xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision de1f01a8eabd1da08d85e77ff99ba85e03cfd1ad)
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 	sh->default_miss_action =
478 			mlx5_glue->dr_create_flow_action_default_miss();
479 	if (!sh->default_miss_action)
480 		DRV_LOG(WARNING, "Default miss action is not supported.");
481 	/* The resources below are only valid with DV support. */
482 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
483 	/* Init shared flex parsers list, no need lcore_share */
484 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
485 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
486 					       mlx5_flex_parser_create_cb,
487 					       mlx5_flex_parser_match_cb,
488 					       mlx5_flex_parser_remove_cb,
489 					       mlx5_flex_parser_clone_cb,
490 					       mlx5_flex_parser_clone_free_cb);
491 	if (!sh->flex_parsers_dv)
492 		goto error;
493 	if (priv->sh->config.dv_flow_en == 2)
494 		return 0;
495 	/* Init port id action list. */
496 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
497 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
498 						   flow_dv_port_id_create_cb,
499 						   flow_dv_port_id_match_cb,
500 						   flow_dv_port_id_remove_cb,
501 						   flow_dv_port_id_clone_cb,
502 						 flow_dv_port_id_clone_free_cb);
503 	if (!sh->port_id_action_list)
504 		goto error;
505 	/* Init push vlan action list. */
506 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
507 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
508 						    flow_dv_push_vlan_create_cb,
509 						    flow_dv_push_vlan_match_cb,
510 						    flow_dv_push_vlan_remove_cb,
511 						    flow_dv_push_vlan_clone_cb,
512 					       flow_dv_push_vlan_clone_free_cb);
513 	if (!sh->push_vlan_action_list)
514 		goto error;
515 	/* Init sample action list. */
516 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
517 	sh->sample_action_list = mlx5_list_create(s, sh, true,
518 						  flow_dv_sample_create_cb,
519 						  flow_dv_sample_match_cb,
520 						  flow_dv_sample_remove_cb,
521 						  flow_dv_sample_clone_cb,
522 						  flow_dv_sample_clone_free_cb);
523 	if (!sh->sample_action_list)
524 		goto error;
525 	/* Init dest array action list. */
526 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
527 	sh->dest_array_list = mlx5_list_create(s, sh, true,
528 					       flow_dv_dest_array_create_cb,
529 					       flow_dv_dest_array_match_cb,
530 					       flow_dv_dest_array_remove_cb,
531 					       flow_dv_dest_array_clone_cb,
532 					      flow_dv_dest_array_clone_free_cb);
533 	if (!sh->dest_array_list)
534 		goto error;
535 #else
536 	if (priv->sh->config.dv_flow_en == 2)
537 		return 0;
538 #endif
539 #ifdef HAVE_MLX5DV_DR
540 	void *domain;
541 
542 	/* Reference counter is zero, we should initialize structures. */
543 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
544 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
545 	if (!domain) {
546 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
547 		err = errno;
548 		goto error;
549 	}
550 	sh->rx_domain = domain;
551 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
552 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
553 	if (!domain) {
554 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
555 		err = errno;
556 		goto error;
557 	}
558 	sh->tx_domain = domain;
559 #ifdef HAVE_MLX5DV_DR_ESWITCH
560 	if (sh->config.dv_esw_en) {
561 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
562 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
563 		if (!domain) {
564 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
565 			err = errno;
566 			goto error;
567 		}
568 		sh->fdb_domain = domain;
569 	}
570 	/*
571 	 * The drop action is just some dummy placeholder in rdma-core. It
572 	 * does not belong to domains and has no any attributes, and, can be
573 	 * shared by the entire device.
574 	 */
575 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
576 	if (!sh->dr_drop_action) {
577 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
578 		err = errno;
579 		goto error;
580 	}
581 #endif
582 	if (!sh->tunnel_hub && sh->config.dv_miss_info)
583 		err = mlx5_alloc_tunnel_hub(sh);
584 	if (err) {
585 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
586 		goto error;
587 	}
588 	if (sh->config.reclaim_mode == MLX5_RCM_AGGR) {
589 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
590 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
591 		if (sh->fdb_domain)
592 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
593 	}
594 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
595 	if (!sh->config.allow_duplicate_pattern) {
596 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
597 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
598 #endif
599 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
600 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
601 		if (sh->fdb_domain)
602 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
603 	}
604 
605 	__mlx5_discovery_misc5_cap(priv);
606 #endif /* HAVE_MLX5DV_DR */
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 	/*
963 	 * Assuming Multiport E-Switch device was detected,
964 	 * if spawned port is an uplink, check if the port
965 	 * was requested through representor devarg.
966 	 */
967 	if (mlx5_is_probed_port_on_mpesw_device(spawn) &&
968 	    switch_info->name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
969 		for (p = 0; p < eth_da->nb_ports; ++p)
970 			if (switch_info->port_name == eth_da->ports[p])
971 				return true;
972 		rte_errno = EBUSY;
973 		return false;
974 	}
975 	switch (eth_da->type) {
976 	case RTE_ETH_REPRESENTOR_PF:
977 		/*
978 		 * PF representors provided in devargs translate to uplink ports, but
979 		 * if and only if the device is a part of MPESW device.
980 		 */
981 		if (!mlx5_is_probed_port_on_mpesw_device(spawn)) {
982 			rte_errno = EBUSY;
983 			return false;
984 		}
985 		break;
986 	case RTE_ETH_REPRESENTOR_SF:
987 		if (!(spawn->info.port_name == -1 &&
988 		      switch_info->name_type ==
989 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
990 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
991 			rte_errno = EBUSY;
992 			return false;
993 		}
994 		break;
995 	case RTE_ETH_REPRESENTOR_VF:
996 		/* Allows HPF representor index -1 as exception. */
997 		if (!(spawn->info.port_name == -1 &&
998 		      switch_info->name_type ==
999 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
1000 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
1001 			rte_errno = EBUSY;
1002 			return false;
1003 		}
1004 		break;
1005 	case RTE_ETH_REPRESENTOR_NONE:
1006 		rte_errno = EBUSY;
1007 		return false;
1008 	default:
1009 		rte_errno = ENOTSUP;
1010 		DRV_LOG(ERR, "unsupported representor type");
1011 		return false;
1012 	}
1013 	/* Check representor ID: */
1014 	for (p = 0; p < eth_da->nb_ports; ++p) {
1015 		if (!mlx5_is_probed_port_on_mpesw_device(spawn) && spawn->pf_bond < 0) {
1016 			/* For non-LAG mode, allow and ignore pf. */
1017 			switch_info->pf_num = eth_da->ports[p];
1018 			repr_id = mlx5_representor_id_encode(switch_info,
1019 							     eth_da->type);
1020 		}
1021 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
1022 			id = MLX5_REPRESENTOR_ID
1023 				(eth_da->ports[p], eth_da->type,
1024 				 eth_da->representor_ports[f]);
1025 			if (repr_id == id)
1026 				return true;
1027 		}
1028 	}
1029 	rte_errno = EBUSY;
1030 	return false;
1031 }
1032 
1033 /**
1034  * Spawn an Ethernet device from Verbs information.
1035  *
1036  * @param dpdk_dev
1037  *   Backing DPDK device.
1038  * @param spawn
1039  *   Verbs device parameters (name, port, switch_info) to spawn.
1040  * @param eth_da
1041  *   Device arguments.
1042  * @param mkvlist
1043  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
1044  *
1045  * @return
1046  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1047  *   is set. The following errors are defined:
1048  *
1049  *   EBUSY: device is not supposed to be spawned.
1050  *   EEXIST: device is already spawned
1051  */
1052 static struct rte_eth_dev *
1053 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1054 	       struct mlx5_dev_spawn_data *spawn,
1055 	       struct rte_eth_devargs *eth_da,
1056 	       struct mlx5_kvargs_ctrl *mkvlist)
1057 {
1058 	const struct mlx5_switch_info *switch_info = &spawn->info;
1059 	struct mlx5_dev_ctx_shared *sh = NULL;
1060 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
1061 	struct rte_eth_dev *eth_dev = NULL;
1062 	struct mlx5_priv *priv = NULL;
1063 	int err = 0;
1064 	struct rte_ether_addr mac;
1065 	char name[RTE_ETH_NAME_MAX_LEN];
1066 	int own_domain_id = 0;
1067 	uint16_t port_id;
1068 	struct mlx5_port_info vport_info = { .query_flags = 0 };
1069 	int nl_rdma;
1070 	int i;
1071 
1072 	/* Determine if this port representor is supposed to be spawned. */
1073 	if (switch_info->representor && dpdk_dev->devargs &&
1074 	    !mlx5_representor_match(spawn, eth_da))
1075 		return NULL;
1076 	/* Build device name. */
1077 	if (spawn->pf_bond >= 0) {
1078 		/* Bonding device. */
1079 		if (!switch_info->representor) {
1080 			err = snprintf(name, sizeof(name), "%s_%s",
1081 				       dpdk_dev->name, spawn->phys_dev_name);
1082 		} else {
1083 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1084 				dpdk_dev->name, spawn->phys_dev_name,
1085 				switch_info->ctrl_num,
1086 				switch_info->pf_num,
1087 				switch_info->name_type ==
1088 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1089 				switch_info->port_name);
1090 		}
1091 	} else if (mlx5_is_probed_port_on_mpesw_device(spawn)) {
1092 		/* MPESW device. */
1093 		if (switch_info->name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
1094 			err = snprintf(name, sizeof(name), "%s_p%d",
1095 				       dpdk_dev->name, spawn->mpesw_port);
1096 		} else {
1097 			err = snprintf(name, sizeof(name), "%s_representor_c%dpf%d%s%u",
1098 				dpdk_dev->name,
1099 				switch_info->ctrl_num,
1100 				switch_info->pf_num,
1101 				switch_info->name_type ==
1102 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1103 				switch_info->port_name);
1104 		}
1105 	} else {
1106 		/* Single device. */
1107 		if (!switch_info->representor)
1108 			strlcpy(name, dpdk_dev->name, sizeof(name));
1109 		else
1110 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
1111 				 dpdk_dev->name,
1112 				 switch_info->name_type ==
1113 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1114 				 switch_info->port_name);
1115 	}
1116 	if (err >= (int)sizeof(name))
1117 		DRV_LOG(WARNING, "device name overflow %s", name);
1118 	/* check if the device is already spawned */
1119 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1120 		/*
1121 		 * When device is already spawned, its devargs should be set
1122 		 * as used. otherwise, mlx5_kvargs_validate() will fail.
1123 		 */
1124 		if (mkvlist)
1125 			mlx5_port_args_set_used(name, port_id, mkvlist);
1126 		rte_errno = EEXIST;
1127 		return NULL;
1128 	}
1129 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1130 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1131 		struct mlx5_mp_id mp_id;
1132 		int fd;
1133 
1134 		eth_dev = rte_eth_dev_attach_secondary(name);
1135 		if (eth_dev == NULL) {
1136 			DRV_LOG(ERR, "can not attach rte ethdev");
1137 			rte_errno = ENOMEM;
1138 			return NULL;
1139 		}
1140 		eth_dev->device = dpdk_dev;
1141 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1142 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1143 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1144 		err = mlx5_proc_priv_init(eth_dev);
1145 		if (err)
1146 			return NULL;
1147 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1148 		/* Receive command fd from primary process */
1149 		fd = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1150 		if (fd < 0)
1151 			goto err_secondary;
1152 		/* Remap UAR for Tx queues. */
1153 		err = mlx5_tx_uar_init_secondary(eth_dev, fd);
1154 		close(fd);
1155 		if (err)
1156 			goto err_secondary;
1157 		/*
1158 		 * Ethdev pointer is still required as input since
1159 		 * the primary device is not accessible from the
1160 		 * secondary process.
1161 		 */
1162 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1163 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1164 		return eth_dev;
1165 err_secondary:
1166 		mlx5_dev_close(eth_dev);
1167 		return NULL;
1168 	}
1169 	sh = mlx5_alloc_shared_dev_ctx(spawn, mkvlist);
1170 	if (!sh)
1171 		return NULL;
1172 	nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
1173 	/* Check port status. */
1174 	if (spawn->phys_port <= UINT8_MAX) {
1175 		/* Legacy Verbs api only support u8 port number. */
1176 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1177 					    &port_attr);
1178 		if (err) {
1179 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1180 			goto error;
1181 		}
1182 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1183 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1184 			err = EINVAL;
1185 			goto error;
1186 		}
1187 	} else if (nl_rdma >= 0) {
1188 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1189 		err = mlx5_nl_port_state(nl_rdma,
1190 			spawn->phys_dev_name,
1191 			spawn->phys_port);
1192 		if (err < 0) {
1193 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1194 				strerror(rte_errno));
1195 			err = -rte_errno;
1196 			goto error;
1197 		}
1198 		port_attr.state = (enum ibv_port_state)err;
1199 	}
1200 	if (port_attr.state != IBV_PORT_ACTIVE)
1201 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1202 			mlx5_glue->port_state_str(port_attr.state),
1203 			port_attr.state);
1204 	/* Allocate private eth device data. */
1205 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1206 			   sizeof(*priv),
1207 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1208 	if (priv == NULL) {
1209 		DRV_LOG(ERR, "priv allocation failure");
1210 		err = ENOMEM;
1211 		goto error;
1212 	}
1213 	/*
1214 	 * When user configures remote PD and CTX and device creates RxQ by
1215 	 * DevX, external RxQ is both supported and requested.
1216 	 */
1217 	if (mlx5_imported_pd_and_ctx(sh->cdev) && mlx5_devx_obj_ops_en(sh)) {
1218 		priv->ext_rxqs = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1219 					     sizeof(struct mlx5_external_rxq) *
1220 					     MLX5_MAX_EXT_RX_QUEUES, 0,
1221 					     SOCKET_ID_ANY);
1222 		if (priv->ext_rxqs == NULL) {
1223 			DRV_LOG(ERR, "Fail to allocate external RxQ array.");
1224 			err = ENOMEM;
1225 			goto error;
1226 		}
1227 		DRV_LOG(DEBUG, "External RxQ is supported.");
1228 	}
1229 	priv->sh = sh;
1230 	priv->dev_port = spawn->phys_port;
1231 	priv->pci_dev = spawn->pci_dev;
1232 	priv->mtu = RTE_ETHER_MTU;
1233 	/* Some internal functions rely on Netlink sockets, open them now. */
1234 	priv->nl_socket_rdma = nl_rdma;
1235 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE, 0);
1236 	priv->representor = !!switch_info->representor;
1237 	priv->master = !!switch_info->master;
1238 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1239 	priv->vport_meta_tag = 0;
1240 	priv->vport_meta_mask = 0;
1241 	priv->pf_bond = spawn->pf_bond;
1242 	priv->mpesw_port = spawn->mpesw_port;
1243 	priv->mpesw_uplink = false;
1244 	priv->mpesw_owner = spawn->info.mpesw_owner;
1245 	if (mlx5_is_port_on_mpesw_device(priv))
1246 		priv->mpesw_uplink = (spawn->info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK);
1247 
1248 	DRV_LOG(DEBUG,
1249 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d "
1250 		"mpesw_port=%d mpesw_uplink=%d",
1251 		priv->dev_port, dpdk_dev->bus->name,
1252 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1253 		priv->master, priv->representor, priv->pf_bond,
1254 		priv->mpesw_port, priv->mpesw_uplink);
1255 
1256 	if (mlx5_is_port_on_mpesw_device(priv) && priv->sh->config.dv_flow_en != 2) {
1257 		DRV_LOG(ERR, "MPESW device is supported only with HWS");
1258 		err = ENOTSUP;
1259 		goto error;
1260 	}
1261 	/*
1262 	 * If we have E-Switch we should determine the vport attributes.
1263 	 * E-Switch may use either source vport field or reg_c[0] metadata
1264 	 * register to match on vport index. The engaged part of metadata
1265 	 * register is defined by mask.
1266 	 */
1267 	if (sh->esw_mode) {
1268 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1269 						 spawn->phys_port,
1270 						 &vport_info);
1271 		if (err) {
1272 			DRV_LOG(WARNING,
1273 				"Cannot query devx port %d on device %s",
1274 				spawn->phys_port, spawn->phys_dev_name);
1275 			vport_info.query_flags = 0;
1276 		}
1277 	}
1278 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1279 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1280 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1281 		if (!priv->vport_meta_mask) {
1282 			DRV_LOG(ERR,
1283 				"vport zero mask for port %d on bonding device %s",
1284 				spawn->phys_port, spawn->phys_dev_name);
1285 			err = ENOTSUP;
1286 			goto error;
1287 		}
1288 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1289 			DRV_LOG(ERR,
1290 				"Invalid vport tag for port %d on bonding device %s",
1291 				spawn->phys_port, spawn->phys_dev_name);
1292 			err = ENOTSUP;
1293 			goto error;
1294 		}
1295 	}
1296 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1297 		priv->vport_id = vport_info.vport_id;
1298 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1299 		DRV_LOG(ERR,
1300 			"Cannot deduce vport index for port %d on bonding device %s",
1301 			spawn->phys_port, spawn->phys_dev_name);
1302 		err = ENOTSUP;
1303 		goto error;
1304 	} else {
1305 		/*
1306 		 * Suppose vport index in compatible way. Kernel/rdma_core
1307 		 * support single E-Switch per PF configurations only and
1308 		 * vport_id field contains the vport index for associated VF,
1309 		 * which is deduced from representor port name.
1310 		 * For example, let's have the IB device port 10, it has
1311 		 * attached network device eth0, which has port name attribute
1312 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1313 		 * as 3 (2+1). This assigning schema should be changed if the
1314 		 * multiple E-Switch instances per PF configurations or/and PCI
1315 		 * subfunctions are added.
1316 		 */
1317 		priv->vport_id = switch_info->representor ?
1318 				 switch_info->port_name + 1 : -1;
1319 	}
1320 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1321 							  eth_da->type);
1322 	/*
1323 	 * Look for sibling devices in order to reuse their switch domain
1324 	 * if any, otherwise allocate one.
1325 	 */
1326 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1327 		const struct mlx5_priv *opriv =
1328 			rte_eth_devices[port_id].data->dev_private;
1329 
1330 		if (!opriv ||
1331 		    opriv->sh != priv->sh ||
1332 			opriv->domain_id ==
1333 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1334 			continue;
1335 		priv->domain_id = opriv->domain_id;
1336 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1337 			priv->dev_port, priv->domain_id);
1338 		break;
1339 	}
1340 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1341 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1342 		if (err) {
1343 			err = rte_errno;
1344 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1345 				strerror(rte_errno));
1346 			goto error;
1347 		}
1348 		own_domain_id = 1;
1349 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1350 			priv->dev_port, priv->domain_id);
1351 	}
1352 	if (sh->cdev->config.devx) {
1353 		struct mlx5_hca_attr *hca_attr = &sh->cdev->config.hca_attr;
1354 
1355 		sh->steering_format_version = hca_attr->steering_format_version;
1356 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO_EXT)
1357 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1358 		    sh->config.dv_flow_en) {
1359 			if (sh->registers.aso_reg != REG_NON) {
1360 				priv->mtr_en = 1;
1361 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1362 			}
1363 		}
1364 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1365 			uint32_t log_obj_size =
1366 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1367 			if (log_obj_size >=
1368 			    hca_attr->qos.log_meter_aso_granularity &&
1369 			    log_obj_size <=
1370 			    hca_attr->qos.log_meter_aso_max_alloc)
1371 				sh->meter_aso_en = 1;
1372 		}
1373 		if (priv->mtr_en) {
1374 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1375 			if (err) {
1376 				err = -err;
1377 				goto error;
1378 			}
1379 		}
1380 		if (hca_attr->flow.tunnel_header_0_1)
1381 			sh->tunnel_header_0_1 = 1;
1382 		if (hca_attr->flow.tunnel_header_2_3)
1383 			sh->tunnel_header_2_3 = 1;
1384 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO_EXT */
1385 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1386 		if (hca_attr->flow_hit_aso && sh->registers.aso_reg == REG_C_3) {
1387 			sh->flow_hit_aso_en = 1;
1388 			err = mlx5_flow_aso_age_mng_init(sh);
1389 			if (err) {
1390 				err = -err;
1391 				goto error;
1392 			}
1393 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1394 		}
1395 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1396 #if defined (HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1397     defined (HAVE_MLX5_DR_ACTION_ASO_CT)
1398 		/* HWS create CT ASO SQ based on HWS configure queue number. */
1399 		if (sh->config.dv_flow_en != 2 &&
1400 		    hca_attr->ct_offload && sh->registers.aso_reg == REG_C_3) {
1401 			err = mlx5_flow_aso_ct_mng_init(sh);
1402 			if (err) {
1403 				err = -err;
1404 				goto error;
1405 			}
1406 			DRV_LOG(DEBUG, "CT ASO is supported.");
1407 			sh->ct_aso_en = 1;
1408 		}
1409 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1410 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1411 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1412 		    sh->config.dv_flow_en) {
1413 			priv->sampler_en = 1;
1414 			DRV_LOG(DEBUG, "Sampler enabled!");
1415 		} else {
1416 			priv->sampler_en = 0;
1417 			if (!hca_attr->log_max_ft_sampler_num)
1418 				DRV_LOG(WARNING,
1419 					"No available register for sampler.");
1420 			else
1421 				DRV_LOG(DEBUG, "DV flow is not supported!");
1422 		}
1423 #endif
1424 		if (hca_attr->lag_rx_port_affinity) {
1425 			sh->lag_rx_port_affinity_en = 1;
1426 			DRV_LOG(DEBUG, "LAG Rx Port Affinity enabled");
1427 		}
1428 		priv->num_lag_ports = hca_attr->num_lag_ports;
1429 		DRV_LOG(DEBUG, "The number of lag ports is %d", priv->num_lag_ports);
1430 	}
1431 	/* Process parameters and store port configuration on priv structure. */
1432 	err = mlx5_port_args_config(priv, mkvlist, &priv->config);
1433 	if (err) {
1434 		err = rte_errno;
1435 		DRV_LOG(ERR, "Failed to process port configure: %s",
1436 			strerror(rte_errno));
1437 		goto error;
1438 	}
1439 	eth_dev = rte_eth_dev_allocate(name);
1440 	if (eth_dev == NULL) {
1441 		DRV_LOG(ERR, "can not allocate rte ethdev");
1442 		err = ENOMEM;
1443 		goto error;
1444 	}
1445 	if (priv->representor) {
1446 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1447 		eth_dev->data->representor_id = priv->representor_id;
1448 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1449 			struct mlx5_priv *opriv =
1450 				rte_eth_devices[port_id].data->dev_private;
1451 			if (opriv &&
1452 			    opriv->master &&
1453 			    opriv->domain_id == priv->domain_id &&
1454 			    opriv->sh == priv->sh) {
1455 				eth_dev->data->backer_port_id = port_id;
1456 				break;
1457 			}
1458 		}
1459 		if (port_id >= RTE_MAX_ETHPORTS)
1460 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1461 	}
1462 	priv->mp_id.port_id = eth_dev->data->port_id;
1463 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1464 	/*
1465 	 * Store associated network device interface index. This index
1466 	 * is permanent throughout the lifetime of device. So, we may store
1467 	 * the ifindex here and use the cached value further.
1468 	 */
1469 	MLX5_ASSERT(spawn->ifindex);
1470 	priv->if_index = spawn->ifindex;
1471 	priv->lag_affinity_idx = sh->refcnt - 1;
1472 	eth_dev->data->dev_private = priv;
1473 	priv->dev_data = eth_dev->data;
1474 	eth_dev->data->mac_addrs = priv->mac;
1475 	eth_dev->device = dpdk_dev;
1476 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1477 	/* Configure the first MAC address by default. */
1478 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1479 		DRV_LOG(ERR,
1480 			"port %u cannot get MAC address, is mlx5_en"
1481 			" loaded? (errno: %s)",
1482 			eth_dev->data->port_id, strerror(rte_errno));
1483 		err = ENODEV;
1484 		goto error;
1485 	}
1486 	DRV_LOG(INFO,
1487 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1488 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1489 #ifdef RTE_LIBRTE_MLX5_DEBUG
1490 	{
1491 		char ifname[MLX5_NAMESIZE];
1492 
1493 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1494 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1495 				eth_dev->data->port_id, ifname);
1496 		else
1497 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1498 				eth_dev->data->port_id);
1499 	}
1500 #endif
1501 	/* Get actual MTU if possible. */
1502 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1503 	if (err) {
1504 		err = rte_errno;
1505 		goto error;
1506 	}
1507 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1508 		priv->mtu);
1509 	/* Initialize burst functions to prevent crashes before link-up. */
1510 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1511 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1512 	eth_dev->dev_ops = &mlx5_dev_ops;
1513 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1514 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1515 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1516 	/* Register MAC address. */
1517 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1518 	if (sh->dev_cap.vf && sh->config.vf_nl_en)
1519 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1520 				      mlx5_ifindex(eth_dev),
1521 				      eth_dev->data->mac_addrs,
1522 				      MLX5_MAX_MAC_ADDRESSES);
1523 	priv->ctrl_flows = 0;
1524 	rte_spinlock_init(&priv->flow_list_lock);
1525 	TAILQ_INIT(&priv->flow_meters);
1526 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1527 	if (!priv->mtr_profile_tbl)
1528 		goto error;
1529 	/* Bring Ethernet device up. */
1530 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1531 		eth_dev->data->port_id);
1532 	/* Read link status in case it is up and there will be no event. */
1533 	mlx5_link_update(eth_dev, 0);
1534 	/* Watch LSC interrupts between port probe and port start. */
1535 	priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1536 							eth_dev->data->port_id;
1537 	mlx5_set_link_up(eth_dev);
1538 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1539 		icfg[i].release_mem_en = !!sh->config.reclaim_mode;
1540 		if (sh->config.reclaim_mode)
1541 			icfg[i].per_core_cache = 0;
1542 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1543 		if (!priv->flows[i])
1544 			goto error;
1545 	}
1546 	/* Create context for virtual machine VLAN workaround. */
1547 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1548 	if (sh->config.dv_flow_en) {
1549 		err = mlx5_alloc_shared_dr(priv);
1550 		if (err)
1551 			goto error;
1552 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1553 			goto error;
1554 	}
1555 	if (mlx5_devx_obj_ops_en(sh)) {
1556 		priv->obj_ops = devx_obj_ops;
1557 		mlx5_queue_counter_id_prepare(eth_dev);
1558 		priv->obj_ops.lb_dummy_queue_create =
1559 					mlx5_rxq_ibv_obj_dummy_lb_create;
1560 		priv->obj_ops.lb_dummy_queue_release =
1561 					mlx5_rxq_ibv_obj_dummy_lb_release;
1562 	} else if (spawn->max_port > UINT8_MAX) {
1563 		/* Verbs can't support ports larger than 255 by design. */
1564 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1565 		err = ENOTSUP;
1566 		goto error;
1567 	} else {
1568 		priv->obj_ops = ibv_obj_ops;
1569 	}
1570 	if (sh->config.tx_pp &&
1571 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1572 		/*
1573 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1574 		 * packet pacing and already checked above.
1575 		 * Hence, we should only make sure the SQs will be created
1576 		 * with DevX, not with Verbs.
1577 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1578 		 * with Clock Queue UAR as required for Tx scheduling.
1579 		 */
1580 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1581 		err = ENODEV;
1582 		goto error;
1583 	}
1584 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1585 	if (!priv->drop_queue.hrxq)
1586 		goto error;
1587 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1588 				       mlx5_hrxq_create_cb,
1589 				       mlx5_hrxq_match_cb,
1590 				       mlx5_hrxq_remove_cb,
1591 				       mlx5_hrxq_clone_cb,
1592 				       mlx5_hrxq_clone_free_cb);
1593 	if (!priv->hrxqs)
1594 		goto error;
1595 	mlx5_set_metadata_mask(eth_dev);
1596 	if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1597 	    !priv->sh->dv_regc0_mask) {
1598 		DRV_LOG(ERR, "metadata mode %u is not supported "
1599 			     "(no metadata reg_c[0] is available)",
1600 			     sh->config.dv_xmeta_en);
1601 			err = ENOTSUP;
1602 			goto error;
1603 	}
1604 	rte_rwlock_init(&priv->ind_tbls_lock);
1605 	if (sh->phdev->config.ipv6_tc_fallback == MLX5_IPV6_TC_UNKNOWN) {
1606 		if (!sh->cdev->config.hca_attr.modify_outer_ipv6_traffic_class ||
1607 		    (sh->config.dv_flow_en == 1 && mlx5_flow_discover_ipv6_tc_support(eth_dev)))
1608 			sh->phdev->config.ipv6_tc_fallback = MLX5_IPV6_TC_FALLBACK;
1609 		else
1610 			sh->phdev->config.ipv6_tc_fallback = MLX5_IPV6_TC_OK;
1611 	}
1612 	if (priv->sh->config.dv_flow_en == 2) {
1613 #ifdef HAVE_MLX5_HWS_SUPPORT
1614 		if (priv->sh->config.dv_esw_en) {
1615 			uint32_t usable_bits;
1616 			uint32_t required_bits;
1617 
1618 			if (priv->sh->dv_regc0_mask == UINT32_MAX) {
1619 				DRV_LOG(ERR, "E-Switch port metadata is required when using HWS "
1620 					     "but it is disabled (configure it through devlink)");
1621 				err = ENOTSUP;
1622 				goto error;
1623 			}
1624 			if (priv->sh->dv_regc0_mask == 0) {
1625 				DRV_LOG(ERR, "E-Switch with HWS is not supported "
1626 					     "(no available bits in reg_c[0])");
1627 				err = ENOTSUP;
1628 				goto error;
1629 			}
1630 			usable_bits = rte_popcount32(priv->sh->dv_regc0_mask);
1631 			required_bits = rte_popcount32(priv->vport_meta_mask);
1632 			if (usable_bits < required_bits) {
1633 				DRV_LOG(ERR, "Not enough bits available in reg_c[0] to provide "
1634 					     "representor matching.");
1635 				err = ENOTSUP;
1636 				goto error;
1637 			}
1638 		}
1639 		if (priv->vport_meta_mask)
1640 			flow_hw_set_port_info(eth_dev);
1641 		if (priv->sh->config.dv_esw_en &&
1642 		    priv->sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1643 		    priv->sh->config.dv_xmeta_en != MLX5_XMETA_MODE_META32_HWS) {
1644 			DRV_LOG(ERR,
1645 				"metadata mode %u is not supported in HWS eswitch mode",
1646 				priv->sh->config.dv_xmeta_en);
1647 				err = ENOTSUP;
1648 				goto error;
1649 		}
1650 		if (priv->sh->config.dv_esw_en &&
1651 		    flow_hw_create_vport_action(eth_dev)) {
1652 			DRV_LOG(ERR, "port %u failed to create vport action",
1653 				eth_dev->data->port_id);
1654 			err = EINVAL;
1655 			goto error;
1656 		}
1657 		/*
1658 		 * If representor matching is disabled, PMD cannot create default flow rules
1659 		 * to receive traffic for all ports, since implicit source port match is not added.
1660 		 * Isolated mode is forced.
1661 		 */
1662 		if (priv->sh->config.dv_esw_en && !priv->sh->config.repr_matching) {
1663 			err = mlx5_flow_isolate(eth_dev, 1, NULL);
1664 			if (err < 0) {
1665 				err = -err;
1666 				goto error;
1667 			}
1668 			DRV_LOG(WARNING, "port %u ingress traffic is restricted to defined "
1669 					 "flow rules (isolated mode) since representor "
1670 					 "matching is disabled",
1671 				eth_dev->data->port_id);
1672 		}
1673 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1674 		return eth_dev;
1675 #else
1676 		DRV_LOG(ERR, "DV support is missing for HWS.");
1677 		goto error;
1678 #endif
1679 	}
1680 	if (!priv->sh->flow_priority_check_flag) {
1681 		/* Supported Verbs flow priority number detection. */
1682 		err = mlx5_flow_discover_priorities(eth_dev);
1683 		priv->sh->flow_max_priority = err;
1684 		priv->sh->flow_priority_check_flag = 1;
1685 	} else {
1686 		err = priv->sh->flow_max_priority;
1687 	}
1688 	if (err < 0) {
1689 		err = -err;
1690 		goto error;
1691 	}
1692 	/* Query availability of metadata reg_c's. */
1693 	if (!priv->sh->metadata_regc_check_flag) {
1694 		err = mlx5_flow_discover_mreg_c(eth_dev);
1695 		if (err < 0) {
1696 			err = -err;
1697 			goto error;
1698 		}
1699 	}
1700 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1701 		DRV_LOG(DEBUG,
1702 			"port %u extensive metadata register is not supported",
1703 			eth_dev->data->port_id);
1704 		if (sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1705 			DRV_LOG(ERR, "metadata mode %u is not supported "
1706 				     "(no metadata registers available)",
1707 				     sh->config.dv_xmeta_en);
1708 			err = ENOTSUP;
1709 			goto error;
1710 		}
1711 	}
1712 	if (sh->config.dv_flow_en &&
1713 	    sh->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1714 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1715 	    priv->sh->dv_regc0_mask) {
1716 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1717 						      MLX5_FLOW_MREG_HTABLE_SZ,
1718 						      false, true, eth_dev,
1719 						      flow_dv_mreg_create_cb,
1720 						      flow_dv_mreg_match_cb,
1721 						      flow_dv_mreg_remove_cb,
1722 						      flow_dv_mreg_clone_cb,
1723 						    flow_dv_mreg_clone_free_cb);
1724 		if (!priv->mreg_cp_tbl) {
1725 			err = ENOMEM;
1726 			goto error;
1727 		}
1728 	}
1729 	rte_spinlock_init(&priv->shared_act_sl);
1730 	mlx5_flow_counter_mode_config(eth_dev);
1731 	mlx5_flow_drop_action_config(eth_dev);
1732 	if (sh->config.dv_flow_en)
1733 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1734 	return eth_dev;
1735 error:
1736 	if (priv) {
1737 		priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1738 							       RTE_MAX_ETHPORTS;
1739 		rte_io_wmb();
1740 #ifdef HAVE_MLX5_HWS_SUPPORT
1741 		if (eth_dev &&
1742 		    priv->sh &&
1743 		    priv->sh->config.dv_flow_en == 2 &&
1744 		    priv->sh->config.dv_esw_en)
1745 			flow_hw_destroy_vport_action(eth_dev);
1746 #endif
1747 		if (priv->mreg_cp_tbl)
1748 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1749 		if (priv->sh)
1750 			mlx5_os_free_shared_dr(priv);
1751 		if (priv->nl_socket_route >= 0)
1752 			close(priv->nl_socket_route);
1753 		if (priv->vmwa_context)
1754 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1755 		if (eth_dev && priv->drop_queue.hrxq)
1756 			mlx5_drop_action_destroy(eth_dev);
1757 		if (priv->mtr_profile_tbl)
1758 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1759 		if (own_domain_id)
1760 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1761 		if (priv->hrxqs)
1762 			mlx5_list_destroy(priv->hrxqs);
1763 		if (eth_dev && priv->flex_item_map)
1764 			mlx5_flex_item_port_cleanup(eth_dev);
1765 		mlx5_free(priv->ext_rxqs);
1766 		mlx5_free(priv);
1767 		if (eth_dev != NULL)
1768 			eth_dev->data->dev_private = NULL;
1769 	}
1770 	if (eth_dev != NULL) {
1771 		/* mac_addrs must not be freed alone because part of
1772 		 * dev_private
1773 		 **/
1774 		eth_dev->data->mac_addrs = NULL;
1775 		rte_eth_dev_release_port(eth_dev);
1776 	}
1777 	if (sh)
1778 		mlx5_free_shared_dev_ctx(sh);
1779 	if (nl_rdma >= 0)
1780 		close(nl_rdma);
1781 	MLX5_ASSERT(err > 0);
1782 	rte_errno = err;
1783 	return NULL;
1784 }
1785 
1786 /**
1787  * Comparison callback to sort device data.
1788  *
1789  * This is meant to be used with qsort().
1790  *
1791  * @param a[in]
1792  *   Pointer to pointer to first data object.
1793  * @param b[in]
1794  *   Pointer to pointer to second data object.
1795  *
1796  * @return
1797  *   0 if both objects are equal, less than 0 if the first argument is less
1798  *   than the second, greater than 0 otherwise.
1799  */
1800 static int
1801 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1802 {
1803 	const struct mlx5_switch_info *si_a =
1804 		&((const struct mlx5_dev_spawn_data *)a)->info;
1805 	const struct mlx5_switch_info *si_b =
1806 		&((const struct mlx5_dev_spawn_data *)b)->info;
1807 	int uplink_a = si_a->name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1808 	int uplink_b = si_b->name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1809 	int ret;
1810 
1811 	/* Uplink ports first. */
1812 	ret = uplink_b - uplink_a;
1813 	if (ret)
1814 		return ret;
1815 	/* Then master devices. */
1816 	ret = si_b->master - si_a->master;
1817 	if (ret)
1818 		return ret;
1819 	/* Then representor devices. */
1820 	ret = si_b->representor - si_a->representor;
1821 	if (ret)
1822 		return ret;
1823 	/* Unidentified devices come last in no specific order. */
1824 	if (!si_a->representor)
1825 		return 0;
1826 	/* Order representors by name. */
1827 	return si_a->port_name - si_b->port_name;
1828 }
1829 
1830 /**
1831  * Match PCI information for possible slaves of bonding device.
1832  *
1833  * @param[in] ibdev_name
1834  *   Name of Infiniband device.
1835  * @param[in] pci_dev
1836  *   Pointer to primary PCI address structure to match.
1837  * @param[in] nl_rdma
1838  *   Netlink RDMA group socket handle.
1839  * @param[in] owner
1840  *   Representor owner PF index.
1841  * @param[out] bond_info
1842  *   Pointer to bonding information.
1843  *
1844  * @return
1845  *   negative value if no bonding device found, otherwise
1846  *   positive index of slave PF in bonding.
1847  */
1848 static int
1849 mlx5_device_bond_pci_match(const char *ibdev_name,
1850 			   const struct rte_pci_addr *pci_dev,
1851 			   int nl_rdma, uint16_t owner,
1852 			   struct mlx5_bond_info *bond_info)
1853 {
1854 	char ifname[IF_NAMESIZE + 1];
1855 	unsigned int ifindex;
1856 	unsigned int np, i;
1857 	FILE *bond_file = NULL, *file;
1858 	int pf = -1;
1859 	int ret;
1860 	uint8_t cur_guid[32] = {0};
1861 	uint8_t guid[32] = {0};
1862 
1863 	/*
1864 	 * Try to get master device name. If something goes wrong suppose
1865 	 * the lack of kernel support and no bonding devices.
1866 	 */
1867 	memset(bond_info, 0, sizeof(*bond_info));
1868 	if (nl_rdma < 0)
1869 		return -1;
1870 	if (!strstr(ibdev_name, "bond"))
1871 		return -1;
1872 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1873 	if (!np)
1874 		return -1;
1875 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1876 		return -1;
1877 	/*
1878 	 * The master device might not be on the predefined port(not on port
1879 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1880 	 * device ports and find master.
1881 	 */
1882 	for (i = 1; i <= np; ++i) {
1883 		/* Check whether Infiniband port is populated. */
1884 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1885 		if (!ifindex)
1886 			continue;
1887 		if (!if_indextoname(ifindex, ifname))
1888 			continue;
1889 		/* Try to read bonding slave names from sysfs. */
1890 		MKSTR(slaves,
1891 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1892 		bond_file = fopen(slaves, "r");
1893 		if (bond_file)
1894 			break;
1895 	}
1896 	if (!bond_file)
1897 		return -1;
1898 	/* Use safe format to check maximal buffer length. */
1899 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1900 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1901 		char tmp_str[IF_NAMESIZE + 32];
1902 		struct rte_pci_addr pci_addr;
1903 		struct mlx5_switch_info	info;
1904 		int ret;
1905 
1906 		/* Process slave interface names in the loop. */
1907 		snprintf(tmp_str, sizeof(tmp_str),
1908 			 "/sys/class/net/%s", ifname);
1909 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1910 			DRV_LOG(WARNING,
1911 				"Cannot get PCI address for netdev \"%s\".",
1912 				ifname);
1913 			continue;
1914 		}
1915 		/* Slave interface PCI address match found. */
1916 		snprintf(tmp_str, sizeof(tmp_str),
1917 			 "/sys/class/net/%s/phys_port_name", ifname);
1918 		file = fopen(tmp_str, "rb");
1919 		if (!file)
1920 			break;
1921 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1922 		if (fscanf(file, "%32s", tmp_str) == 1)
1923 			mlx5_translate_port_name(tmp_str, &info);
1924 		fclose(file);
1925 		/* Only process PF ports. */
1926 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1927 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1928 			continue;
1929 		/* Check max bonding member. */
1930 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1931 			DRV_LOG(WARNING, "bonding index out of range, "
1932 				"please increase MLX5_BOND_MAX_PORTS: %s",
1933 				tmp_str);
1934 			break;
1935 		}
1936 		/* Get ifindex. */
1937 		snprintf(tmp_str, sizeof(tmp_str),
1938 			 "/sys/class/net/%s/ifindex", ifname);
1939 		file = fopen(tmp_str, "rb");
1940 		if (!file)
1941 			break;
1942 		ret = fscanf(file, "%u", &ifindex);
1943 		fclose(file);
1944 		if (ret != 1)
1945 			break;
1946 		/* Save bonding info. */
1947 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1948 			sizeof(bond_info->ports[0].ifname));
1949 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1950 		bond_info->ports[info.port_name].ifindex = ifindex;
1951 		bond_info->n_port++;
1952 		/*
1953 		 * Under socket direct mode, bonding will use
1954 		 * system_image_guid as identification.
1955 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1956 		 * All bonding members should have the same guid even if driver
1957 		 * is using PCIe BDF.
1958 		 */
1959 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1960 		if (ret < 0)
1961 			break;
1962 		else if (ret > 0) {
1963 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1964 			    owner == info.port_name &&
1965 			    (owner != 0 || (owner == 0 &&
1966 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1967 				pf = info.port_name;
1968 		} else if (pci_dev->domain == pci_addr.domain &&
1969 		    pci_dev->bus == pci_addr.bus &&
1970 		    pci_dev->devid == pci_addr.devid &&
1971 		    ((pci_dev->function == 0 &&
1972 		      pci_dev->function + owner == pci_addr.function) ||
1973 		     (pci_dev->function == owner &&
1974 		      pci_addr.function == owner)))
1975 			pf = info.port_name;
1976 	}
1977 	if (pf >= 0) {
1978 		/* Get bond interface info */
1979 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1980 					   bond_info->ifname);
1981 		if (ret)
1982 			DRV_LOG(ERR, "unable to get bond info: %s",
1983 				strerror(rte_errno));
1984 		else
1985 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1986 				ifindex, bond_info->ifindex, bond_info->ifname);
1987 	}
1988 	if (owner == 0 && pf != 0) {
1989 		DRV_LOG(INFO, "PCIe instance " PCI_PRI_FMT " isn't bonding owner",
1990 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1991 				pci_dev->function);
1992 	}
1993 	return pf;
1994 }
1995 
1996 static int
1997 mlx5_nl_esw_multiport_get(struct rte_pci_addr *pci_addr, int *enabled)
1998 {
1999 	char pci_addr_str[PCI_PRI_STR_SIZE] = { 0 };
2000 	int nlsk_fd;
2001 	int devlink_id;
2002 	int ret;
2003 
2004 	/* Provide correct value to have defined enabled state in case of an error. */
2005 	*enabled = 0;
2006 	rte_pci_device_name(pci_addr, pci_addr_str, sizeof(pci_addr_str));
2007 	nlsk_fd = mlx5_nl_init(NETLINK_GENERIC, 0);
2008 	if (nlsk_fd < 0)
2009 		return nlsk_fd;
2010 	devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
2011 	if (devlink_id < 0) {
2012 		ret = devlink_id;
2013 		DRV_LOG(DEBUG, "Unable to get devlink family id for Multiport E-Switch checks "
2014 			       "by netlink, for PCI device %s", pci_addr_str);
2015 		goto close_nlsk_fd;
2016 	}
2017 	ret = mlx5_nl_devlink_esw_multiport_get(nlsk_fd, devlink_id, pci_addr_str, enabled);
2018 	if (ret < 0)
2019 		DRV_LOG(DEBUG, "Unable to get Multiport E-Switch state by Netlink.");
2020 close_nlsk_fd:
2021 	close(nlsk_fd);
2022 	return ret;
2023 }
2024 
2025 #define SYSFS_MPESW_PARAM_MAX_LEN 16
2026 
2027 static int
2028 mlx5_sysfs_esw_multiport_get(struct ibv_device *ibv, struct rte_pci_addr *pci_addr, int *enabled)
2029 {
2030 	int nl_rdma;
2031 	unsigned int n_ports;
2032 	unsigned int i;
2033 	int ret;
2034 
2035 	/* Provide correct value to have defined enabled state in case of an error. */
2036 	*enabled = 0;
2037 	nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
2038 	if (nl_rdma < 0)
2039 		return nl_rdma;
2040 	n_ports = mlx5_nl_portnum(nl_rdma, ibv->name);
2041 	if (!n_ports) {
2042 		ret = -rte_errno;
2043 		goto close_nl_rdma;
2044 	}
2045 	for (i = 1; i <= n_ports; ++i) {
2046 		unsigned int ifindex;
2047 		char ifname[IF_NAMESIZE + 1];
2048 		struct rte_pci_addr if_pci_addr;
2049 		char mpesw[SYSFS_MPESW_PARAM_MAX_LEN + 1];
2050 		FILE *sysfs;
2051 		int n;
2052 
2053 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv->name, i);
2054 		if (!ifindex)
2055 			continue;
2056 		if (!if_indextoname(ifindex, ifname))
2057 			continue;
2058 		MKSTR(sysfs_if_path, "/sys/class/net/%s", ifname);
2059 		if (mlx5_get_pci_addr(sysfs_if_path, &if_pci_addr))
2060 			continue;
2061 		if (pci_addr->domain != if_pci_addr.domain ||
2062 		    pci_addr->bus != if_pci_addr.bus ||
2063 		    pci_addr->devid != if_pci_addr.devid ||
2064 		    pci_addr->function != if_pci_addr.function)
2065 			continue;
2066 		MKSTR(sysfs_mpesw_path,
2067 		      "/sys/class/net/%s/compat/devlink/lag_port_select_mode", ifname);
2068 		sysfs = fopen(sysfs_mpesw_path, "r");
2069 		if (!sysfs)
2070 			continue;
2071 		n = fscanf(sysfs, "%" RTE_STR(SYSFS_MPESW_PARAM_MAX_LEN) "s", mpesw);
2072 		fclose(sysfs);
2073 		if (n != 1)
2074 			continue;
2075 		ret = 0;
2076 		if (strcmp(mpesw, "multiport_esw") == 0) {
2077 			*enabled = 1;
2078 			break;
2079 		}
2080 		*enabled = 0;
2081 		break;
2082 	}
2083 	if (i > n_ports) {
2084 		DRV_LOG(DEBUG, "Unable to get Multiport E-Switch state by sysfs.");
2085 		rte_errno = ENOENT;
2086 		ret = -rte_errno;
2087 	}
2088 
2089 close_nl_rdma:
2090 	close(nl_rdma);
2091 	return ret;
2092 }
2093 
2094 static int
2095 mlx5_is_mpesw_enabled(struct ibv_device *ibv, struct rte_pci_addr *ibv_pci_addr, int *enabled)
2096 {
2097 	/*
2098 	 * Try getting Multiport E-Switch state through netlink interface
2099 	 * If unable, try sysfs interface. If that is unable as well,
2100 	 * assume that Multiport E-Switch is disabled and return an error.
2101 	 */
2102 	if (mlx5_nl_esw_multiport_get(ibv_pci_addr, enabled) >= 0 ||
2103 	    mlx5_sysfs_esw_multiport_get(ibv, ibv_pci_addr, enabled) >= 0)
2104 		return 0;
2105 	DRV_LOG(DEBUG, "Unable to check MPESW state for IB device %s "
2106 		       "(PCI: " PCI_PRI_FMT ")",
2107 		       ibv->name,
2108 		       ibv_pci_addr->domain, ibv_pci_addr->bus,
2109 		       ibv_pci_addr->devid, ibv_pci_addr->function);
2110 	*enabled = 0;
2111 	return -rte_errno;
2112 }
2113 
2114 static int
2115 mlx5_device_mpesw_pci_match(struct ibv_device *ibv,
2116 			    const struct rte_pci_addr *owner_pci,
2117 			    int nl_rdma)
2118 {
2119 	struct rte_pci_addr ibdev_pci_addr = { 0 };
2120 	char ifname[IF_NAMESIZE + 1] = { 0 };
2121 	unsigned int ifindex;
2122 	unsigned int np;
2123 	unsigned int i;
2124 	int enabled = 0;
2125 	int ret;
2126 
2127 	/* Check if IB device's PCI address matches the probed PCI address. */
2128 	if (mlx5_get_pci_addr(ibv->ibdev_path, &ibdev_pci_addr)) {
2129 		DRV_LOG(DEBUG, "Skipping MPESW check for IB device %s since "
2130 			       "there is no underlying PCI device", ibv->name);
2131 		rte_errno = ENOENT;
2132 		return -rte_errno;
2133 	}
2134 	if (ibdev_pci_addr.domain != owner_pci->domain ||
2135 	    ibdev_pci_addr.bus != owner_pci->bus ||
2136 	    ibdev_pci_addr.devid != owner_pci->devid ||
2137 	    ibdev_pci_addr.function != owner_pci->function) {
2138 		return -1;
2139 	}
2140 	/* Check if IB device has MPESW enabled. */
2141 	if (mlx5_is_mpesw_enabled(ibv, &ibdev_pci_addr, &enabled))
2142 		return -1;
2143 	if (!enabled)
2144 		return -1;
2145 	/* Iterate through IB ports to find MPESW master uplink port. */
2146 	if (nl_rdma < 0)
2147 		return -1;
2148 	np = mlx5_nl_portnum(nl_rdma, ibv->name);
2149 	if (!np)
2150 		return -1;
2151 	for (i = 1; i <= np; ++i) {
2152 		struct rte_pci_addr pci_addr;
2153 		FILE *file;
2154 		char port_name[IF_NAMESIZE + 1];
2155 		struct mlx5_switch_info	info;
2156 
2157 		/* Check whether IB port has a corresponding netdev. */
2158 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv->name, i);
2159 		if (!ifindex)
2160 			continue;
2161 		if (!if_indextoname(ifindex, ifname))
2162 			continue;
2163 		/* Read port name and determine its type. */
2164 		MKSTR(ifphysportname, "/sys/class/net/%s/phys_port_name", ifname);
2165 		file = fopen(ifphysportname, "rb");
2166 		if (!file)
2167 			continue;
2168 		ret = fscanf(file, "%16s", port_name);
2169 		fclose(file);
2170 		if (ret != 1)
2171 			continue;
2172 		memset(&info, 0, sizeof(info));
2173 		mlx5_translate_port_name(port_name, &info);
2174 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2175 			continue;
2176 		/* Fetch PCI address of the device to which the netdev is bound. */
2177 		MKSTR(ifpath, "/sys/class/net/%s", ifname);
2178 		if (mlx5_get_pci_addr(ifpath, &pci_addr))
2179 			continue;
2180 		if (pci_addr.domain == ibdev_pci_addr.domain &&
2181 		    pci_addr.bus == ibdev_pci_addr.bus &&
2182 		    pci_addr.devid == ibdev_pci_addr.devid &&
2183 		    pci_addr.function == ibdev_pci_addr.function) {
2184 			MLX5_ASSERT(info.port_name >= 0);
2185 			return info.port_name;
2186 		}
2187 	}
2188 	/* No matching MPESW uplink port was found. */
2189 	return -1;
2190 }
2191 
2192 /**
2193  * Register a PCI device within bonding.
2194  *
2195  * This function spawns Ethernet devices out of a given PCI device and
2196  * bonding owner PF index.
2197  *
2198  * @param[in] cdev
2199  *   Pointer to common mlx5 device structure.
2200  * @param[in] req_eth_da
2201  *   Requested ethdev device argument.
2202  * @param[in] owner_id
2203  *   Requested owner PF port ID within bonding device, default to 0.
2204  * @param[in, out] mkvlist
2205  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2206  *
2207  * @return
2208  *   0 on success, a negative errno value otherwise and rte_errno is set.
2209  */
2210 static int
2211 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
2212 		     struct rte_eth_devargs *req_eth_da,
2213 		     uint16_t owner_id, struct mlx5_kvargs_ctrl *mkvlist)
2214 {
2215 	struct ibv_device **ibv_list;
2216 	/*
2217 	 * Number of found IB Devices matching with requested PCI BDF.
2218 	 * nd != 1 means there are multiple IB devices over the same
2219 	 * PCI device and we have representors and master.
2220 	 */
2221 	unsigned int nd = 0;
2222 	/*
2223 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2224 	 * we have the single multiport IB device, and there may be
2225 	 * representors attached to some of found ports.
2226 	 */
2227 	unsigned int np = 0;
2228 	/*
2229 	 * Number of DPDK ethernet devices to Spawn - either over
2230 	 * multiple IB devices or multiple ports of single IB device.
2231 	 * Actually this is the number of iterations to spawn.
2232 	 */
2233 	unsigned int ns = 0;
2234 	/*
2235 	 * Bonding device
2236 	 *   < 0 - no bonding device (single one)
2237 	 *  >= 0 - bonding device (value is slave PF index)
2238 	 */
2239 	int bd = -1;
2240 	/*
2241 	 * Multiport E-Switch (MPESW) device:
2242 	 *   < 0 - no MPESW device or could not determine if it is MPESW device,
2243 	 *  >= 0 - MPESW device. Value is the port index of the MPESW owner.
2244 	 */
2245 	int mpesw = MLX5_MPESW_PORT_INVALID;
2246 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2247 	struct mlx5_dev_spawn_data *list = NULL;
2248 	struct rte_eth_devargs eth_da = *req_eth_da;
2249 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2250 	struct mlx5_bond_info bond_info;
2251 	int ret = -1;
2252 
2253 	errno = 0;
2254 	ibv_list = mlx5_glue->get_device_list(&ret);
2255 	if (!ibv_list) {
2256 		rte_errno = errno ? errno : ENOSYS;
2257 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2258 		return -rte_errno;
2259 	}
2260 	/*
2261 	 * First scan the list of all Infiniband devices to find
2262 	 * matching ones, gathering into the list.
2263 	 */
2264 	struct ibv_device *ibv_match[ret + 1];
2265 	int nl_route = mlx5_nl_init(NETLINK_ROUTE, 0);
2266 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA, 0);
2267 	unsigned int i;
2268 
2269 	while (ret-- > 0) {
2270 		struct rte_pci_addr pci_addr;
2271 
2272 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2273 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2274 						nl_rdma, owner_id, &bond_info);
2275 		if (bd >= 0) {
2276 			/*
2277 			 * Bonding device detected. Only one match is allowed,
2278 			 * the bonding is supported over multi-port IB device,
2279 			 * there should be no matches on representor PCI
2280 			 * functions or non VF LAG bonding devices with
2281 			 * specified address.
2282 			 */
2283 			if (nd) {
2284 				DRV_LOG(ERR,
2285 					"multiple PCI match on bonding device"
2286 					"\"%s\" found", ibv_list[ret]->name);
2287 				rte_errno = ENOENT;
2288 				ret = -rte_errno;
2289 				goto exit;
2290 			}
2291 			/* Amend owner pci address if owner PF ID specified. */
2292 			if (eth_da.nb_representor_ports)
2293 				owner_pci.function += owner_id;
2294 			DRV_LOG(INFO,
2295 				"PCI information matches for slave %d bonding device \"%s\"",
2296 				bd, ibv_list[ret]->name);
2297 			ibv_match[nd++] = ibv_list[ret];
2298 			break;
2299 		}
2300 		mpesw = mlx5_device_mpesw_pci_match(ibv_list[ret], &owner_pci, nl_rdma);
2301 		if (mpesw >= 0) {
2302 			/*
2303 			 * MPESW device detected. Only one matching IB device is allowed,
2304 			 * so if any matches were found previously, fail gracefully.
2305 			 */
2306 			if (nd) {
2307 				DRV_LOG(ERR,
2308 					"PCI information matches MPESW device \"%s\", "
2309 					"but multiple matching PCI devices were found. "
2310 					"Probing failed.",
2311 					ibv_list[ret]->name);
2312 				rte_errno = ENOENT;
2313 				ret = -rte_errno;
2314 				goto exit;
2315 			}
2316 			DRV_LOG(INFO,
2317 				"PCI information matches MPESW device \"%s\"",
2318 				ibv_list[ret]->name);
2319 			ibv_match[nd++] = ibv_list[ret];
2320 			break;
2321 		}
2322 		/* Bonding or MPESW device was not found. */
2323 		if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2324 					&pci_addr))
2325 			continue;
2326 		if (rte_pci_addr_cmp(&owner_pci, &pci_addr) != 0)
2327 			continue;
2328 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2329 			ibv_list[ret]->name);
2330 		ibv_match[nd++] = ibv_list[ret];
2331 	}
2332 	ibv_match[nd] = NULL;
2333 	if (!nd) {
2334 		/* No device matches, just complain and bail out. */
2335 		DRV_LOG(WARNING,
2336 			"PF %u doesn't have Verbs device matches PCI device " PCI_PRI_FMT ","
2337 			" are kernel drivers loaded?",
2338 			owner_id, owner_pci.domain, owner_pci.bus,
2339 			owner_pci.devid, owner_pci.function);
2340 		rte_errno = ENOENT;
2341 		ret = -rte_errno;
2342 		goto exit;
2343 	}
2344 	if (nd == 1) {
2345 		/*
2346 		 * Found single matching device may have multiple ports.
2347 		 * Each port may be representor, we have to check the port
2348 		 * number and check the representors existence.
2349 		 */
2350 		if (nl_rdma >= 0)
2351 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2352 		if (!np)
2353 			DRV_LOG(WARNING,
2354 				"Cannot get IB device \"%s\" ports number.",
2355 				ibv_match[0]->name);
2356 		if (bd >= 0 && !np) {
2357 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2358 			rte_errno = ENOENT;
2359 			ret = -rte_errno;
2360 			goto exit;
2361 		}
2362 		if (mpesw >= 0 && !np) {
2363 			DRV_LOG(ERR, "Cannot get ports for MPESW device.");
2364 			rte_errno = ENOENT;
2365 			ret = -rte_errno;
2366 			goto exit;
2367 		}
2368 	}
2369 	/* Now we can determine the maximal amount of devices to be spawned. */
2370 	list = mlx5_malloc(MLX5_MEM_ZERO,
2371 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2372 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2373 	if (!list) {
2374 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2375 		rte_errno = ENOMEM;
2376 		ret = -rte_errno;
2377 		goto exit;
2378 	}
2379 	if (bd >= 0 || mpesw >= 0 || np > 1) {
2380 		/*
2381 		 * Single IB device with multiple ports found,
2382 		 * it may be E-Switch master device and representors.
2383 		 * We have to perform identification through the ports.
2384 		 */
2385 		MLX5_ASSERT(nl_rdma >= 0);
2386 		MLX5_ASSERT(ns == 0);
2387 		MLX5_ASSERT(nd == 1);
2388 		MLX5_ASSERT(np);
2389 		for (i = 1; i <= np; ++i) {
2390 			list[ns].bond_info = &bond_info;
2391 			list[ns].max_port = np;
2392 			list[ns].phys_port = i;
2393 			list[ns].phys_dev_name = ibv_match[0]->name;
2394 			list[ns].eth_dev = NULL;
2395 			list[ns].pci_dev = pci_dev;
2396 			list[ns].cdev = cdev;
2397 			list[ns].pf_bond = bd;
2398 			list[ns].mpesw_port = MLX5_MPESW_PORT_INVALID;
2399 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2400 							   ibv_match[0]->name,
2401 							   i);
2402 			if (!list[ns].ifindex) {
2403 				/*
2404 				 * No network interface index found for the
2405 				 * specified port, it means there is no
2406 				 * representor on this port. It's OK,
2407 				 * there can be disabled ports, for example
2408 				 * if sriov_numvfs < sriov_totalvfs.
2409 				 */
2410 				continue;
2411 			}
2412 			ret = -1;
2413 			if (nl_route >= 0)
2414 				ret = mlx5_nl_switch_info(nl_route,
2415 							  list[ns].ifindex,
2416 							  &list[ns].info);
2417 			if (ret || (!list[ns].info.representor &&
2418 				    !list[ns].info.master)) {
2419 				/*
2420 				 * We failed to recognize representors with
2421 				 * Netlink, let's try to perform the task
2422 				 * with sysfs.
2423 				 */
2424 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2425 							     &list[ns].info);
2426 			}
2427 			if (!ret && bd >= 0) {
2428 				switch (list[ns].info.name_type) {
2429 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2430 					if (np == 1) {
2431 						/*
2432 						 * Force standalone bonding
2433 						 * device for ROCE LAG
2434 						 * configurations.
2435 						 */
2436 						list[ns].info.master = 0;
2437 						list[ns].info.representor = 0;
2438 					}
2439 					if (list[ns].info.port_name == bd)
2440 						ns++;
2441 					break;
2442 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2443 					/* Fallthrough */
2444 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2445 					/* Fallthrough */
2446 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2447 					if (list[ns].info.pf_num == bd)
2448 						ns++;
2449 					break;
2450 				default:
2451 					break;
2452 				}
2453 				continue;
2454 			}
2455 			if (!ret && mpesw >= 0) {
2456 				switch (list[ns].info.name_type) {
2457 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2458 					/* Owner port is treated as master port. */
2459 					if (list[ns].info.port_name == mpesw) {
2460 						list[ns].info.master = 1;
2461 						list[ns].info.representor = 0;
2462 					} else {
2463 						list[ns].info.master = 0;
2464 						list[ns].info.representor = 1;
2465 					}
2466 					/*
2467 					 * Ports of this type have uplink port index
2468 					 * encoded in the name. This index is also a PF index.
2469 					 */
2470 					list[ns].info.pf_num = list[ns].info.port_name;
2471 					list[ns].mpesw_port = list[ns].info.port_name;
2472 					list[ns].info.mpesw_owner = mpesw;
2473 					ns++;
2474 					break;
2475 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2476 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2477 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2478 					/* Only spawn representors related to the probed PF. */
2479 					if (list[ns].info.pf_num == owner_id) {
2480 						/*
2481 						 * Ports of this type have PF index encoded in name,
2482 						 * which translate to the related uplink port index.
2483 						 */
2484 						list[ns].mpesw_port = list[ns].info.pf_num;
2485 						/* MPESW owner is also saved but not used now. */
2486 						list[ns].info.mpesw_owner = mpesw;
2487 						ns++;
2488 					}
2489 					break;
2490 				default:
2491 					break;
2492 				}
2493 				continue;
2494 			}
2495 			if (!ret && (list[ns].info.representor ^
2496 				     list[ns].info.master))
2497 				ns++;
2498 		}
2499 		if (!ns) {
2500 			DRV_LOG(ERR,
2501 				"Unable to recognize master/representors on the IB device with multiple ports.");
2502 			rte_errno = ENOENT;
2503 			ret = -rte_errno;
2504 			goto exit;
2505 		}
2506 	} else {
2507 		/*
2508 		 * The existence of several matching entries (nd > 1) means
2509 		 * port representors have been instantiated. No existing Verbs
2510 		 * call nor sysfs entries can tell them apart, this can only
2511 		 * be done through Netlink calls assuming kernel drivers are
2512 		 * recent enough to support them.
2513 		 *
2514 		 * In the event of identification failure through Netlink,
2515 		 * try again through sysfs, then:
2516 		 *
2517 		 * 1. A single IB device matches (nd == 1) with single
2518 		 *    port (np=0/1) and is not a representor, assume
2519 		 *    no switch support.
2520 		 *
2521 		 * 2. Otherwise no safe assumptions can be made;
2522 		 *    complain louder and bail out.
2523 		 */
2524 		for (i = 0; i != nd; ++i) {
2525 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2526 			list[ns].bond_info = NULL;
2527 			list[ns].max_port = 1;
2528 			list[ns].phys_port = 1;
2529 			list[ns].phys_dev_name = ibv_match[i]->name;
2530 			list[ns].eth_dev = NULL;
2531 			list[ns].pci_dev = pci_dev;
2532 			list[ns].cdev = cdev;
2533 			list[ns].pf_bond = -1;
2534 			list[ns].mpesw_port = MLX5_MPESW_PORT_INVALID;
2535 			list[ns].ifindex = 0;
2536 			if (nl_rdma >= 0)
2537 				list[ns].ifindex = mlx5_nl_ifindex
2538 							    (nl_rdma,
2539 							     ibv_match[i]->name,
2540 							     1);
2541 			if (!list[ns].ifindex) {
2542 				char ifname[IF_NAMESIZE];
2543 
2544 				/*
2545 				 * Netlink failed, it may happen with old
2546 				 * ib_core kernel driver (before 4.16).
2547 				 * We can assume there is old driver because
2548 				 * here we are processing single ports IB
2549 				 * devices. Let's try sysfs to retrieve
2550 				 * the ifindex. The method works for
2551 				 * master device only.
2552 				 */
2553 				if (nd > 1) {
2554 					/*
2555 					 * Multiple devices found, assume
2556 					 * representors, can not distinguish
2557 					 * master/representor and retrieve
2558 					 * ifindex via sysfs.
2559 					 */
2560 					continue;
2561 				}
2562 				ret = mlx5_get_ifname_sysfs
2563 					(ibv_match[i]->ibdev_path, ifname);
2564 				if (!ret)
2565 					list[ns].ifindex =
2566 						if_nametoindex(ifname);
2567 				if (!list[ns].ifindex) {
2568 					/*
2569 					 * No network interface index found
2570 					 * for the specified device, it means
2571 					 * there it is neither representor
2572 					 * nor master.
2573 					 */
2574 					continue;
2575 				}
2576 			}
2577 			ret = -1;
2578 			if (nl_route >= 0)
2579 				ret = mlx5_nl_switch_info(nl_route,
2580 							  list[ns].ifindex,
2581 							  &list[ns].info);
2582 			if (ret || (!list[ns].info.representor &&
2583 				    !list[ns].info.master)) {
2584 				/*
2585 				 * We failed to recognize representors with
2586 				 * Netlink, let's try to perform the task
2587 				 * with sysfs.
2588 				 */
2589 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2590 							     &list[ns].info);
2591 			}
2592 			if (!ret && (list[ns].info.representor ^
2593 				     list[ns].info.master)) {
2594 				ns++;
2595 			} else if ((nd == 1) &&
2596 				   !list[ns].info.representor &&
2597 				   !list[ns].info.master) {
2598 				/*
2599 				 * Single IB device with one physical port and
2600 				 * attached network device.
2601 				 * May be SRIOV is not enabled or there is no
2602 				 * representors.
2603 				 */
2604 				DRV_LOG(INFO, "No E-Switch support detected.");
2605 				ns++;
2606 				break;
2607 			}
2608 		}
2609 		if (!ns) {
2610 			DRV_LOG(ERR,
2611 				"Unable to recognize master/representors on the multiple IB devices.");
2612 			rte_errno = ENOENT;
2613 			ret = -rte_errno;
2614 			goto exit;
2615 		}
2616 		/*
2617 		 * New kernels may add the switch_id attribute for the case
2618 		 * there is no E-Switch and we wrongly recognized the only
2619 		 * device as master. Override this if there is the single
2620 		 * device with single port and new device name format present.
2621 		 */
2622 		if (nd == 1 &&
2623 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2624 			list[0].info.master = 0;
2625 			list[0].info.representor = 0;
2626 		}
2627 	}
2628 	MLX5_ASSERT(ns);
2629 	/*
2630 	 * Sort list to probe devices in natural order for users convenience
2631 	 * (i.e. master first, then representors from lowest to highest ID).
2632 	 */
2633 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2634 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2635 		/* Set devargs default values. */
2636 		if (eth_da.nb_mh_controllers == 0) {
2637 			eth_da.nb_mh_controllers = 1;
2638 			eth_da.mh_controllers[0] = 0;
2639 		}
2640 		if (eth_da.nb_ports == 0 && ns > 0) {
2641 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2642 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2643 					pci_dev->device.devargs->args);
2644 			eth_da.nb_ports = 1;
2645 			eth_da.ports[0] = list[0].info.pf_num;
2646 		}
2647 		if (eth_da.nb_representor_ports == 0) {
2648 			eth_da.nb_representor_ports = 1;
2649 			eth_da.representor_ports[0] = 0;
2650 		}
2651 	}
2652 	for (i = 0; i != ns; ++i) {
2653 		uint32_t restore;
2654 
2655 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i], &eth_da,
2656 						 mkvlist);
2657 		if (!list[i].eth_dev) {
2658 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2659 				break;
2660 			/* Device is disabled or already spawned. Ignore it. */
2661 			continue;
2662 		}
2663 		restore = list[i].eth_dev->data->dev_flags;
2664 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2665 		/**
2666 		 * Each representor has a dedicated interrupts vector.
2667 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2668 		 * representor eth_dev object because representor and PF
2669 		 * share the same PCI address.
2670 		 * Override representor device with a dedicated
2671 		 * interrupts handle here.
2672 		 * Representor interrupts handle is released in mlx5_dev_stop().
2673 		 */
2674 		if (list[i].info.representor) {
2675 			struct rte_intr_handle *intr_handle =
2676 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2677 			if (intr_handle == NULL) {
2678 				DRV_LOG(ERR,
2679 					"port %u failed to allocate memory for interrupt handler "
2680 					"Rx interrupts will not be supported",
2681 					i);
2682 				rte_errno = ENOMEM;
2683 				ret = -rte_errno;
2684 				goto exit;
2685 			}
2686 			list[i].eth_dev->intr_handle = intr_handle;
2687 		}
2688 		/* Restore non-PCI flags cleared by the above call. */
2689 		list[i].eth_dev->data->dev_flags |= restore;
2690 		rte_eth_dev_probing_finish(list[i].eth_dev);
2691 	}
2692 	if (i != ns) {
2693 		DRV_LOG(ERR,
2694 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2695 			" encountering an error: %s",
2696 			owner_pci.domain, owner_pci.bus,
2697 			owner_pci.devid, owner_pci.function,
2698 			strerror(rte_errno));
2699 		ret = -rte_errno;
2700 		/* Roll back. */
2701 		while (i--) {
2702 			if (!list[i].eth_dev)
2703 				continue;
2704 			mlx5_dev_close(list[i].eth_dev);
2705 			/* mac_addrs must not be freed because in dev_private */
2706 			list[i].eth_dev->data->mac_addrs = NULL;
2707 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2708 		}
2709 		/* Restore original error. */
2710 		rte_errno = -ret;
2711 	} else {
2712 		ret = 0;
2713 	}
2714 exit:
2715 	/*
2716 	 * Do the routine cleanup:
2717 	 * - close opened Netlink sockets
2718 	 * - free allocated spawn data array
2719 	 * - free the Infiniband device list
2720 	 */
2721 	if (nl_rdma >= 0)
2722 		close(nl_rdma);
2723 	if (nl_route >= 0)
2724 		close(nl_route);
2725 	if (list)
2726 		mlx5_free(list);
2727 	MLX5_ASSERT(ibv_list);
2728 	mlx5_glue->free_device_list(ibv_list);
2729 	return ret;
2730 }
2731 
2732 static int
2733 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2734 			  struct rte_eth_devargs *eth_da)
2735 {
2736 	int ret = 0;
2737 
2738 	if (dev->devargs == NULL)
2739 		return 0;
2740 	memset(eth_da, 0, sizeof(*eth_da));
2741 	/* Parse representor information first from class argument. */
2742 	if (dev->devargs->cls_str)
2743 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da, 1);
2744 	if (ret < 0) {
2745 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2746 			dev->devargs->cls_str);
2747 		return -rte_errno;
2748 	}
2749 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE && dev->devargs->args) {
2750 		/* Parse legacy device argument */
2751 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da, 1);
2752 		if (ret < 0) {
2753 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2754 				dev->devargs->args);
2755 			return -rte_errno;
2756 		}
2757 	}
2758 	return 0;
2759 }
2760 
2761 /**
2762  * Callback to register a PCI device.
2763  *
2764  * This function spawns Ethernet devices out of a given PCI device.
2765  *
2766  * @param[in] cdev
2767  *   Pointer to common mlx5 device structure.
2768  * @param[in, out] mkvlist
2769  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2770  *
2771  * @return
2772  *   0 on success, a negative errno value otherwise and rte_errno is set.
2773  */
2774 static int
2775 mlx5_os_pci_probe(struct mlx5_common_device *cdev,
2776 		  struct mlx5_kvargs_ctrl *mkvlist)
2777 {
2778 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2779 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2780 	int ret = 0;
2781 	uint16_t p;
2782 
2783 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2784 	if (ret != 0)
2785 		return ret;
2786 
2787 	if (eth_da.nb_ports > 0) {
2788 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2789 		for (p = 0; p < eth_da.nb_ports; p++) {
2790 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2791 						   eth_da.ports[p], mkvlist);
2792 			if (ret) {
2793 				DRV_LOG(INFO, "Probe of PCI device " PCI_PRI_FMT " "
2794 					"aborted due to proding failure of PF %u",
2795 					pci_dev->addr.domain, pci_dev->addr.bus,
2796 					pci_dev->addr.devid, pci_dev->addr.function,
2797 					eth_da.ports[p]);
2798 				mlx5_net_remove(cdev);
2799 				if (p != 0)
2800 					break;
2801 			}
2802 		}
2803 	} else {
2804 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0, mkvlist);
2805 	}
2806 	return ret;
2807 }
2808 
2809 /* Probe a single SF device on auxiliary bus, no representor support. */
2810 static int
2811 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev,
2812 			struct mlx5_kvargs_ctrl *mkvlist)
2813 {
2814 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2815 	struct mlx5_dev_spawn_data spawn = {
2816 		.pf_bond = -1,
2817 		.mpesw_port = MLX5_MPESW_PORT_INVALID,
2818 	};
2819 	struct rte_device *dev = cdev->dev;
2820 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2821 	struct rte_eth_dev *eth_dev;
2822 	int ret = 0;
2823 
2824 	/* Parse ethdev devargs. */
2825 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2826 	if (ret != 0)
2827 		return ret;
2828 	/* Init spawn data. */
2829 	spawn.max_port = 1;
2830 	spawn.phys_port = 1;
2831 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2832 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2833 	if (ret < 0) {
2834 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2835 		return ret;
2836 	}
2837 	spawn.ifindex = ret;
2838 	spawn.cdev = cdev;
2839 	/* Spawn device. */
2840 	eth_dev = mlx5_dev_spawn(dev, &spawn, &eth_da, mkvlist);
2841 	if (eth_dev == NULL)
2842 		return -rte_errno;
2843 	/* Post create. */
2844 	eth_dev->intr_handle = adev->intr_handle;
2845 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2846 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2847 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2848 		eth_dev->data->numa_node = dev->numa_node;
2849 	}
2850 	rte_eth_dev_probing_finish(eth_dev);
2851 	return 0;
2852 }
2853 
2854 /**
2855  * Net class driver callback to probe a device.
2856  *
2857  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2858  *
2859  * @param[in] cdev
2860  *   Pointer to the common mlx5 device.
2861  * @param[in, out] mkvlist
2862  *   Pointer to mlx5 kvargs control, can be NULL if there is no devargs.
2863  *
2864  * @return
2865  *   0 on success, a negative errno value otherwise and rte_errno is set.
2866  */
2867 int
2868 mlx5_os_net_probe(struct mlx5_common_device *cdev,
2869 		  struct mlx5_kvargs_ctrl *mkvlist)
2870 {
2871 	int ret;
2872 
2873 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2874 		mlx5_pmd_socket_init();
2875 	ret = mlx5_init_once();
2876 	if (ret) {
2877 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2878 			strerror(rte_errno));
2879 		return -rte_errno;
2880 	}
2881 	ret = mlx5_probe_again_args_validate(cdev, mkvlist);
2882 	if (ret) {
2883 		DRV_LOG(ERR, "Probe again parameters are not compatible : %s",
2884 			strerror(rte_errno));
2885 		return -rte_errno;
2886 	}
2887 	if (mlx5_dev_is_pci(cdev->dev))
2888 		return mlx5_os_pci_probe(cdev, mkvlist);
2889 	else
2890 		return mlx5_os_auxiliary_probe(cdev, mkvlist);
2891 }
2892 
2893 /**
2894  * Cleanup resources when the last device is closed.
2895  */
2896 void
2897 mlx5_os_net_cleanup(void)
2898 {
2899 	mlx5_pmd_socket_uninit();
2900 }
2901 
2902 /**
2903  * Install shared asynchronous device events handler.
2904  * This function is implemented to support event sharing
2905  * between multiple ports of single IB device.
2906  *
2907  * @param sh
2908  *   Pointer to mlx5_dev_ctx_shared object.
2909  */
2910 void
2911 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2912 {
2913 	struct ibv_context *ctx = sh->cdev->ctx;
2914 	int nlsk_fd;
2915 
2916 	sh->intr_handle = mlx5_os_interrupt_handler_create
2917 		(RTE_INTR_INSTANCE_F_SHARED, true,
2918 		 ctx->async_fd, mlx5_dev_interrupt_handler, sh);
2919 	if (!sh->intr_handle) {
2920 		DRV_LOG(ERR, "Failed to allocate intr_handle.");
2921 		return;
2922 	}
2923 	nlsk_fd = mlx5_nl_init(NETLINK_ROUTE, RTMGRP_LINK);
2924 	if (nlsk_fd < 0) {
2925 		DRV_LOG(ERR, "Failed to create a socket for Netlink events: %s",
2926 			rte_strerror(rte_errno));
2927 		return;
2928 	}
2929 	sh->intr_handle_nl = mlx5_os_interrupt_handler_create
2930 		(RTE_INTR_INSTANCE_F_SHARED, true,
2931 		 nlsk_fd, mlx5_dev_interrupt_handler_nl, sh);
2932 	if (sh->intr_handle_nl == NULL) {
2933 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2934 		return;
2935 	}
2936 	if (sh->cdev->config.devx) {
2937 #ifdef HAVE_IBV_DEVX_ASYNC
2938 		struct mlx5dv_devx_cmd_comp *devx_comp;
2939 
2940 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2941 		devx_comp = sh->devx_comp;
2942 		if (!devx_comp) {
2943 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2944 			return;
2945 		}
2946 		sh->intr_handle_devx = mlx5_os_interrupt_handler_create
2947 			(RTE_INTR_INSTANCE_F_SHARED, true,
2948 			 devx_comp->fd,
2949 			 mlx5_dev_interrupt_handler_devx, sh);
2950 		if (!sh->intr_handle_devx) {
2951 			DRV_LOG(ERR, "Failed to allocate intr_handle.");
2952 			return;
2953 		}
2954 #endif /* HAVE_IBV_DEVX_ASYNC */
2955 	}
2956 }
2957 
2958 /**
2959  * Uninstall shared asynchronous device events handler.
2960  * This function is implemented to support event sharing
2961  * between multiple ports of single IB device.
2962  *
2963  * @param dev
2964  *   Pointer to mlx5_dev_ctx_shared object.
2965  */
2966 void
2967 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2968 {
2969 	mlx5_os_interrupt_handler_destroy(sh->intr_handle,
2970 					  mlx5_dev_interrupt_handler, sh);
2971 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_nl,
2972 					  mlx5_dev_interrupt_handler_nl, sh);
2973 #ifdef HAVE_IBV_DEVX_ASYNC
2974 	mlx5_os_interrupt_handler_destroy(sh->intr_handle_devx,
2975 					  mlx5_dev_interrupt_handler_devx, sh);
2976 	if (sh->devx_comp)
2977 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2978 #endif
2979 }
2980 
2981 /**
2982  * Read statistics by a named counter.
2983  *
2984  * @param[in] priv
2985  *   Pointer to the private device data structure.
2986  * @param[in] ctr_name
2987  *   Pointer to the name of the statistic counter to read
2988  * @param[out] stat
2989  *   Pointer to read statistic value.
2990  * @return
2991  *   0 on success and stat is valud, 1 if failed to read the value
2992  *   rte_errno is set.
2993  *
2994  */
2995 int
2996 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2997 		      uint64_t *stat)
2998 {
2999 	int fd;
3000 
3001 	if (priv->sh) {
3002 		if (priv->q_counters != NULL &&
3003 		    strcmp(ctr_name, "out_of_buffer") == 0) {
3004 			if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
3005 				DRV_LOG(WARNING, "Devx out_of_buffer counter is not supported in the secondary process");
3006 				rte_errno = ENOTSUP;
3007 				return 1;
3008 			}
3009 			return mlx5_devx_cmd_queue_counter_query
3010 					(priv->q_counters, 0, (uint32_t *)stat);
3011 		}
3012 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
3013 		      priv->sh->ibdev_path,
3014 		      priv->dev_port,
3015 		      ctr_name);
3016 		fd = open(path, O_RDONLY);
3017 		/*
3018 		 * in switchdev the file location is not per port
3019 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
3020 		 */
3021 		if (fd == -1) {
3022 			MKSTR(path1, "%s/hw_counters/%s",
3023 			      priv->sh->ibdev_path,
3024 			      ctr_name);
3025 			fd = open(path1, O_RDONLY);
3026 		}
3027 		if (fd != -1) {
3028 			char buf[21] = {'\0'};
3029 			ssize_t n = read(fd, buf, sizeof(buf));
3030 
3031 			close(fd);
3032 			if (n != -1) {
3033 				*stat = strtoull(buf, NULL, 10);
3034 				return 0;
3035 			}
3036 		}
3037 	}
3038 	*stat = 0;
3039 	return 1;
3040 }
3041 
3042 /**
3043  * Remove a MAC address from device
3044  *
3045  * @param dev
3046  *   Pointer to Ethernet device structure.
3047  * @param index
3048  *   MAC address index.
3049  */
3050 void
3051 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3052 {
3053 	struct mlx5_priv *priv = dev->data->dev_private;
3054 	const int vf = priv->sh->dev_cap.vf;
3055 
3056 	if (vf)
3057 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
3058 					mlx5_ifindex(dev), priv->mac_own,
3059 					&dev->data->mac_addrs[index], index);
3060 }
3061 
3062 /**
3063  * Adds a MAC address to the device
3064  *
3065  * @param dev
3066  *   Pointer to Ethernet device structure.
3067  * @param mac_addr
3068  *   MAC address to register.
3069  * @param index
3070  *   MAC address index.
3071  *
3072  * @return
3073  *   0 on success, a negative errno value otherwise
3074  */
3075 int
3076 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
3077 		     uint32_t index)
3078 {
3079 	struct mlx5_priv *priv = dev->data->dev_private;
3080 	const int vf = priv->sh->dev_cap.vf;
3081 	int ret = 0;
3082 
3083 	if (vf)
3084 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
3085 					   mlx5_ifindex(dev), priv->mac_own,
3086 					   mac, index);
3087 	return ret;
3088 }
3089 
3090 /**
3091  * Modify a VF MAC address
3092  *
3093  * @param priv
3094  *   Pointer to device private data.
3095  * @param mac_addr
3096  *   MAC address to modify into.
3097  * @param iface_idx
3098  *   Net device interface index
3099  * @param vf_index
3100  *   VF index
3101  *
3102  * @return
3103  *   0 on success, a negative errno value otherwise
3104  */
3105 int
3106 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
3107 			   unsigned int iface_idx,
3108 			   struct rte_ether_addr *mac_addr,
3109 			   int vf_index)
3110 {
3111 	return mlx5_nl_vf_mac_addr_modify
3112 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
3113 }
3114 
3115 /**
3116  * Set device promiscuous mode
3117  *
3118  * @param dev
3119  *   Pointer to Ethernet device structure.
3120  * @param enable
3121  *   0 - promiscuous is disabled, otherwise - enabled
3122  *
3123  * @return
3124  *   0 on success, a negative error value otherwise
3125  */
3126 int
3127 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
3128 {
3129 	struct mlx5_priv *priv = dev->data->dev_private;
3130 
3131 	return mlx5_nl_promisc(priv->nl_socket_route,
3132 			       mlx5_ifindex(dev), !!enable);
3133 }
3134 
3135 /**
3136  * Set device promiscuous mode
3137  *
3138  * @param dev
3139  *   Pointer to Ethernet device structure.
3140  * @param enable
3141  *   0 - all multicase is disabled, otherwise - enabled
3142  *
3143  * @return
3144  *   0 on success, a negative error value otherwise
3145  */
3146 int
3147 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
3148 {
3149 	struct mlx5_priv *priv = dev->data->dev_private;
3150 
3151 	return mlx5_nl_allmulti(priv->nl_socket_route,
3152 				mlx5_ifindex(dev), !!enable);
3153 }
3154 
3155 /**
3156  * Flush device MAC addresses
3157  *
3158  * @param dev
3159  *   Pointer to Ethernet device structure.
3160  *
3161  */
3162 void
3163 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
3164 {
3165 	struct mlx5_priv *priv = dev->data->dev_private;
3166 
3167 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
3168 			       dev->data->mac_addrs,
3169 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
3170 }
3171