 	 * - free allocated spawn data array
2414 	 * - free the Infiniband device list
2415 	 */
2416 	if (nl_rdma >= 0)
2417 		close(nl_rdma);
2418 	if (nl_route >= 0)
2419 		close(nl_route);
2420 	if (list)
2421 		mlx5_free(list);
2422 	MLX5_ASSERT(ibv_list);
2423 	mlx5_glue->free_device_list(ibv_list);
2424 	return ret;
2425 }
2426 
2427 static int
2428 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2429 			  struct rte_eth_devargs *eth_da)
2430 {
2431 	int ret = 0;
2432 
2433 	if (dev->devargs == NULL)
2434 		return 0;
2435 	memset(eth_da, 0, sizeof(*eth_da));
2436 	/* Parse representor information first from class argument. */
2437 	if (dev->devargs->cls_str)
2438 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2439 	if (ret != 0) {
2440 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2441 			dev->devargs->cls_str);
2442 		return -rte_errno;
2443 	}
2444 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE && dev->devargs->args) {
2445 		/* Parse legacy device argument */
2446 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2447 		if (ret) {
2448 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2449 				dev->devargs->args);
2450 			return -rte_errno;
2451 		}
2452 	}
2453 	return 0;
2454 }
2455 
2456 /**
2457  * Callback to register a PCI device.
2458  *
2459  * This function spawns Ethernet devices out of a given PCI device.
2460  *
2461  * @param[in] cdev
2462  *   Pointer to common mlx5 device structure.
2463  * @param[in, out] mkvlist
2464  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2465  *
2466  * @return
2467  *   0 on success, a negative errno value otherwise and rte_errno is set.
2468  */
2469 static int
2470 mlx5_os_pci_probe(struct mlx5_common_device *cdev,
2471 		  struct mlx5_kvargs_ctrl *mkvlist)
2472 {
2473 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2474 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2475 	int ret = 0;
2476 	uint16_t p;
2477 
2478 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2479 	if (ret != 0)
2480 		return ret;
2481 
2482 	if (eth_da.nb_ports > 0) {
2483 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2484 		for (p = 0; p < eth_da.nb_ports; p++) {
2485 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2486 						   eth_da.ports[p], mkvlist);
2487 			if (ret) {
2488 				DRV_LOG(INFO, "Probe of PCI device " PCI_PRI_FMT " "
2489 					"aborted due to proding failure of PF %u",
2490 					pci_dev->addr.domain, pci_dev->addr.bus,
2491 					pci_dev->addr.devid, pci_dev->addr.function,
2492 					eth_da.ports[p]);
2493 				mlx5_net_remove(cdev);
2494 				if (p != 0)
2495 					break;
2496 			}
2497 		}
2498 	} else {
2499 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0, mkvlist);
2500 	}
2501 	return ret;
2502 }
2503 
2504 /* Probe a single SF device on auxiliary bus, no representor support. */
2505 static int
2506 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev,
2507 			struct mlx5_kvargs_ctrl *mkvlist)
2508 {
2509 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2510 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2511 	struct rte_device *dev = cdev->dev;
2512 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2513 	struct rte_eth_dev *eth_dev;
2514 	int ret = 0;
2515 
2516 	/* Parse ethdev devargs. */
2517 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2518 	if (ret != 0)
2519 		return ret;
2520 	/* Init spawn data. */
2521 	spawn.max_port = 1;
2522 	spawn.phys_port = 1;
2523 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2524 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2525 	if (ret < 0) {
2526 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2527 		return ret;
2528 	}
2529 	spawn.ifindex = ret;
2530 	spawn.cdev = cdev;
2531 	/* Spawn device. */
2532 	eth_dev = mlx5_dev_spawn(dev, &spawn, &eth_da, mkvlist);
2533 	if (eth_dev == NULL)
2534 		return -rte_errno;
2535 	/* Post create. */
2536 	eth_dev->intr_handle = adev->intr_handle;
2537 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2538 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2539 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2540 		eth_dev->data->numa_node = dev->numa_node;
2541 	}
2542 	rte_eth_dev_probing_finish(eth_dev);
2543 	return 0;
2544 }
2545 
2546 /**
2547  * Net class driver callback to probe a device.
2548  *
2549  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2550  *
2551  * @param[in] cdev
2552  *   Pointer to the common mlx5 device.
2553  * @param[in, out] mkvlist
2554  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2555  *
2556  * @return
2557  *   0 on success, a negative errno value otherwise and rte_errno is set.
2558  */
2559 int
2560 mlx5_os_net_probe(struct mlx5_common_device *cdev,
2561 		  struct mlx5_kvargs_ctrl *mkvlist)
2562 {
2563 	int ret;
2564 
2565 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2566 		mlx5_pmd_socket_init();
2567 	ret = mlx5_init_once();
2568 	if (ret) {
2569 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2570 			strerror(rte_errno));
2571 		return -rte_errno;
2572 	}
2573 	ret = mlx5_probe_again_args_validate(cdev, mkvlist);
2574 	if (ret) {
2575 		DRV_LOG(ERR, "Probe again parameters are not compatible : %s",
2576 			strerror(rte_errno));
2577 		return -rte_errno;
2578 	}
2579 	if (mlx5_dev_is_pci(cdev->dev))
2580 		return mlx5_os_pci_probe(cdev, mkvlist);
2581 	else
2582 		return mlx5_os_auxiliary_probe(cdev, mkvlist);
2583 }
2584 
2585 /**
2586  * Cleanup resources when the last device is closed.
2587  */
2588 void
2589 mlx5_os_net_cleanup(void)
2590 {
2591 	mlx5_pmd_socket_uninit();
2592 }
2593 
2594 /**
2595  * Install shared asynchronous device events handler.
2596  * This function is implemented to support event sharing
2597  * between multiple ports of single IB device.
2598  *
2599  * @param sh
2600  *   Pointer to mlx5_dev_ctx_shared object.
2601  */
2602 void
2603 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2604 {
2605 	struct ibv_context *ctx = sh->cdev->ctx;
2606 	int nlsk_fd;
2607 
2608 	sh->intr_handle = mlx5_os_interrupt_handler_create
2609 		(RTE_INTR_INSTANCE_F_SHARED, true,
2610 		 ctx->async_fd, mlx5_dev_interrupt_handler, sh);
2611 	if (!sh->intr_handle) {
2612 		DRV_LOG(ERR, "Failed to allocate intr_handle.");
2613 		return;
2614 	}
2615 	nlsk_fd = mlx5_nl_init(NETLINK_ROUTE, RTMGRP_LINK);
2616 	if (nlsk_fd < 0) {
2617 		DRV_LOG(ERR, "Failed to create a socket for Netlink events: %s",
2618 			rte_strerror(rte_errno));
2619 		return;
2620 	}
2621 	sh->intr_handle_nl = mlx5_os_interrupt_handler_create
2622 		(RTE_INTR_INSTANCE_F_SHARED, true,
2623 		 nlsk_fd, mlx5_dev_interrupt_handler_nl, sh);
2624 	if (sh->intr_handle_nl == NULL) {
2625 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2626 		return;
2627 	}
2628 	if (sh->cdev->config.devx) {
2629 #ifdef HAVE_IBV_DEVX_ASYNC
2630 		struct mlx5dv_devx_cmd_comp *devx_comp;
2631 
2632 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2633 		devx_comp = sh->devx_comp;
2634 		if (!devx_comp) {
2635 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2636 			return;
2637 		}
2638 		sh->intr_handle_devx = mlx5_os_interrupt_handler_create
2639 			(RTE_INTR_INSTANCE_F_SHARED, true,
2640 			 devx_comp->fd,
2641 			 mlx5_dev_interrupt_handler_devx, sh);
2642 		if (!sh->intr_handle_devx) {
2643 			DRV_LOG(ERR, "Failed to allocate intr_handle.");
2644 			return;
2645 		}
2646 #endif /* HAVE_IBV_DEVX_ASYNC */
2647 	}
2648 }
2649 
2650 /**
2651  * Uninstall shared asynchronous device events handler.
2652  * This function is implemented to support event sharing
2653  * between multiple ports of single IB device.
2654  *
2655  * @param dev
2656  *   Pointer to mlx5_dev_ctx_shared object.
2657  */
2658 void
2659 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2660 {
2661 	mlx5_os_interrupt_handler_destroy(sh->intr_handle,
2662 					  mlx5_dev_interrupt_handler, sh);
2663 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_nl,
2664 					  mlx5_dev_interrupt_handler_nl, sh);
2665 #ifdef HAVE_IBV_DEVX_ASYNC
2666 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_devx,
2667 					  mlx5_dev_interrupt_handler_devx, sh);
2668 	if (sh->devx_comp)
2669 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2670 #endif
2671 }
2672 
2673 /**
2674  * Read statistics by a named counter.
2675  *
2676  * @param[in] priv
2677  *   Pointer to the private device data structure.
2678  * @param[in] ctr_name
2679  *   Pointer to the name of the statistic counter to read
2680  * @param[out] stat
2681  *   Pointer to read statistic value.
2682  * @return
2683  *   0 on success and stat is valud, 1 if failed to read the value
2684  *   rte_errno is set.
2685  *
2686  */
2687 int
2688 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2689 		      uint64_t *stat)
2690 {
2691 	int fd;
2692 
2693 	if (priv->sh) {
2694 		if (priv->q_counters != NULL &&
2695 		    strcmp(ctr_name, "out_of_buffer") == 0)
2696 			return mlx5_devx_cmd_queue_counter_query
2697 					(priv->q_counters, 0, (uint32_t *)stat);
2698 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2699 		      priv->sh->ibdev_path,
2700 		      priv->dev_port,
2701 		      ctr_name);
2702 		fd = open(path, O_RDONLY);
2703 		/*
2704 		 * in switchdev the file location is not per port
2705 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2706 		 */
2707 		if (fd == -1) {
2708 			MKSTR(path1, "%s/hw_counters/%s",
2709 			      priv->sh->ibdev_path,
2710 			      ctr_name);
2711 			fd = open(path1, O_RDONLY);
2712 		}
2713 		if (fd != -1) {
2714 			char buf[21] = {'\0'};
2715 			ssize_t n = read(fd, buf, sizeof(buf));
2716 
2717 			close(fd);
2718 			if (n != -1) {
2719 				*stat = strtoull(buf, NULL, 10);
2720 				return 0;
2721 			}
2722 		}
2723 	}
2724 	*stat = 0;
2725 	return 1;
2726 }
2727 
2728 /**
2729  * Remove a MAC address from device
2730  *
2731  * @param dev
2732  *   Pointer to Ethernet device structure.
2733  * @param index
2734  *   MAC address index.
2735  */
2736 void
2737 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2738 {
2739 	struct mlx5_priv *priv = dev->data->dev_private;
2740 	const int vf = priv->sh->dev_cap.vf;
2741 
2742 	if (vf)
2743 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2744 					mlx5_ifindex(dev), priv->mac_own,
2745 					&dev->data->mac_addrs[index], index);
2746 }
2747 
2748 /**
2749  * Adds a MAC address to the device
2750  *
2751  * @param dev
2752  *   Pointer to Ethernet device structure.
2753  * @param mac_addr
2754  *   MAC address to register.
2755  * @param index
2756  *   MAC address index.
2757  *
2758  * @return
2759  *   0 on success, a negative errno value otherwise
2760  */
2761 int
2762 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2763 		     uint32_t index)
2764 {
2765 	struct mlx5_priv *priv = dev->data->dev_private;
2766 	const int vf = priv->sh->dev_cap.vf;
2767 	int ret = 0;
2768 
2769 	if (vf)
2770 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2771 					   mlx5_ifindex(dev), priv->mac_own,
2772 					   mac, index);
2773 	return ret;
2774 }
2775 
2776 /**
2777  * Modify a VF MAC address
2778  *
2779  * @param priv
2780  *   Pointer to device private data.
2781  * @param mac_addr
2782  *   MAC address to modify into.
2783  * @param iface_idx
2784  *   Net device interface index
2785  * @param vf_index
2786  *   VF index
2787  *
2788  * @return
2789  *   0 on success, a negative errno value otherwise
2790  */
2791 int
2792 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2793 			   unsigned int iface_idx,
2794 			   struct rte_ether_addr *mac_addr,
2795 			   int vf_index)
2796 {
2797 	return mlx5_nl_vf_mac_addr_modify
2798 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2799 }
2800 
2801 /**
2802  * Set device promiscuous mode
2803  *
2804  * @param dev
2805  *   Pointer to Ethernet device structure.
2806  * @param enable
2807  *   0 - promiscuous is disabled, otherwise - enabled
2808  *
2809  * @return
2810  *   0 on success, a negative error value otherwise
2811  */
2812 int
2813 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2814 {
2815 	struct mlx5_priv *priv = dev->data->dev_private;
2816 
2817 	return mlx5_nl_promisc(priv->nl_socket_route,
2818 			       mlx5_ifindex(dev), !!enable);
2819 }
2820 
2821 /**
2822  * Set device promiscuous mode
2823  *
2824  * @param dev
2825  *   Pointer to Ethernet device structure.
2826  * @param enable
2827  *   0 - all multicase is disabled, otherwise - enabled
2828  *
2829  * @return
2830  *   0 on success, a negative error value otherwise
2831  */
2832 int
2833 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2834 {
2835 	struct mlx5_priv *priv = dev->data->dev_private;
2836 
2837 	return mlx5_nl_allmulti(priv->nl_socket_route,
2838 				mlx5_ifindex(dev), !!enable);
2839 }
2840 
2841 /**
2842  * Flush device MAC addresses
2843  *
2844  * @param dev
2845  *   Pointer to Ethernet device structure.
2846  *
2847  */
2848 void
2849 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2850 {
2851 	struct mlx5_priv *priv = dev->data->dev_private;
2852 
2853 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2854 			       dev->data->mac_addrs,
2855 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2856 }
2857