xref: /dpdk/drivers/net/mlx5/mlx5.c (revision e11bdd37745229bf26b557305c07d118c3dbaad7)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 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 <sys/mman.h>
14 #include <linux/rtnetlink.h>
15 
16 /* Verbs header. */
17 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
18 #ifdef PEDANTIC
19 #pragma GCC diagnostic ignored "-Wpedantic"
20 #endif
21 #include <infiniband/verbs.h>
22 #ifdef PEDANTIC
23 #pragma GCC diagnostic error "-Wpedantic"
24 #endif
25 
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_ethdev_pci.h>
29 #include <rte_pci.h>
30 #include <rte_bus_pci.h>
31 #include <rte_common.h>
32 #include <rte_kvargs.h>
33 #include <rte_rwlock.h>
34 #include <rte_spinlock.h>
35 #include <rte_string_fns.h>
36 #include <rte_alarm.h>
37 
38 #include <mlx5_glue.h>
39 #include <mlx5_devx_cmds.h>
40 #include <mlx5_common.h>
41 #include <mlx5_common_mp.h>
42 
43 #include "mlx5_defs.h"
44 #include "mlx5.h"
45 #include "mlx5_utils.h"
46 #include "mlx5_rxtx.h"
47 #include "mlx5_autoconf.h"
48 #include "mlx5_mr.h"
49 #include "mlx5_flow.h"
50 #include "rte_pmd_mlx5.h"
51 
52 /* Device parameter to enable RX completion queue compression. */
53 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
54 
55 /* Device parameter to enable RX completion entry padding to 128B. */
56 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
57 
58 /* Device parameter to enable padding Rx packet to cacheline size. */
59 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
60 
61 /* Device parameter to enable Multi-Packet Rx queue. */
62 #define MLX5_RX_MPRQ_EN "mprq_en"
63 
64 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
65 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
66 
67 /* Device parameter to configure log 2 of the stride size for MPRQ. */
68 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
69 
70 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
71 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
72 
73 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
74 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
75 
76 /* Device parameter to configure inline send. Deprecated, ignored.*/
77 #define MLX5_TXQ_INLINE "txq_inline"
78 
79 /* Device parameter to limit packet size to inline with ordinary SEND. */
80 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
81 
82 /* Device parameter to configure minimal data size to inline. */
83 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
84 
85 /* Device parameter to limit packet size to inline with Enhanced MPW. */
86 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
87 
88 /*
89  * Device parameter to configure the number of TX queues threshold for
90  * enabling inline send.
91  */
92 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
93 
94 /*
95  * Device parameter to configure the number of TX queues threshold for
96  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
97  */
98 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
99 
100 /* Device parameter to enable multi-packet send WQEs. */
101 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
102 
103 /*
104  * Device parameter to force doorbell register mapping
105  * to non-cahed region eliminating the extra write memory barrier.
106  */
107 #define MLX5_TX_DB_NC "tx_db_nc"
108 
109 /*
110  * Device parameter to include 2 dsegs in the title WQEBB.
111  * Deprecated, ignored.
112  */
113 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
114 
115 /*
116  * Device parameter to limit the size of inlining packet.
117  * Deprecated, ignored.
118  */
119 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
120 
121 /*
122  * Device parameter to enable hardware Tx vector.
123  * Deprecated, ignored (no vectorized Tx routines anymore).
124  */
125 #define MLX5_TX_VEC_EN "tx_vec_en"
126 
127 /* Device parameter to enable hardware Rx vector. */
128 #define MLX5_RX_VEC_EN "rx_vec_en"
129 
130 /* Allow L3 VXLAN flow creation. */
131 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
132 
133 /* Activate DV E-Switch flow steering. */
134 #define MLX5_DV_ESW_EN "dv_esw_en"
135 
136 /* Activate DV flow steering. */
137 #define MLX5_DV_FLOW_EN "dv_flow_en"
138 
139 /* Enable extensive flow metadata support. */
140 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
141 
142 /* Activate Netlink support in VF mode. */
143 #define MLX5_VF_NL_EN "vf_nl_en"
144 
145 /* Enable extending memsegs when creating a MR. */
146 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
147 
148 /* Select port representors to instantiate. */
149 #define MLX5_REPRESENTOR "representor"
150 
151 /* Device parameter to configure the maximum number of dump files per queue. */
152 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
153 
154 /* Configure timeout of LRO session (in microseconds). */
155 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
156 
157 /*
158  * Device parameter to configure the total data buffer size for a single
159  * hairpin queue (logarithm value).
160  */
161 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
162 
163 #ifndef HAVE_IBV_MLX5_MOD_MPW
164 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
165 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
166 #endif
167 
168 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
169 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
170 #endif
171 
172 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
173 
174 /* Shared memory between primary and secondary processes. */
175 struct mlx5_shared_data *mlx5_shared_data;
176 
177 /* Spinlock for mlx5_shared_data allocation. */
178 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
179 
180 /* Process local data for secondary processes. */
181 static struct mlx5_local_data mlx5_local_data;
182 
183 /** Driver-specific log messages type. */
184 int mlx5_logtype;
185 
186 /** Data associated with devices to spawn. */
187 struct mlx5_dev_spawn_data {
188 	uint32_t ifindex; /**< Network interface index. */
189 	uint32_t max_port; /**< IB device maximal port index. */
190 	uint32_t ibv_port; /**< IB device physical port index. */
191 	int pf_bond; /**< bonding device PF index. < 0 - no bonding */
192 	struct mlx5_switch_info info; /**< Switch information. */
193 	struct ibv_device *ibv_dev; /**< Associated IB device. */
194 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
195 	struct rte_pci_device *pci_dev; /**< Backend PCI device. */
196 };
197 
198 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
199 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
200 
201 static struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
202 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
203 	{
204 		.size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
205 		.trunk_size = 64,
206 		.grow_trunk = 3,
207 		.grow_shift = 2,
208 		.need_lock = 0,
209 		.release_mem_en = 1,
210 		.malloc = rte_malloc_socket,
211 		.free = rte_free,
212 		.type = "mlx5_encap_decap_ipool",
213 	},
214 	{
215 		.size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
216 		.trunk_size = 64,
217 		.grow_trunk = 3,
218 		.grow_shift = 2,
219 		.need_lock = 0,
220 		.release_mem_en = 1,
221 		.malloc = rte_malloc_socket,
222 		.free = rte_free,
223 		.type = "mlx5_push_vlan_ipool",
224 	},
225 	{
226 		.size = sizeof(struct mlx5_flow_dv_tag_resource),
227 		.trunk_size = 64,
228 		.grow_trunk = 3,
229 		.grow_shift = 2,
230 		.need_lock = 0,
231 		.release_mem_en = 1,
232 		.malloc = rte_malloc_socket,
233 		.free = rte_free,
234 		.type = "mlx5_tag_ipool",
235 	},
236 	{
237 		.size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
238 		.trunk_size = 64,
239 		.grow_trunk = 3,
240 		.grow_shift = 2,
241 		.need_lock = 0,
242 		.release_mem_en = 1,
243 		.malloc = rte_malloc_socket,
244 		.free = rte_free,
245 		.type = "mlx5_port_id_ipool",
246 	},
247 	{
248 		.size = sizeof(struct mlx5_flow_tbl_data_entry),
249 		.trunk_size = 64,
250 		.grow_trunk = 3,
251 		.grow_shift = 2,
252 		.need_lock = 0,
253 		.release_mem_en = 1,
254 		.malloc = rte_malloc_socket,
255 		.free = rte_free,
256 		.type = "mlx5_jump_ipool",
257 	},
258 #endif
259 	{
260 		.size = sizeof(struct mlx5_flow_meter),
261 		.trunk_size = 64,
262 		.grow_trunk = 3,
263 		.grow_shift = 2,
264 		.need_lock = 0,
265 		.release_mem_en = 1,
266 		.malloc = rte_malloc_socket,
267 		.free = rte_free,
268 		.type = "mlx5_meter_ipool",
269 	},
270 	{
271 		.size = sizeof(struct mlx5_flow_mreg_copy_resource),
272 		.trunk_size = 64,
273 		.grow_trunk = 3,
274 		.grow_shift = 2,
275 		.need_lock = 0,
276 		.release_mem_en = 1,
277 		.malloc = rte_malloc_socket,
278 		.free = rte_free,
279 		.type = "mlx5_mcp_ipool",
280 	},
281 	{
282 		.size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
283 		.trunk_size = 64,
284 		.grow_trunk = 3,
285 		.grow_shift = 2,
286 		.need_lock = 0,
287 		.release_mem_en = 1,
288 		.malloc = rte_malloc_socket,
289 		.free = rte_free,
290 		.type = "mlx5_hrxq_ipool",
291 	},
292 	{
293 		.size = sizeof(struct mlx5_flow_handle),
294 		.trunk_size = 64,
295 		.grow_trunk = 3,
296 		.grow_shift = 2,
297 		.need_lock = 0,
298 		.release_mem_en = 1,
299 		.malloc = rte_malloc_socket,
300 		.free = rte_free,
301 		.type = "mlx5_flow_handle_ipool",
302 	},
303 	{
304 		.size = sizeof(struct rte_flow),
305 		.trunk_size = 4096,
306 		.need_lock = 1,
307 		.release_mem_en = 1,
308 		.malloc = rte_malloc_socket,
309 		.free = rte_free,
310 		.type = "rte_flow_ipool",
311 	},
312 };
313 
314 
315 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
316 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
317 
318 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
319 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
320 
321 /**
322  * Allocate ID pool structure.
323  *
324  * @param[in] max_id
325  *   The maximum id can be allocated from the pool.
326  *
327  * @return
328  *   Pointer to pool object, NULL value otherwise.
329  */
330 struct mlx5_flow_id_pool *
331 mlx5_flow_id_pool_alloc(uint32_t max_id)
332 {
333 	struct mlx5_flow_id_pool *pool;
334 	void *mem;
335 
336 	pool = rte_zmalloc("id pool allocation", sizeof(*pool),
337 			   RTE_CACHE_LINE_SIZE);
338 	if (!pool) {
339 		DRV_LOG(ERR, "can't allocate id pool");
340 		rte_errno  = ENOMEM;
341 		return NULL;
342 	}
343 	mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
344 			  RTE_CACHE_LINE_SIZE);
345 	if (!mem) {
346 		DRV_LOG(ERR, "can't allocate mem for id pool");
347 		rte_errno  = ENOMEM;
348 		goto error;
349 	}
350 	pool->free_arr = mem;
351 	pool->curr = pool->free_arr;
352 	pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
353 	pool->base_index = 0;
354 	pool->max_id = max_id;
355 	return pool;
356 error:
357 	rte_free(pool);
358 	return NULL;
359 }
360 
361 /**
362  * Release ID pool structure.
363  *
364  * @param[in] pool
365  *   Pointer to flow id pool object to free.
366  */
367 void
368 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
369 {
370 	rte_free(pool->free_arr);
371 	rte_free(pool);
372 }
373 
374 /**
375  * Generate ID.
376  *
377  * @param[in] pool
378  *   Pointer to flow id pool.
379  * @param[out] id
380  *   The generated ID.
381  *
382  * @return
383  *   0 on success, error value otherwise.
384  */
385 uint32_t
386 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
387 {
388 	if (pool->curr == pool->free_arr) {
389 		if (pool->base_index == pool->max_id) {
390 			rte_errno  = ENOMEM;
391 			DRV_LOG(ERR, "no free id");
392 			return -rte_errno;
393 		}
394 		*id = ++pool->base_index;
395 		return 0;
396 	}
397 	*id = *(--pool->curr);
398 	return 0;
399 }
400 
401 /**
402  * Release ID.
403  *
404  * @param[in] pool
405  *   Pointer to flow id pool.
406  * @param[out] id
407  *   The generated ID.
408  *
409  * @return
410  *   0 on success, error value otherwise.
411  */
412 uint32_t
413 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
414 {
415 	uint32_t size;
416 	uint32_t size2;
417 	void *mem;
418 
419 	if (pool->curr == pool->last) {
420 		size = pool->curr - pool->free_arr;
421 		size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
422 		MLX5_ASSERT(size2 > size);
423 		mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
424 		if (!mem) {
425 			DRV_LOG(ERR, "can't allocate mem for id pool");
426 			rte_errno  = ENOMEM;
427 			return -rte_errno;
428 		}
429 		memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
430 		rte_free(pool->free_arr);
431 		pool->free_arr = mem;
432 		pool->curr = pool->free_arr + size;
433 		pool->last = pool->free_arr + size2;
434 	}
435 	*pool->curr = id;
436 	pool->curr++;
437 	return 0;
438 }
439 
440 /**
441  * Initialize the shared aging list information per port.
442  *
443  * @param[in] sh
444  *   Pointer to mlx5_ibv_shared object.
445  */
446 static void
447 mlx5_flow_aging_init(struct mlx5_ibv_shared *sh)
448 {
449 	uint32_t i;
450 	struct mlx5_age_info *age_info;
451 
452 	for (i = 0; i < sh->max_port; i++) {
453 		age_info = &sh->port[i].age_info;
454 		age_info->flags = 0;
455 		TAILQ_INIT(&age_info->aged_counters);
456 		rte_spinlock_init(&age_info->aged_sl);
457 		MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
458 	}
459 }
460 
461 /**
462  * Initialize the counters management structure.
463  *
464  * @param[in] sh
465  *   Pointer to mlx5_ibv_shared object to free
466  */
467 static void
468 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
469 {
470 	uint8_t i, age;
471 
472 	sh->cmng.age = 0;
473 	TAILQ_INIT(&sh->cmng.flow_counters);
474 	for (age = 0; age < RTE_DIM(sh->cmng.ccont[0]); ++age) {
475 		for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
476 			TAILQ_INIT(&sh->cmng.ccont[i][age].pool_list);
477 	}
478 }
479 
480 /**
481  * Destroy all the resources allocated for a counter memory management.
482  *
483  * @param[in] mng
484  *   Pointer to the memory management structure.
485  */
486 static void
487 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
488 {
489 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
490 
491 	LIST_REMOVE(mng, next);
492 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
493 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
494 	rte_free(mem);
495 }
496 
497 /**
498  * Close and release all the resources of the counters management.
499  *
500  * @param[in] sh
501  *   Pointer to mlx5_ibv_shared object to free.
502  */
503 static void
504 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
505 {
506 	struct mlx5_counter_stats_mem_mng *mng;
507 	uint8_t i, age = 0;
508 	int j;
509 	int retries = 1024;
510 
511 	rte_errno = 0;
512 	while (--retries) {
513 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
514 		if (rte_errno != EINPROGRESS)
515 			break;
516 		rte_pause();
517 	}
518 
519 	for (age = 0; age < RTE_DIM(sh->cmng.ccont[0]); ++age) {
520 		for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
521 			struct mlx5_flow_counter_pool *pool;
522 			uint32_t batch = !!(i % 2);
523 
524 			if (!sh->cmng.ccont[i][age].pools)
525 				continue;
526 			pool = TAILQ_FIRST(&sh->cmng.ccont[i][age].pool_list);
527 			while (pool) {
528 				if (batch) {
529 					if (pool->min_dcs)
530 						claim_zero
531 						(mlx5_devx_cmd_destroy
532 						(pool->min_dcs));
533 				}
534 				for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
535 					if (MLX5_POOL_GET_CNT(pool, j)->action)
536 						claim_zero
537 						(mlx5_glue->destroy_flow_action
538 						 (MLX5_POOL_GET_CNT
539 						  (pool, j)->action));
540 					if (!batch && MLX5_GET_POOL_CNT_EXT
541 					    (pool, j)->dcs)
542 						claim_zero(mlx5_devx_cmd_destroy
543 							  (MLX5_GET_POOL_CNT_EXT
544 							  (pool, j)->dcs));
545 				}
546 				TAILQ_REMOVE(&sh->cmng.ccont[i][age].pool_list,
547 					pool, next);
548 				rte_free(pool);
549 				pool = TAILQ_FIRST
550 					(&sh->cmng.ccont[i][age].pool_list);
551 			}
552 			rte_free(sh->cmng.ccont[i][age].pools);
553 		}
554 	}
555 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
556 	while (mng) {
557 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
558 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
559 	}
560 	memset(&sh->cmng, 0, sizeof(sh->cmng));
561 }
562 
563 /**
564  * Initialize the flow resources' indexed mempool.
565  *
566  * @param[in] sh
567  *   Pointer to mlx5_ibv_shared object.
568  * @param[in] sh
569  *   Pointer to user dev config.
570  */
571 static void
572 mlx5_flow_ipool_create(struct mlx5_ibv_shared *sh,
573 		       const struct mlx5_dev_config *config __rte_unused)
574 {
575 	uint8_t i;
576 
577 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
578 	/*
579 	 * While DV is supported, user chooses the verbs mode,
580 	 * the mlx5 flow handle size is different with the
581 	 * MLX5_FLOW_HANDLE_VERBS_SIZE.
582 	 */
583 	if (!config->dv_flow_en)
584 		mlx5_ipool_cfg[MLX5_IPOOL_MLX5_FLOW].size =
585 					MLX5_FLOW_HANDLE_VERBS_SIZE;
586 #endif
587 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
588 		sh->ipool[i] = mlx5_ipool_create(&mlx5_ipool_cfg[i]);
589 }
590 
591 /**
592  * Release the flow resources' indexed mempool.
593  *
594  * @param[in] sh
595  *   Pointer to mlx5_ibv_shared object.
596  */
597 static void
598 mlx5_flow_ipool_destroy(struct mlx5_ibv_shared *sh)
599 {
600 	uint8_t i;
601 
602 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
603 		mlx5_ipool_destroy(sh->ipool[i]);
604 }
605 
606 /**
607  * Extract pdn of PD object using DV API.
608  *
609  * @param[in] pd
610  *   Pointer to the verbs PD object.
611  * @param[out] pdn
612  *   Pointer to the PD object number variable.
613  *
614  * @return
615  *   0 on success, error value otherwise.
616  */
617 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
618 static int
619 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
620 {
621 	struct mlx5dv_obj obj;
622 	struct mlx5dv_pd pd_info;
623 	int ret = 0;
624 
625 	obj.pd.in = pd;
626 	obj.pd.out = &pd_info;
627 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
628 	if (ret) {
629 		DRV_LOG(DEBUG, "Fail to get PD object info");
630 		return ret;
631 	}
632 	*pdn = pd_info.pdn;
633 	return 0;
634 }
635 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
636 
637 static int
638 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
639 {
640 	char *env;
641 	int value;
642 
643 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
644 	/* Get environment variable to store. */
645 	env = getenv(MLX5_SHUT_UP_BF);
646 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
647 	if (config->dbnc == MLX5_ARG_UNSET)
648 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
649 	else
650 		setenv(MLX5_SHUT_UP_BF,
651 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
652 	return value;
653 }
654 
655 static void
656 mlx5_restore_doorbell_mapping_env(int value)
657 {
658 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
659 	/* Restore the original environment variable state. */
660 	if (value == MLX5_ARG_UNSET)
661 		unsetenv(MLX5_SHUT_UP_BF);
662 	else
663 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
664 }
665 
666 /**
667  * Allocate shared IB device context. If there is multiport device the
668  * master and representors will share this context, if there is single
669  * port dedicated IB device, the context will be used by only given
670  * port due to unification.
671  *
672  * Routine first searches the context for the specified IB device name,
673  * if found the shared context assumed and reference counter is incremented.
674  * If no context found the new one is created and initialized with specified
675  * IB device context and parameters.
676  *
677  * @param[in] spawn
678  *   Pointer to the IB device attributes (name, port, etc).
679  * @param[in] config
680  *   Pointer to device configuration structure.
681  *
682  * @return
683  *   Pointer to mlx5_ibv_shared object on success,
684  *   otherwise NULL and rte_errno is set.
685  */
686 static struct mlx5_ibv_shared *
687 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
688 			const struct mlx5_dev_config *config)
689 {
690 	struct mlx5_ibv_shared *sh;
691 	int dbmap_env;
692 	int err = 0;
693 	uint32_t i;
694 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
695 	struct mlx5_devx_tis_attr tis_attr = { 0 };
696 #endif
697 
698 	MLX5_ASSERT(spawn);
699 	/* Secondary process should not create the shared context. */
700 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
701 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
702 	/* Search for IB context by device name. */
703 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
704 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
705 			sh->refcnt++;
706 			goto exit;
707 		}
708 	}
709 	/* No device found, we have to create new shared context. */
710 	MLX5_ASSERT(spawn->max_port);
711 	sh = rte_zmalloc("ethdev shared ib context",
712 			 sizeof(struct mlx5_ibv_shared) +
713 			 spawn->max_port *
714 			 sizeof(struct mlx5_ibv_shared_port),
715 			 RTE_CACHE_LINE_SIZE);
716 	if (!sh) {
717 		DRV_LOG(ERR, "shared context allocation failure");
718 		rte_errno  = ENOMEM;
719 		goto exit;
720 	}
721 	/*
722 	 * Configure environment variable "MLX5_BF_SHUT_UP"
723 	 * before the device creation. The rdma_core library
724 	 * checks the variable at device creation and
725 	 * stores the result internally.
726 	 */
727 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
728 	/* Try to open IB device with DV first, then usual Verbs. */
729 	errno = 0;
730 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
731 	if (sh->ctx) {
732 		sh->devx = 1;
733 		DRV_LOG(DEBUG, "DevX is supported");
734 		/* The device is created, no need for environment. */
735 		mlx5_restore_doorbell_mapping_env(dbmap_env);
736 	} else {
737 		/* The environment variable is still configured. */
738 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
739 		err = errno ? errno : ENODEV;
740 		/*
741 		 * The environment variable is not needed anymore,
742 		 * all device creation attempts are completed.
743 		 */
744 		mlx5_restore_doorbell_mapping_env(dbmap_env);
745 		if (!sh->ctx)
746 			goto error;
747 		DRV_LOG(DEBUG, "DevX is NOT supported");
748 	}
749 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
750 	if (err) {
751 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
752 		goto error;
753 	}
754 	sh->refcnt = 1;
755 	sh->max_port = spawn->max_port;
756 	strncpy(sh->ibdev_name, sh->ctx->device->name,
757 		sizeof(sh->ibdev_name));
758 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
759 		sizeof(sh->ibdev_path));
760 	pthread_mutex_init(&sh->intr_mutex, NULL);
761 	/*
762 	 * Setting port_id to max unallowed value means
763 	 * there is no interrupt subhandler installed for
764 	 * the given port index i.
765 	 */
766 	for (i = 0; i < sh->max_port; i++) {
767 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
768 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
769 	}
770 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
771 	if (sh->pd == NULL) {
772 		DRV_LOG(ERR, "PD allocation failure");
773 		err = ENOMEM;
774 		goto error;
775 	}
776 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
777 	if (sh->devx) {
778 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
779 		if (err) {
780 			DRV_LOG(ERR, "Fail to extract pdn from PD");
781 			goto error;
782 		}
783 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
784 		if (!sh->td) {
785 			DRV_LOG(ERR, "TD allocation failure");
786 			err = ENOMEM;
787 			goto error;
788 		}
789 		tis_attr.transport_domain = sh->td->id;
790 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
791 		if (!sh->tis) {
792 			DRV_LOG(ERR, "TIS allocation failure");
793 			err = ENOMEM;
794 			goto error;
795 		}
796 	}
797 	sh->flow_id_pool = mlx5_flow_id_pool_alloc
798 					((1 << HAIRPIN_FLOW_ID_BITS) - 1);
799 	if (!sh->flow_id_pool) {
800 		DRV_LOG(ERR, "can't create flow id pool");
801 		err = ENOMEM;
802 		goto error;
803 	}
804 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
805 	/*
806 	 * Once the device is added to the list of memory event
807 	 * callback, its global MR cache table cannot be expanded
808 	 * on the fly because of deadlock. If it overflows, lookup
809 	 * should be done by searching MR list linearly, which is slow.
810 	 *
811 	 * At this point the device is not added to the memory
812 	 * event list yet, context is just being created.
813 	 */
814 	err = mlx5_mr_btree_init(&sh->share_cache.cache,
815 				 MLX5_MR_BTREE_CACHE_N * 2,
816 				 spawn->pci_dev->device.numa_node);
817 	if (err) {
818 		err = rte_errno;
819 		goto error;
820 	}
821 	mlx5_flow_aging_init(sh);
822 	mlx5_flow_counters_mng_init(sh);
823 	mlx5_flow_ipool_create(sh, config);
824 	/* Add device to memory callback list. */
825 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
826 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
827 			 sh, mem_event_cb);
828 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
829 	/* Add context to the global device list. */
830 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
831 exit:
832 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
833 	return sh;
834 error:
835 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
836 	MLX5_ASSERT(sh);
837 	if (sh->tis)
838 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
839 	if (sh->td)
840 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
841 	if (sh->pd)
842 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
843 	if (sh->ctx)
844 		claim_zero(mlx5_glue->close_device(sh->ctx));
845 	if (sh->flow_id_pool)
846 		mlx5_flow_id_pool_release(sh->flow_id_pool);
847 	rte_free(sh);
848 	MLX5_ASSERT(err > 0);
849 	rte_errno = err;
850 	return NULL;
851 }
852 
853 /**
854  * Free shared IB device context. Decrement counter and if zero free
855  * all allocated resources and close handles.
856  *
857  * @param[in] sh
858  *   Pointer to mlx5_ibv_shared object to free
859  */
860 static void
861 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
862 {
863 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
864 #ifdef RTE_LIBRTE_MLX5_DEBUG
865 	/* Check the object presence in the list. */
866 	struct mlx5_ibv_shared *lctx;
867 
868 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
869 		if (lctx == sh)
870 			break;
871 	MLX5_ASSERT(lctx);
872 	if (lctx != sh) {
873 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
874 		goto exit;
875 	}
876 #endif
877 	MLX5_ASSERT(sh);
878 	MLX5_ASSERT(sh->refcnt);
879 	/* Secondary process should not free the shared context. */
880 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
881 	if (--sh->refcnt)
882 		goto exit;
883 	/* Remove from memory callback device list. */
884 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
885 	LIST_REMOVE(sh, mem_event_cb);
886 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
887 	/* Release created Memory Regions. */
888 	mlx5_mr_release_cache(&sh->share_cache);
889 	/* Remove context from the global device list. */
890 	LIST_REMOVE(sh, next);
891 	/*
892 	 *  Ensure there is no async event handler installed.
893 	 *  Only primary process handles async device events.
894 	 **/
895 	mlx5_flow_counters_mng_close(sh);
896 	mlx5_flow_ipool_destroy(sh);
897 	MLX5_ASSERT(!sh->intr_cnt);
898 	if (sh->intr_cnt)
899 		mlx5_intr_callback_unregister
900 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
901 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
902 	if (sh->devx_intr_cnt) {
903 		if (sh->intr_handle_devx.fd)
904 			rte_intr_callback_unregister(&sh->intr_handle_devx,
905 					  mlx5_dev_interrupt_handler_devx, sh);
906 		if (sh->devx_comp)
907 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
908 	}
909 #endif
910 	pthread_mutex_destroy(&sh->intr_mutex);
911 	if (sh->pd)
912 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
913 	if (sh->tis)
914 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
915 	if (sh->td)
916 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
917 	if (sh->ctx)
918 		claim_zero(mlx5_glue->close_device(sh->ctx));
919 	if (sh->flow_id_pool)
920 		mlx5_flow_id_pool_release(sh->flow_id_pool);
921 	rte_free(sh);
922 exit:
923 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
924 }
925 
926 /**
927  * Destroy table hash list and all the root entries per domain.
928  *
929  * @param[in] priv
930  *   Pointer to the private device data structure.
931  */
932 static void
933 mlx5_free_table_hash_list(struct mlx5_priv *priv)
934 {
935 	struct mlx5_ibv_shared *sh = priv->sh;
936 	struct mlx5_flow_tbl_data_entry *tbl_data;
937 	union mlx5_flow_tbl_key table_key = {
938 		{
939 			.table_id = 0,
940 			.reserved = 0,
941 			.domain = 0,
942 			.direction = 0,
943 		}
944 	};
945 	struct mlx5_hlist_entry *pos;
946 
947 	if (!sh->flow_tbls)
948 		return;
949 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
950 	if (pos) {
951 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
952 					entry);
953 		MLX5_ASSERT(tbl_data);
954 		mlx5_hlist_remove(sh->flow_tbls, pos);
955 		rte_free(tbl_data);
956 	}
957 	table_key.direction = 1;
958 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
959 	if (pos) {
960 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
961 					entry);
962 		MLX5_ASSERT(tbl_data);
963 		mlx5_hlist_remove(sh->flow_tbls, pos);
964 		rte_free(tbl_data);
965 	}
966 	table_key.direction = 0;
967 	table_key.domain = 1;
968 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
969 	if (pos) {
970 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
971 					entry);
972 		MLX5_ASSERT(tbl_data);
973 		mlx5_hlist_remove(sh->flow_tbls, pos);
974 		rte_free(tbl_data);
975 	}
976 	mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
977 }
978 
979 /**
980  * Initialize flow table hash list and create the root tables entry
981  * for each domain.
982  *
983  * @param[in] priv
984  *   Pointer to the private device data structure.
985  *
986  * @return
987  *   Zero on success, positive error code otherwise.
988  */
989 static int
990 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
991 {
992 	struct mlx5_ibv_shared *sh = priv->sh;
993 	char s[MLX5_HLIST_NAMESIZE];
994 	int err = 0;
995 
996 	MLX5_ASSERT(sh);
997 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
998 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
999 	if (!sh->flow_tbls) {
1000 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
1001 		err = ENOMEM;
1002 		return err;
1003 	}
1004 #ifndef HAVE_MLX5DV_DR
1005 	/*
1006 	 * In case we have not DR support, the zero tables should be created
1007 	 * because DV expect to see them even if they cannot be created by
1008 	 * RDMA-CORE.
1009 	 */
1010 	union mlx5_flow_tbl_key table_key = {
1011 		{
1012 			.table_id = 0,
1013 			.reserved = 0,
1014 			.domain = 0,
1015 			.direction = 0,
1016 		}
1017 	};
1018 	struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
1019 							  sizeof(*tbl_data), 0);
1020 
1021 	if (!tbl_data) {
1022 		err = ENOMEM;
1023 		goto error;
1024 	}
1025 	tbl_data->entry.key = table_key.v64;
1026 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1027 	if (err)
1028 		goto error;
1029 	rte_atomic32_init(&tbl_data->tbl.refcnt);
1030 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
1031 	table_key.direction = 1;
1032 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
1033 	if (!tbl_data) {
1034 		err = ENOMEM;
1035 		goto error;
1036 	}
1037 	tbl_data->entry.key = table_key.v64;
1038 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1039 	if (err)
1040 		goto error;
1041 	rte_atomic32_init(&tbl_data->tbl.refcnt);
1042 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
1043 	table_key.direction = 0;
1044 	table_key.domain = 1;
1045 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
1046 	if (!tbl_data) {
1047 		err = ENOMEM;
1048 		goto error;
1049 	}
1050 	tbl_data->entry.key = table_key.v64;
1051 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1052 	if (err)
1053 		goto error;
1054 	rte_atomic32_init(&tbl_data->tbl.refcnt);
1055 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
1056 	return err;
1057 error:
1058 	mlx5_free_table_hash_list(priv);
1059 #endif /* HAVE_MLX5DV_DR */
1060 	return err;
1061 }
1062 
1063 /**
1064  * Initialize DR related data within private structure.
1065  * Routine checks the reference counter and does actual
1066  * resources creation/initialization only if counter is zero.
1067  *
1068  * @param[in] priv
1069  *   Pointer to the private device data structure.
1070  *
1071  * @return
1072  *   Zero on success, positive error code otherwise.
1073  */
1074 static int
1075 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
1076 {
1077 	struct mlx5_ibv_shared *sh = priv->sh;
1078 	char s[MLX5_HLIST_NAMESIZE];
1079 	int err = 0;
1080 
1081 	if (!sh->flow_tbls)
1082 		err = mlx5_alloc_table_hash_list(priv);
1083 	else
1084 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
1085 			(void *)sh->flow_tbls);
1086 	if (err)
1087 		return err;
1088 	/* Create tags hash list table. */
1089 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
1090 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
1091 	if (!sh->tag_table) {
1092 		DRV_LOG(ERR, "tags with hash creation failed.\n");
1093 		err = ENOMEM;
1094 		goto error;
1095 	}
1096 #ifdef HAVE_MLX5DV_DR
1097 	void *domain;
1098 
1099 	if (sh->dv_refcnt) {
1100 		/* Shared DV/DR structures is already initialized. */
1101 		sh->dv_refcnt++;
1102 		priv->dr_shared = 1;
1103 		return 0;
1104 	}
1105 	/* Reference counter is zero, we should initialize structures. */
1106 	domain = mlx5_glue->dr_create_domain(sh->ctx,
1107 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
1108 	if (!domain) {
1109 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
1110 		err = errno;
1111 		goto error;
1112 	}
1113 	sh->rx_domain = domain;
1114 	domain = mlx5_glue->dr_create_domain(sh->ctx,
1115 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
1116 	if (!domain) {
1117 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
1118 		err = errno;
1119 		goto error;
1120 	}
1121 	pthread_mutex_init(&sh->dv_mutex, NULL);
1122 	sh->tx_domain = domain;
1123 #ifdef HAVE_MLX5DV_DR_ESWITCH
1124 	if (priv->config.dv_esw_en) {
1125 		domain  = mlx5_glue->dr_create_domain
1126 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
1127 		if (!domain) {
1128 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
1129 			err = errno;
1130 			goto error;
1131 		}
1132 		sh->fdb_domain = domain;
1133 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
1134 	}
1135 #endif
1136 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
1137 #endif /* HAVE_MLX5DV_DR */
1138 	sh->dv_refcnt++;
1139 	priv->dr_shared = 1;
1140 	return 0;
1141 error:
1142 	/* Rollback the created objects. */
1143 	if (sh->rx_domain) {
1144 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
1145 		sh->rx_domain = NULL;
1146 	}
1147 	if (sh->tx_domain) {
1148 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
1149 		sh->tx_domain = NULL;
1150 	}
1151 	if (sh->fdb_domain) {
1152 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1153 		sh->fdb_domain = NULL;
1154 	}
1155 	if (sh->esw_drop_action) {
1156 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1157 		sh->esw_drop_action = NULL;
1158 	}
1159 	if (sh->pop_vlan_action) {
1160 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1161 		sh->pop_vlan_action = NULL;
1162 	}
1163 	if (sh->tag_table) {
1164 		/* tags should be destroyed with flow before. */
1165 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1166 		sh->tag_table = NULL;
1167 	}
1168 	mlx5_free_table_hash_list(priv);
1169 	return err;
1170 }
1171 
1172 /**
1173  * Destroy DR related data within private structure.
1174  *
1175  * @param[in] priv
1176  *   Pointer to the private device data structure.
1177  */
1178 static void
1179 mlx5_free_shared_dr(struct mlx5_priv *priv)
1180 {
1181 	struct mlx5_ibv_shared *sh;
1182 
1183 	if (!priv->dr_shared)
1184 		return;
1185 	priv->dr_shared = 0;
1186 	sh = priv->sh;
1187 	MLX5_ASSERT(sh);
1188 #ifdef HAVE_MLX5DV_DR
1189 	MLX5_ASSERT(sh->dv_refcnt);
1190 	if (sh->dv_refcnt && --sh->dv_refcnt)
1191 		return;
1192 	if (sh->rx_domain) {
1193 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
1194 		sh->rx_domain = NULL;
1195 	}
1196 	if (sh->tx_domain) {
1197 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
1198 		sh->tx_domain = NULL;
1199 	}
1200 #ifdef HAVE_MLX5DV_DR_ESWITCH
1201 	if (sh->fdb_domain) {
1202 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1203 		sh->fdb_domain = NULL;
1204 	}
1205 	if (sh->esw_drop_action) {
1206 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1207 		sh->esw_drop_action = NULL;
1208 	}
1209 #endif
1210 	if (sh->pop_vlan_action) {
1211 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1212 		sh->pop_vlan_action = NULL;
1213 	}
1214 	pthread_mutex_destroy(&sh->dv_mutex);
1215 #endif /* HAVE_MLX5DV_DR */
1216 	if (sh->tag_table) {
1217 		/* tags should be destroyed with flow before. */
1218 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1219 		sh->tag_table = NULL;
1220 	}
1221 	mlx5_free_table_hash_list(priv);
1222 }
1223 
1224 /**
1225  * Initialize shared data between primary and secondary process.
1226  *
1227  * A memzone is reserved by primary process and secondary processes attach to
1228  * the memzone.
1229  *
1230  * @return
1231  *   0 on success, a negative errno value otherwise and rte_errno is set.
1232  */
1233 static int
1234 mlx5_init_shared_data(void)
1235 {
1236 	const struct rte_memzone *mz;
1237 	int ret = 0;
1238 
1239 	rte_spinlock_lock(&mlx5_shared_data_lock);
1240 	if (mlx5_shared_data == NULL) {
1241 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1242 			/* Allocate shared memory. */
1243 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1244 						 sizeof(*mlx5_shared_data),
1245 						 SOCKET_ID_ANY, 0);
1246 			if (mz == NULL) {
1247 				DRV_LOG(ERR,
1248 					"Cannot allocate mlx5 shared data");
1249 				ret = -rte_errno;
1250 				goto error;
1251 			}
1252 			mlx5_shared_data = mz->addr;
1253 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1254 			rte_spinlock_init(&mlx5_shared_data->lock);
1255 		} else {
1256 			/* Lookup allocated shared memory. */
1257 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1258 			if (mz == NULL) {
1259 				DRV_LOG(ERR,
1260 					"Cannot attach mlx5 shared data");
1261 				ret = -rte_errno;
1262 				goto error;
1263 			}
1264 			mlx5_shared_data = mz->addr;
1265 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1266 		}
1267 	}
1268 error:
1269 	rte_spinlock_unlock(&mlx5_shared_data_lock);
1270 	return ret;
1271 }
1272 
1273 /**
1274  * Retrieve integer value from environment variable.
1275  *
1276  * @param[in] name
1277  *   Environment variable name.
1278  *
1279  * @return
1280  *   Integer value, 0 if the variable is not set.
1281  */
1282 int
1283 mlx5_getenv_int(const char *name)
1284 {
1285 	const char *val = getenv(name);
1286 
1287 	if (val == NULL)
1288 		return 0;
1289 	return atoi(val);
1290 }
1291 
1292 /**
1293  * Verbs callback to allocate a memory. This function should allocate the space
1294  * according to the size provided residing inside a huge page.
1295  * Please note that all allocation must respect the alignment from libmlx5
1296  * (i.e. currently sysconf(_SC_PAGESIZE)).
1297  *
1298  * @param[in] size
1299  *   The size in bytes of the memory to allocate.
1300  * @param[in] data
1301  *   A pointer to the callback data.
1302  *
1303  * @return
1304  *   Allocated buffer, NULL otherwise and rte_errno is set.
1305  */
1306 static void *
1307 mlx5_alloc_verbs_buf(size_t size, void *data)
1308 {
1309 	struct mlx5_priv *priv = data;
1310 	void *ret;
1311 	size_t alignment = sysconf(_SC_PAGESIZE);
1312 	unsigned int socket = SOCKET_ID_ANY;
1313 
1314 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1315 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1316 
1317 		socket = ctrl->socket;
1318 	} else if (priv->verbs_alloc_ctx.type ==
1319 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1320 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1321 
1322 		socket = ctrl->socket;
1323 	}
1324 	MLX5_ASSERT(data != NULL);
1325 	ret = rte_malloc_socket(__func__, size, alignment, socket);
1326 	if (!ret && size)
1327 		rte_errno = ENOMEM;
1328 	return ret;
1329 }
1330 
1331 /**
1332  * Verbs callback to free a memory.
1333  *
1334  * @param[in] ptr
1335  *   A pointer to the memory to free.
1336  * @param[in] data
1337  *   A pointer to the callback data.
1338  */
1339 static void
1340 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1341 {
1342 	MLX5_ASSERT(data != NULL);
1343 	rte_free(ptr);
1344 }
1345 
1346 /**
1347  * DPDK callback to add udp tunnel port
1348  *
1349  * @param[in] dev
1350  *   A pointer to eth_dev
1351  * @param[in] udp_tunnel
1352  *   A pointer to udp tunnel
1353  *
1354  * @return
1355  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1356  */
1357 int
1358 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1359 			 struct rte_eth_udp_tunnel *udp_tunnel)
1360 {
1361 	MLX5_ASSERT(udp_tunnel != NULL);
1362 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1363 	    udp_tunnel->udp_port == 4789)
1364 		return 0;
1365 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1366 	    udp_tunnel->udp_port == 4790)
1367 		return 0;
1368 	return -ENOTSUP;
1369 }
1370 
1371 /**
1372  * Initialize process private data structure.
1373  *
1374  * @param dev
1375  *   Pointer to Ethernet device structure.
1376  *
1377  * @return
1378  *   0 on success, a negative errno value otherwise and rte_errno is set.
1379  */
1380 int
1381 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1382 {
1383 	struct mlx5_priv *priv = dev->data->dev_private;
1384 	struct mlx5_proc_priv *ppriv;
1385 	size_t ppriv_size;
1386 
1387 	/*
1388 	 * UAR register table follows the process private structure. BlueFlame
1389 	 * registers for Tx queues are stored in the table.
1390 	 */
1391 	ppriv_size =
1392 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1393 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1394 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1395 	if (!ppriv) {
1396 		rte_errno = ENOMEM;
1397 		return -rte_errno;
1398 	}
1399 	ppriv->uar_table_sz = ppriv_size;
1400 	dev->process_private = ppriv;
1401 	return 0;
1402 }
1403 
1404 /**
1405  * Un-initialize process private data structure.
1406  *
1407  * @param dev
1408  *   Pointer to Ethernet device structure.
1409  */
1410 static void
1411 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1412 {
1413 	if (!dev->process_private)
1414 		return;
1415 	rte_free(dev->process_private);
1416 	dev->process_private = NULL;
1417 }
1418 
1419 /**
1420  * DPDK callback to close the device.
1421  *
1422  * Destroy all queues and objects, free memory.
1423  *
1424  * @param dev
1425  *   Pointer to Ethernet device structure.
1426  */
1427 static void
1428 mlx5_dev_close(struct rte_eth_dev *dev)
1429 {
1430 	struct mlx5_priv *priv = dev->data->dev_private;
1431 	unsigned int i;
1432 	int ret;
1433 
1434 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1435 		dev->data->port_id,
1436 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1437 	/* In case mlx5_dev_stop() has not been called. */
1438 	mlx5_dev_interrupt_handler_uninstall(dev);
1439 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1440 	/*
1441 	 * If default mreg copy action is removed at the stop stage,
1442 	 * the search will return none and nothing will be done anymore.
1443 	 */
1444 	mlx5_flow_stop_default(dev);
1445 	mlx5_traffic_disable(dev);
1446 	/*
1447 	 * If all the flows are already flushed in the device stop stage,
1448 	 * then this will return directly without any action.
1449 	 */
1450 	mlx5_flow_list_flush(dev, &priv->flows, true);
1451 	mlx5_flow_meter_flush(dev, NULL);
1452 	/* Free the intermediate buffers for flow creation. */
1453 	mlx5_flow_free_intermediate(dev);
1454 	/* Prevent crashes when queues are still in use. */
1455 	dev->rx_pkt_burst = removed_rx_burst;
1456 	dev->tx_pkt_burst = removed_tx_burst;
1457 	rte_wmb();
1458 	/* Disable datapath on secondary process. */
1459 	mlx5_mp_req_stop_rxtx(dev);
1460 	if (priv->rxqs != NULL) {
1461 		/* XXX race condition if mlx5_rx_burst() is still running. */
1462 		usleep(1000);
1463 		for (i = 0; (i != priv->rxqs_n); ++i)
1464 			mlx5_rxq_release(dev, i);
1465 		priv->rxqs_n = 0;
1466 		priv->rxqs = NULL;
1467 	}
1468 	if (priv->txqs != NULL) {
1469 		/* XXX race condition if mlx5_tx_burst() is still running. */
1470 		usleep(1000);
1471 		for (i = 0; (i != priv->txqs_n); ++i)
1472 			mlx5_txq_release(dev, i);
1473 		priv->txqs_n = 0;
1474 		priv->txqs = NULL;
1475 	}
1476 	mlx5_proc_priv_uninit(dev);
1477 	if (priv->mreg_cp_tbl)
1478 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1479 	mlx5_mprq_free_mp(dev);
1480 	mlx5_free_shared_dr(priv);
1481 	if (priv->rss_conf.rss_key != NULL)
1482 		rte_free(priv->rss_conf.rss_key);
1483 	if (priv->reta_idx != NULL)
1484 		rte_free(priv->reta_idx);
1485 	if (priv->config.vf)
1486 		mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1487 				       dev->data->mac_addrs,
1488 				       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1489 	if (priv->nl_socket_route >= 0)
1490 		close(priv->nl_socket_route);
1491 	if (priv->nl_socket_rdma >= 0)
1492 		close(priv->nl_socket_rdma);
1493 	if (priv->vmwa_context)
1494 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1495 	ret = mlx5_hrxq_verify(dev);
1496 	if (ret)
1497 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1498 			dev->data->port_id);
1499 	ret = mlx5_ind_table_obj_verify(dev);
1500 	if (ret)
1501 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1502 			dev->data->port_id);
1503 	ret = mlx5_rxq_obj_verify(dev);
1504 	if (ret)
1505 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1506 			dev->data->port_id);
1507 	ret = mlx5_rxq_verify(dev);
1508 	if (ret)
1509 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1510 			dev->data->port_id);
1511 	ret = mlx5_txq_obj_verify(dev);
1512 	if (ret)
1513 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1514 			dev->data->port_id);
1515 	ret = mlx5_txq_verify(dev);
1516 	if (ret)
1517 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1518 			dev->data->port_id);
1519 	ret = mlx5_flow_verify(dev);
1520 	if (ret)
1521 		DRV_LOG(WARNING, "port %u some flows still remain",
1522 			dev->data->port_id);
1523 	if (priv->sh) {
1524 		/*
1525 		 * Free the shared context in last turn, because the cleanup
1526 		 * routines above may use some shared fields, like
1527 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1528 		 * ifindex if Netlink fails.
1529 		 */
1530 		mlx5_free_shared_ibctx(priv->sh);
1531 		priv->sh = NULL;
1532 	}
1533 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1534 		unsigned int c = 0;
1535 		uint16_t port_id;
1536 
1537 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1538 			struct mlx5_priv *opriv =
1539 				rte_eth_devices[port_id].data->dev_private;
1540 
1541 			if (!opriv ||
1542 			    opriv->domain_id != priv->domain_id ||
1543 			    &rte_eth_devices[port_id] == dev)
1544 				continue;
1545 			++c;
1546 			break;
1547 		}
1548 		if (!c)
1549 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1550 	}
1551 	memset(priv, 0, sizeof(*priv));
1552 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1553 	/*
1554 	 * Reset mac_addrs to NULL such that it is not freed as part of
1555 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1556 	 * it is freed when dev_private is freed.
1557 	 */
1558 	dev->data->mac_addrs = NULL;
1559 }
1560 
1561 const struct eth_dev_ops mlx5_dev_ops = {
1562 	.dev_configure = mlx5_dev_configure,
1563 	.dev_start = mlx5_dev_start,
1564 	.dev_stop = mlx5_dev_stop,
1565 	.dev_set_link_down = mlx5_set_link_down,
1566 	.dev_set_link_up = mlx5_set_link_up,
1567 	.dev_close = mlx5_dev_close,
1568 	.promiscuous_enable = mlx5_promiscuous_enable,
1569 	.promiscuous_disable = mlx5_promiscuous_disable,
1570 	.allmulticast_enable = mlx5_allmulticast_enable,
1571 	.allmulticast_disable = mlx5_allmulticast_disable,
1572 	.link_update = mlx5_link_update,
1573 	.stats_get = mlx5_stats_get,
1574 	.stats_reset = mlx5_stats_reset,
1575 	.xstats_get = mlx5_xstats_get,
1576 	.xstats_reset = mlx5_xstats_reset,
1577 	.xstats_get_names = mlx5_xstats_get_names,
1578 	.fw_version_get = mlx5_fw_version_get,
1579 	.dev_infos_get = mlx5_dev_infos_get,
1580 	.read_clock = mlx5_read_clock,
1581 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1582 	.vlan_filter_set = mlx5_vlan_filter_set,
1583 	.rx_queue_setup = mlx5_rx_queue_setup,
1584 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1585 	.tx_queue_setup = mlx5_tx_queue_setup,
1586 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1587 	.rx_queue_release = mlx5_rx_queue_release,
1588 	.tx_queue_release = mlx5_tx_queue_release,
1589 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1590 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1591 	.mac_addr_remove = mlx5_mac_addr_remove,
1592 	.mac_addr_add = mlx5_mac_addr_add,
1593 	.mac_addr_set = mlx5_mac_addr_set,
1594 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1595 	.mtu_set = mlx5_dev_set_mtu,
1596 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1597 	.vlan_offload_set = mlx5_vlan_offload_set,
1598 	.reta_update = mlx5_dev_rss_reta_update,
1599 	.reta_query = mlx5_dev_rss_reta_query,
1600 	.rss_hash_update = mlx5_rss_hash_update,
1601 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1602 	.filter_ctrl = mlx5_dev_filter_ctrl,
1603 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1604 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1605 	.rxq_info_get = mlx5_rxq_info_get,
1606 	.txq_info_get = mlx5_txq_info_get,
1607 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1608 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1609 	.rx_queue_count = mlx5_rx_queue_count,
1610 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1611 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1612 	.is_removed = mlx5_is_removed,
1613 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1614 	.get_module_info = mlx5_get_module_info,
1615 	.get_module_eeprom = mlx5_get_module_eeprom,
1616 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1617 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1618 };
1619 
1620 /* Available operations from secondary process. */
1621 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1622 	.stats_get = mlx5_stats_get,
1623 	.stats_reset = mlx5_stats_reset,
1624 	.xstats_get = mlx5_xstats_get,
1625 	.xstats_reset = mlx5_xstats_reset,
1626 	.xstats_get_names = mlx5_xstats_get_names,
1627 	.fw_version_get = mlx5_fw_version_get,
1628 	.dev_infos_get = mlx5_dev_infos_get,
1629 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1630 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1631 	.rxq_info_get = mlx5_rxq_info_get,
1632 	.txq_info_get = mlx5_txq_info_get,
1633 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1634 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1635 	.get_module_info = mlx5_get_module_info,
1636 	.get_module_eeprom = mlx5_get_module_eeprom,
1637 };
1638 
1639 /* Available operations in flow isolated mode. */
1640 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1641 	.dev_configure = mlx5_dev_configure,
1642 	.dev_start = mlx5_dev_start,
1643 	.dev_stop = mlx5_dev_stop,
1644 	.dev_set_link_down = mlx5_set_link_down,
1645 	.dev_set_link_up = mlx5_set_link_up,
1646 	.dev_close = mlx5_dev_close,
1647 	.promiscuous_enable = mlx5_promiscuous_enable,
1648 	.promiscuous_disable = mlx5_promiscuous_disable,
1649 	.allmulticast_enable = mlx5_allmulticast_enable,
1650 	.allmulticast_disable = mlx5_allmulticast_disable,
1651 	.link_update = mlx5_link_update,
1652 	.stats_get = mlx5_stats_get,
1653 	.stats_reset = mlx5_stats_reset,
1654 	.xstats_get = mlx5_xstats_get,
1655 	.xstats_reset = mlx5_xstats_reset,
1656 	.xstats_get_names = mlx5_xstats_get_names,
1657 	.fw_version_get = mlx5_fw_version_get,
1658 	.dev_infos_get = mlx5_dev_infos_get,
1659 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1660 	.vlan_filter_set = mlx5_vlan_filter_set,
1661 	.rx_queue_setup = mlx5_rx_queue_setup,
1662 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1663 	.tx_queue_setup = mlx5_tx_queue_setup,
1664 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1665 	.rx_queue_release = mlx5_rx_queue_release,
1666 	.tx_queue_release = mlx5_tx_queue_release,
1667 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1668 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1669 	.mac_addr_remove = mlx5_mac_addr_remove,
1670 	.mac_addr_add = mlx5_mac_addr_add,
1671 	.mac_addr_set = mlx5_mac_addr_set,
1672 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1673 	.mtu_set = mlx5_dev_set_mtu,
1674 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1675 	.vlan_offload_set = mlx5_vlan_offload_set,
1676 	.filter_ctrl = mlx5_dev_filter_ctrl,
1677 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1678 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1679 	.rxq_info_get = mlx5_rxq_info_get,
1680 	.txq_info_get = mlx5_txq_info_get,
1681 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1682 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1683 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1684 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1685 	.is_removed = mlx5_is_removed,
1686 	.get_module_info = mlx5_get_module_info,
1687 	.get_module_eeprom = mlx5_get_module_eeprom,
1688 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1689 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1690 };
1691 
1692 /**
1693  * Verify and store value for device argument.
1694  *
1695  * @param[in] key
1696  *   Key argument to verify.
1697  * @param[in] val
1698  *   Value associated with key.
1699  * @param opaque
1700  *   User data.
1701  *
1702  * @return
1703  *   0 on success, a negative errno value otherwise and rte_errno is set.
1704  */
1705 static int
1706 mlx5_args_check(const char *key, const char *val, void *opaque)
1707 {
1708 	struct mlx5_dev_config *config = opaque;
1709 	unsigned long tmp;
1710 
1711 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1712 	if (!strcmp(MLX5_REPRESENTOR, key))
1713 		return 0;
1714 	errno = 0;
1715 	tmp = strtoul(val, NULL, 0);
1716 	if (errno) {
1717 		rte_errno = errno;
1718 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1719 		return -rte_errno;
1720 	}
1721 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1722 		config->cqe_comp = !!tmp;
1723 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1724 		config->cqe_pad = !!tmp;
1725 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1726 		config->hw_padding = !!tmp;
1727 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1728 		config->mprq.enabled = !!tmp;
1729 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1730 		config->mprq.stride_num_n = tmp;
1731 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1732 		config->mprq.stride_size_n = tmp;
1733 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1734 		config->mprq.max_memcpy_len = tmp;
1735 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1736 		config->mprq.min_rxqs_num = tmp;
1737 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1738 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1739 				 " converted to txq_inline_max", key);
1740 		config->txq_inline_max = tmp;
1741 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1742 		config->txq_inline_max = tmp;
1743 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1744 		config->txq_inline_min = tmp;
1745 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1746 		config->txq_inline_mpw = tmp;
1747 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1748 		config->txqs_inline = tmp;
1749 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1750 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1751 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1752 		config->mps = !!tmp;
1753 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1754 		if (tmp != MLX5_TXDB_CACHED &&
1755 		    tmp != MLX5_TXDB_NCACHED &&
1756 		    tmp != MLX5_TXDB_HEURISTIC) {
1757 			DRV_LOG(ERR, "invalid Tx doorbell "
1758 				     "mapping parameter");
1759 			rte_errno = EINVAL;
1760 			return -rte_errno;
1761 		}
1762 		config->dbnc = tmp;
1763 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1764 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1765 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1766 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1767 				 " converted to txq_inline_mpw", key);
1768 		config->txq_inline_mpw = tmp;
1769 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1770 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1771 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1772 		config->rx_vec_en = !!tmp;
1773 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1774 		config->l3_vxlan_en = !!tmp;
1775 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1776 		config->vf_nl_en = !!tmp;
1777 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1778 		config->dv_esw_en = !!tmp;
1779 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1780 		config->dv_flow_en = !!tmp;
1781 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1782 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1783 		    tmp != MLX5_XMETA_MODE_META16 &&
1784 		    tmp != MLX5_XMETA_MODE_META32) {
1785 			DRV_LOG(ERR, "invalid extensive "
1786 				     "metadata parameter");
1787 			rte_errno = EINVAL;
1788 			return -rte_errno;
1789 		}
1790 		config->dv_xmeta_en = tmp;
1791 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1792 		config->mr_ext_memseg_en = !!tmp;
1793 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1794 		config->max_dump_files_num = tmp;
1795 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1796 		config->lro.timeout = tmp;
1797 	} else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1798 		DRV_LOG(DEBUG, "class argument is %s.", val);
1799 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1800 		config->log_hp_size = tmp;
1801 	} else {
1802 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1803 		rte_errno = EINVAL;
1804 		return -rte_errno;
1805 	}
1806 	return 0;
1807 }
1808 
1809 /**
1810  * Parse device parameters.
1811  *
1812  * @param config
1813  *   Pointer to device configuration structure.
1814  * @param devargs
1815  *   Device arguments structure.
1816  *
1817  * @return
1818  *   0 on success, a negative errno value otherwise and rte_errno is set.
1819  */
1820 static int
1821 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1822 {
1823 	const char **params = (const char *[]){
1824 		MLX5_RXQ_CQE_COMP_EN,
1825 		MLX5_RXQ_CQE_PAD_EN,
1826 		MLX5_RXQ_PKT_PAD_EN,
1827 		MLX5_RX_MPRQ_EN,
1828 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1829 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1830 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1831 		MLX5_RXQS_MIN_MPRQ,
1832 		MLX5_TXQ_INLINE,
1833 		MLX5_TXQ_INLINE_MIN,
1834 		MLX5_TXQ_INLINE_MAX,
1835 		MLX5_TXQ_INLINE_MPW,
1836 		MLX5_TXQS_MIN_INLINE,
1837 		MLX5_TXQS_MAX_VEC,
1838 		MLX5_TXQ_MPW_EN,
1839 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1840 		MLX5_TXQ_MAX_INLINE_LEN,
1841 		MLX5_TX_DB_NC,
1842 		MLX5_TX_VEC_EN,
1843 		MLX5_RX_VEC_EN,
1844 		MLX5_L3_VXLAN_EN,
1845 		MLX5_VF_NL_EN,
1846 		MLX5_DV_ESW_EN,
1847 		MLX5_DV_FLOW_EN,
1848 		MLX5_DV_XMETA_EN,
1849 		MLX5_MR_EXT_MEMSEG_EN,
1850 		MLX5_REPRESENTOR,
1851 		MLX5_MAX_DUMP_FILES_NUM,
1852 		MLX5_LRO_TIMEOUT_USEC,
1853 		MLX5_CLASS_ARG_NAME,
1854 		MLX5_HP_BUF_SIZE,
1855 		NULL,
1856 	};
1857 	struct rte_kvargs *kvlist;
1858 	int ret = 0;
1859 	int i;
1860 
1861 	if (devargs == NULL)
1862 		return 0;
1863 	/* Following UGLY cast is done to pass checkpatch. */
1864 	kvlist = rte_kvargs_parse(devargs->args, params);
1865 	if (kvlist == NULL) {
1866 		rte_errno = EINVAL;
1867 		return -rte_errno;
1868 	}
1869 	/* Process parameters. */
1870 	for (i = 0; (params[i] != NULL); ++i) {
1871 		if (rte_kvargs_count(kvlist, params[i])) {
1872 			ret = rte_kvargs_process(kvlist, params[i],
1873 						 mlx5_args_check, config);
1874 			if (ret) {
1875 				rte_errno = EINVAL;
1876 				rte_kvargs_free(kvlist);
1877 				return -rte_errno;
1878 			}
1879 		}
1880 	}
1881 	rte_kvargs_free(kvlist);
1882 	return 0;
1883 }
1884 
1885 static struct rte_pci_driver mlx5_driver;
1886 
1887 /**
1888  * PMD global initialization.
1889  *
1890  * Independent from individual device, this function initializes global
1891  * per-PMD data structures distinguishing primary and secondary processes.
1892  * Hence, each initialization is called once per a process.
1893  *
1894  * @return
1895  *   0 on success, a negative errno value otherwise and rte_errno is set.
1896  */
1897 static int
1898 mlx5_init_once(void)
1899 {
1900 	struct mlx5_shared_data *sd;
1901 	struct mlx5_local_data *ld = &mlx5_local_data;
1902 	int ret = 0;
1903 
1904 	if (mlx5_init_shared_data())
1905 		return -rte_errno;
1906 	sd = mlx5_shared_data;
1907 	MLX5_ASSERT(sd);
1908 	rte_spinlock_lock(&sd->lock);
1909 	switch (rte_eal_process_type()) {
1910 	case RTE_PROC_PRIMARY:
1911 		if (sd->init_done)
1912 			break;
1913 		LIST_INIT(&sd->mem_event_cb_list);
1914 		rte_rwlock_init(&sd->mem_event_rwlock);
1915 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1916 						mlx5_mr_mem_event_cb, NULL);
1917 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
1918 					   mlx5_mp_primary_handle);
1919 		if (ret)
1920 			goto out;
1921 		sd->init_done = true;
1922 		break;
1923 	case RTE_PROC_SECONDARY:
1924 		if (ld->init_done)
1925 			break;
1926 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
1927 					     mlx5_mp_secondary_handle);
1928 		if (ret)
1929 			goto out;
1930 		++sd->secondary_cnt;
1931 		ld->init_done = true;
1932 		break;
1933 	default:
1934 		break;
1935 	}
1936 out:
1937 	rte_spinlock_unlock(&sd->lock);
1938 	return ret;
1939 }
1940 
1941 /**
1942  * Configures the minimal amount of data to inline into WQE
1943  * while sending packets.
1944  *
1945  * - the txq_inline_min has the maximal priority, if this
1946  *   key is specified in devargs
1947  * - if DevX is enabled the inline mode is queried from the
1948  *   device (HCA attributes and NIC vport context if needed).
1949  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1950  *   and none (0 bytes) for other NICs
1951  *
1952  * @param spawn
1953  *   Verbs device parameters (name, port, switch_info) to spawn.
1954  * @param config
1955  *   Device configuration parameters.
1956  */
1957 static void
1958 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1959 		    struct mlx5_dev_config *config)
1960 {
1961 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1962 		/* Application defines size of inlined data explicitly. */
1963 		switch (spawn->pci_dev->id.device_id) {
1964 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1965 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1966 			if (config->txq_inline_min <
1967 				       (int)MLX5_INLINE_HSIZE_L2) {
1968 				DRV_LOG(DEBUG,
1969 					"txq_inline_mix aligned to minimal"
1970 					" ConnectX-4 required value %d",
1971 					(int)MLX5_INLINE_HSIZE_L2);
1972 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1973 			}
1974 			break;
1975 		}
1976 		goto exit;
1977 	}
1978 	if (config->hca_attr.eth_net_offloads) {
1979 		/* We have DevX enabled, inline mode queried successfully. */
1980 		switch (config->hca_attr.wqe_inline_mode) {
1981 		case MLX5_CAP_INLINE_MODE_L2:
1982 			/* outer L2 header must be inlined. */
1983 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1984 			goto exit;
1985 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1986 			/* No inline data are required by NIC. */
1987 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1988 			config->hw_vlan_insert =
1989 				config->hca_attr.wqe_vlan_insert;
1990 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1991 			goto exit;
1992 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1993 			/* inline mode is defined by NIC vport context. */
1994 			if (!config->hca_attr.eth_virt)
1995 				break;
1996 			switch (config->hca_attr.vport_inline_mode) {
1997 			case MLX5_INLINE_MODE_NONE:
1998 				config->txq_inline_min =
1999 					MLX5_INLINE_HSIZE_NONE;
2000 				goto exit;
2001 			case MLX5_INLINE_MODE_L2:
2002 				config->txq_inline_min =
2003 					MLX5_INLINE_HSIZE_L2;
2004 				goto exit;
2005 			case MLX5_INLINE_MODE_IP:
2006 				config->txq_inline_min =
2007 					MLX5_INLINE_HSIZE_L3;
2008 				goto exit;
2009 			case MLX5_INLINE_MODE_TCP_UDP:
2010 				config->txq_inline_min =
2011 					MLX5_INLINE_HSIZE_L4;
2012 				goto exit;
2013 			case MLX5_INLINE_MODE_INNER_L2:
2014 				config->txq_inline_min =
2015 					MLX5_INLINE_HSIZE_INNER_L2;
2016 				goto exit;
2017 			case MLX5_INLINE_MODE_INNER_IP:
2018 				config->txq_inline_min =
2019 					MLX5_INLINE_HSIZE_INNER_L3;
2020 				goto exit;
2021 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
2022 				config->txq_inline_min =
2023 					MLX5_INLINE_HSIZE_INNER_L4;
2024 				goto exit;
2025 			}
2026 		}
2027 	}
2028 	/*
2029 	 * We get here if we are unable to deduce
2030 	 * inline data size with DevX. Try PCI ID
2031 	 * to determine old NICs.
2032 	 */
2033 	switch (spawn->pci_dev->id.device_id) {
2034 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2035 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2036 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2037 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2038 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2039 		config->hw_vlan_insert = 0;
2040 		break;
2041 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2042 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2043 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2044 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2045 		/*
2046 		 * These NICs support VLAN insertion from WQE and
2047 		 * report the wqe_vlan_insert flag. But there is the bug
2048 		 * and PFC control may be broken, so disable feature.
2049 		 */
2050 		config->hw_vlan_insert = 0;
2051 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2052 		break;
2053 	default:
2054 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2055 		break;
2056 	}
2057 exit:
2058 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2059 }
2060 
2061 /**
2062  * Configures the metadata mask fields in the shared context.
2063  *
2064  * @param [in] dev
2065  *   Pointer to Ethernet device.
2066  */
2067 static void
2068 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2069 {
2070 	struct mlx5_priv *priv = dev->data->dev_private;
2071 	struct mlx5_ibv_shared *sh = priv->sh;
2072 	uint32_t meta, mark, reg_c0;
2073 
2074 	reg_c0 = ~priv->vport_meta_mask;
2075 	switch (priv->config.dv_xmeta_en) {
2076 	case MLX5_XMETA_MODE_LEGACY:
2077 		meta = UINT32_MAX;
2078 		mark = MLX5_FLOW_MARK_MASK;
2079 		break;
2080 	case MLX5_XMETA_MODE_META16:
2081 		meta = reg_c0 >> rte_bsf32(reg_c0);
2082 		mark = MLX5_FLOW_MARK_MASK;
2083 		break;
2084 	case MLX5_XMETA_MODE_META32:
2085 		meta = UINT32_MAX;
2086 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2087 		break;
2088 	default:
2089 		meta = 0;
2090 		mark = 0;
2091 		MLX5_ASSERT(false);
2092 		break;
2093 	}
2094 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2095 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2096 				 sh->dv_mark_mask, mark);
2097 	else
2098 		sh->dv_mark_mask = mark;
2099 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2100 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2101 				 sh->dv_meta_mask, meta);
2102 	else
2103 		sh->dv_meta_mask = meta;
2104 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2105 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2106 				 sh->dv_meta_mask, reg_c0);
2107 	else
2108 		sh->dv_regc0_mask = reg_c0;
2109 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2110 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2111 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2112 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2113 }
2114 
2115 /**
2116  * Allocate page of door-bells and register it using DevX API.
2117  *
2118  * @param [in] dev
2119  *   Pointer to Ethernet device.
2120  *
2121  * @return
2122  *   Pointer to new page on success, NULL otherwise.
2123  */
2124 static struct mlx5_devx_dbr_page *
2125 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
2126 {
2127 	struct mlx5_priv *priv = dev->data->dev_private;
2128 	struct mlx5_devx_dbr_page *page;
2129 
2130 	/* Allocate space for door-bell page and management data. */
2131 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
2132 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
2133 	if (!page) {
2134 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
2135 			dev->data->port_id);
2136 		return NULL;
2137 	}
2138 	/* Register allocated memory. */
2139 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
2140 					      MLX5_DBR_PAGE_SIZE, 0);
2141 	if (!page->umem) {
2142 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
2143 			dev->data->port_id);
2144 		rte_free(page);
2145 		return NULL;
2146 	}
2147 	return page;
2148 }
2149 
2150 /**
2151  * Find the next available door-bell, allocate new page if needed.
2152  *
2153  * @param [in] dev
2154  *   Pointer to Ethernet device.
2155  * @param [out] dbr_page
2156  *   Door-bell page containing the page data.
2157  *
2158  * @return
2159  *   Door-bell address offset on success, a negative error value otherwise.
2160  */
2161 int64_t
2162 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
2163 {
2164 	struct mlx5_priv *priv = dev->data->dev_private;
2165 	struct mlx5_devx_dbr_page *page = NULL;
2166 	uint32_t i, j;
2167 
2168 	LIST_FOREACH(page, &priv->dbrpgs, next)
2169 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
2170 			break;
2171 	if (!page) { /* No page with free door-bell exists. */
2172 		page = mlx5_alloc_dbr_page(dev);
2173 		if (!page) /* Failed to allocate new page. */
2174 			return (-1);
2175 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
2176 	}
2177 	/* Loop to find bitmap part with clear bit. */
2178 	for (i = 0;
2179 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
2180 	     i++)
2181 		; /* Empty. */
2182 	/* Find the first clear bit. */
2183 	MLX5_ASSERT(i < MLX5_DBR_BITMAP_SIZE);
2184 	j = rte_bsf64(~page->dbr_bitmap[i]);
2185 	page->dbr_bitmap[i] |= (1 << j);
2186 	page->dbr_count++;
2187 	*dbr_page = page;
2188 	return (((i * 64) + j) * sizeof(uint64_t));
2189 }
2190 
2191 /**
2192  * Release a door-bell record.
2193  *
2194  * @param [in] dev
2195  *   Pointer to Ethernet device.
2196  * @param [in] umem_id
2197  *   UMEM ID of page containing the door-bell record to release.
2198  * @param [in] offset
2199  *   Offset of door-bell record in page.
2200  *
2201  * @return
2202  *   0 on success, a negative error value otherwise.
2203  */
2204 int32_t
2205 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
2206 {
2207 	struct mlx5_priv *priv = dev->data->dev_private;
2208 	struct mlx5_devx_dbr_page *page = NULL;
2209 	int ret = 0;
2210 
2211 	LIST_FOREACH(page, &priv->dbrpgs, next)
2212 		/* Find the page this address belongs to. */
2213 		if (page->umem->umem_id == umem_id)
2214 			break;
2215 	if (!page)
2216 		return -EINVAL;
2217 	page->dbr_count--;
2218 	if (!page->dbr_count) {
2219 		/* Page not used, free it and remove from list. */
2220 		LIST_REMOVE(page, next);
2221 		if (page->umem)
2222 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
2223 		rte_free(page);
2224 	} else {
2225 		/* Mark in bitmap that this door-bell is not in use. */
2226 		offset /= MLX5_DBR_SIZE;
2227 		int i = offset / 64;
2228 		int j = offset % 64;
2229 
2230 		page->dbr_bitmap[i] &= ~(1 << j);
2231 	}
2232 	return ret;
2233 }
2234 
2235 int
2236 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2237 {
2238 	static const char *const dynf_names[] = {
2239 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2240 		RTE_MBUF_DYNFLAG_METADATA_NAME
2241 	};
2242 	unsigned int i;
2243 
2244 	if (n < RTE_DIM(dynf_names))
2245 		return -ENOMEM;
2246 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2247 		if (names[i] == NULL)
2248 			return -EINVAL;
2249 		strcpy(names[i], dynf_names[i]);
2250 	}
2251 	return RTE_DIM(dynf_names);
2252 }
2253 
2254 /**
2255  * Check sibling device configurations.
2256  *
2257  * Sibling devices sharing the Infiniband device context
2258  * should have compatible configurations. This regards
2259  * representors and bonding slaves.
2260  *
2261  * @param priv
2262  *   Private device descriptor.
2263  * @param config
2264  *   Configuration of the device is going to be created.
2265  *
2266  * @return
2267  *   0 on success, EINVAL otherwise
2268  */
2269 static int
2270 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2271 			      struct mlx5_dev_config *config)
2272 {
2273 	struct mlx5_ibv_shared *sh = priv->sh;
2274 	struct mlx5_dev_config *sh_conf = NULL;
2275 	uint16_t port_id;
2276 
2277 	MLX5_ASSERT(sh);
2278 	/* Nothing to compare for the single/first device. */
2279 	if (sh->refcnt == 1)
2280 		return 0;
2281 	/* Find the device with shared context. */
2282 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2283 		struct mlx5_priv *opriv =
2284 			rte_eth_devices[port_id].data->dev_private;
2285 
2286 		if (opriv && opriv != priv && opriv->sh == sh) {
2287 			sh_conf = &opriv->config;
2288 			break;
2289 		}
2290 	}
2291 	if (!sh_conf)
2292 		return 0;
2293 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2294 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2295 			     " for shared %s context", sh->ibdev_name);
2296 		rte_errno = EINVAL;
2297 		return rte_errno;
2298 	}
2299 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2300 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2301 			     " for shared %s context", sh->ibdev_name);
2302 		rte_errno = EINVAL;
2303 		return rte_errno;
2304 	}
2305 	return 0;
2306 }
2307 /**
2308  * Spawn an Ethernet device from Verbs information.
2309  *
2310  * @param dpdk_dev
2311  *   Backing DPDK device.
2312  * @param spawn
2313  *   Verbs device parameters (name, port, switch_info) to spawn.
2314  * @param config
2315  *   Device configuration parameters.
2316  *
2317  * @return
2318  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2319  *   is set. The following errors are defined:
2320  *
2321  *   EBUSY: device is not supposed to be spawned.
2322  *   EEXIST: device is already spawned
2323  */
2324 static struct rte_eth_dev *
2325 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2326 	       struct mlx5_dev_spawn_data *spawn,
2327 	       struct mlx5_dev_config config)
2328 {
2329 	const struct mlx5_switch_info *switch_info = &spawn->info;
2330 	struct mlx5_ibv_shared *sh = NULL;
2331 	struct ibv_port_attr port_attr;
2332 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2333 	struct rte_eth_dev *eth_dev = NULL;
2334 	struct mlx5_priv *priv = NULL;
2335 	int err = 0;
2336 	unsigned int hw_padding = 0;
2337 	unsigned int mps;
2338 	unsigned int cqe_comp;
2339 	unsigned int cqe_pad = 0;
2340 	unsigned int tunnel_en = 0;
2341 	unsigned int mpls_en = 0;
2342 	unsigned int swp = 0;
2343 	unsigned int mprq = 0;
2344 	unsigned int mprq_min_stride_size_n = 0;
2345 	unsigned int mprq_max_stride_size_n = 0;
2346 	unsigned int mprq_min_stride_num_n = 0;
2347 	unsigned int mprq_max_stride_num_n = 0;
2348 	struct rte_ether_addr mac;
2349 	char name[RTE_ETH_NAME_MAX_LEN];
2350 	int own_domain_id = 0;
2351 	uint16_t port_id;
2352 	unsigned int i;
2353 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2354 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2355 #endif
2356 
2357 	/* Determine if this port representor is supposed to be spawned. */
2358 	if (switch_info->representor && dpdk_dev->devargs) {
2359 		struct rte_eth_devargs eth_da;
2360 
2361 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2362 		if (err) {
2363 			rte_errno = -err;
2364 			DRV_LOG(ERR, "failed to process device arguments: %s",
2365 				strerror(rte_errno));
2366 			return NULL;
2367 		}
2368 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
2369 			if (eth_da.representor_ports[i] ==
2370 			    (uint16_t)switch_info->port_name)
2371 				break;
2372 		if (i == eth_da.nb_representor_ports) {
2373 			rte_errno = EBUSY;
2374 			return NULL;
2375 		}
2376 	}
2377 	/* Build device name. */
2378 	if (spawn->pf_bond <  0) {
2379 		/* Single device. */
2380 		if (!switch_info->representor)
2381 			strlcpy(name, dpdk_dev->name, sizeof(name));
2382 		else
2383 			snprintf(name, sizeof(name), "%s_representor_%u",
2384 				 dpdk_dev->name, switch_info->port_name);
2385 	} else {
2386 		/* Bonding device. */
2387 		if (!switch_info->representor)
2388 			snprintf(name, sizeof(name), "%s_%s",
2389 				 dpdk_dev->name, spawn->ibv_dev->name);
2390 		else
2391 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
2392 				 dpdk_dev->name, spawn->ibv_dev->name,
2393 				 switch_info->port_name);
2394 	}
2395 	/* check if the device is already spawned */
2396 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2397 		rte_errno = EEXIST;
2398 		return NULL;
2399 	}
2400 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2401 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2402 		struct mlx5_mp_id mp_id;
2403 
2404 		eth_dev = rte_eth_dev_attach_secondary(name);
2405 		if (eth_dev == NULL) {
2406 			DRV_LOG(ERR, "can not attach rte ethdev");
2407 			rte_errno = ENOMEM;
2408 			return NULL;
2409 		}
2410 		eth_dev->device = dpdk_dev;
2411 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2412 		err = mlx5_proc_priv_init(eth_dev);
2413 		if (err)
2414 			return NULL;
2415 		mp_id.port_id = eth_dev->data->port_id;
2416 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
2417 		/* Receive command fd from primary process */
2418 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
2419 		if (err < 0)
2420 			return NULL;
2421 		/* Remap UAR for Tx queues. */
2422 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2423 		if (err)
2424 			return NULL;
2425 		/*
2426 		 * Ethdev pointer is still required as input since
2427 		 * the primary device is not accessible from the
2428 		 * secondary process.
2429 		 */
2430 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2431 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2432 		return eth_dev;
2433 	}
2434 	/*
2435 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2436 	 * advance to create dv/verbs device context. We proceed the
2437 	 * devargs here to get ones, and later proceed devargs again
2438 	 * to override some hardware settings.
2439 	 */
2440 	err = mlx5_args(&config, dpdk_dev->devargs);
2441 	if (err) {
2442 		err = rte_errno;
2443 		DRV_LOG(ERR, "failed to process device arguments: %s",
2444 			strerror(rte_errno));
2445 		goto error;
2446 	}
2447 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2448 	if (!sh)
2449 		return NULL;
2450 	config.devx = sh->devx;
2451 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2452 	config.dest_tir = 1;
2453 #endif
2454 #ifdef HAVE_IBV_MLX5_MOD_SWP
2455 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2456 #endif
2457 	/*
2458 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2459 	 * as all ConnectX-5 devices.
2460 	 */
2461 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2462 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2463 #endif
2464 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2465 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2466 #endif
2467 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2468 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2469 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2470 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2471 			mps = MLX5_MPW_ENHANCED;
2472 		} else {
2473 			DRV_LOG(DEBUG, "MPW is supported");
2474 			mps = MLX5_MPW;
2475 		}
2476 	} else {
2477 		DRV_LOG(DEBUG, "MPW isn't supported");
2478 		mps = MLX5_MPW_DISABLED;
2479 	}
2480 #ifdef HAVE_IBV_MLX5_MOD_SWP
2481 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2482 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2483 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2484 #endif
2485 	config.swp = !!swp;
2486 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2487 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2488 		struct mlx5dv_striding_rq_caps mprq_caps =
2489 			dv_attr.striding_rq_caps;
2490 
2491 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2492 			mprq_caps.min_single_stride_log_num_of_bytes);
2493 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2494 			mprq_caps.max_single_stride_log_num_of_bytes);
2495 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2496 			mprq_caps.min_single_wqe_log_num_of_strides);
2497 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2498 			mprq_caps.max_single_wqe_log_num_of_strides);
2499 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2500 			mprq_caps.supported_qpts);
2501 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2502 		mprq = 1;
2503 		mprq_min_stride_size_n =
2504 			mprq_caps.min_single_stride_log_num_of_bytes;
2505 		mprq_max_stride_size_n =
2506 			mprq_caps.max_single_stride_log_num_of_bytes;
2507 		mprq_min_stride_num_n =
2508 			mprq_caps.min_single_wqe_log_num_of_strides;
2509 		mprq_max_stride_num_n =
2510 			mprq_caps.max_single_wqe_log_num_of_strides;
2511 	}
2512 #endif
2513 	if (RTE_CACHE_LINE_SIZE == 128 &&
2514 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2515 		cqe_comp = 0;
2516 	else
2517 		cqe_comp = 1;
2518 	config.cqe_comp = cqe_comp;
2519 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2520 	/* Whether device supports 128B Rx CQE padding. */
2521 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2522 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2523 #endif
2524 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2525 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2526 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2527 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2528 			     (dv_attr.tunnel_offloads_caps &
2529 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2530 			     (dv_attr.tunnel_offloads_caps &
2531 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2532 	}
2533 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2534 		tunnel_en ? "" : "not ");
2535 #else
2536 	DRV_LOG(WARNING,
2537 		"tunnel offloading disabled due to old OFED/rdma-core version");
2538 #endif
2539 	config.tunnel_en = tunnel_en;
2540 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2541 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2542 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2543 		   (dv_attr.tunnel_offloads_caps &
2544 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2545 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2546 		mpls_en ? "" : "not ");
2547 #else
2548 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2549 		" old OFED/rdma-core version or firmware configuration");
2550 #endif
2551 	config.mpls_en = mpls_en;
2552 	/* Check port status. */
2553 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2554 	if (err) {
2555 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2556 		goto error;
2557 	}
2558 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2559 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2560 		err = EINVAL;
2561 		goto error;
2562 	}
2563 	if (port_attr.state != IBV_PORT_ACTIVE)
2564 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2565 			mlx5_glue->port_state_str(port_attr.state),
2566 			port_attr.state);
2567 	/* Allocate private eth device data. */
2568 	priv = rte_zmalloc("ethdev private structure",
2569 			   sizeof(*priv),
2570 			   RTE_CACHE_LINE_SIZE);
2571 	if (priv == NULL) {
2572 		DRV_LOG(ERR, "priv allocation failure");
2573 		err = ENOMEM;
2574 		goto error;
2575 	}
2576 	priv->sh = sh;
2577 	priv->ibv_port = spawn->ibv_port;
2578 	priv->pci_dev = spawn->pci_dev;
2579 	priv->mtu = RTE_ETHER_MTU;
2580 	priv->mp_id.port_id = port_id;
2581 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
2582 #ifndef RTE_ARCH_64
2583 	/* Initialize UAR access locks for 32bit implementations. */
2584 	rte_spinlock_init(&priv->uar_lock_cq);
2585 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2586 		rte_spinlock_init(&priv->uar_lock[i]);
2587 #endif
2588 	/* Some internal functions rely on Netlink sockets, open them now. */
2589 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2590 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2591 	priv->representor = !!switch_info->representor;
2592 	priv->master = !!switch_info->master;
2593 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2594 	priv->vport_meta_tag = 0;
2595 	priv->vport_meta_mask = 0;
2596 	priv->pf_bond = spawn->pf_bond;
2597 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2598 	/*
2599 	 * The DevX port query API is implemented. E-Switch may use
2600 	 * either vport or reg_c[0] metadata register to match on
2601 	 * vport index. The engaged part of metadata register is
2602 	 * defined by mask.
2603 	 */
2604 	if (switch_info->representor || switch_info->master) {
2605 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2606 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2607 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2608 						 &devx_port);
2609 		if (err) {
2610 			DRV_LOG(WARNING,
2611 				"can't query devx port %d on device %s",
2612 				spawn->ibv_port, spawn->ibv_dev->name);
2613 			devx_port.comp_mask = 0;
2614 		}
2615 	}
2616 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2617 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2618 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2619 		if (!priv->vport_meta_mask) {
2620 			DRV_LOG(ERR, "vport zero mask for port %d"
2621 				     " on bonding device %s",
2622 				     spawn->ibv_port, spawn->ibv_dev->name);
2623 			err = ENOTSUP;
2624 			goto error;
2625 		}
2626 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2627 			DRV_LOG(ERR, "invalid vport tag for port %d"
2628 				     " on bonding device %s",
2629 				     spawn->ibv_port, spawn->ibv_dev->name);
2630 			err = ENOTSUP;
2631 			goto error;
2632 		}
2633 	}
2634 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2635 		priv->vport_id = devx_port.vport_num;
2636 	} else if (spawn->pf_bond >= 0) {
2637 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2638 			     " on bonding device %s",
2639 			     spawn->ibv_port, spawn->ibv_dev->name);
2640 		err = ENOTSUP;
2641 		goto error;
2642 	} else {
2643 		/* Suppose vport index in compatible way. */
2644 		priv->vport_id = switch_info->representor ?
2645 				 switch_info->port_name + 1 : -1;
2646 	}
2647 #else
2648 	/*
2649 	 * Kernel/rdma_core support single E-Switch per PF configurations
2650 	 * only and vport_id field contains the vport index for
2651 	 * associated VF, which is deduced from representor port name.
2652 	 * For example, let's have the IB device port 10, it has
2653 	 * attached network device eth0, which has port name attribute
2654 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2655 	 * as 3 (2+1). This assigning schema should be changed if the
2656 	 * multiple E-Switch instances per PF configurations or/and PCI
2657 	 * subfunctions are added.
2658 	 */
2659 	priv->vport_id = switch_info->representor ?
2660 			 switch_info->port_name + 1 : -1;
2661 #endif
2662 	/* representor_id field keeps the unmodified VF index. */
2663 	priv->representor_id = switch_info->representor ?
2664 			       switch_info->port_name : -1;
2665 	/*
2666 	 * Look for sibling devices in order to reuse their switch domain
2667 	 * if any, otherwise allocate one.
2668 	 */
2669 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2670 		const struct mlx5_priv *opriv =
2671 			rte_eth_devices[port_id].data->dev_private;
2672 
2673 		if (!opriv ||
2674 		    opriv->sh != priv->sh ||
2675 			opriv->domain_id ==
2676 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2677 			continue;
2678 		priv->domain_id = opriv->domain_id;
2679 		break;
2680 	}
2681 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2682 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2683 		if (err) {
2684 			err = rte_errno;
2685 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2686 				strerror(rte_errno));
2687 			goto error;
2688 		}
2689 		own_domain_id = 1;
2690 	}
2691 	/* Override some values set by hardware configuration. */
2692 	mlx5_args(&config, dpdk_dev->devargs);
2693 	err = mlx5_dev_check_sibling_config(priv, &config);
2694 	if (err)
2695 		goto error;
2696 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2697 			    IBV_DEVICE_RAW_IP_CSUM);
2698 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2699 		(config.hw_csum ? "" : "not "));
2700 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2701 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2702 	DRV_LOG(DEBUG, "counters are not supported");
2703 #endif
2704 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2705 	if (config.dv_flow_en) {
2706 		DRV_LOG(WARNING, "DV flow is not supported");
2707 		config.dv_flow_en = 0;
2708 	}
2709 #endif
2710 	config.ind_table_max_size =
2711 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2712 	/*
2713 	 * Remove this check once DPDK supports larger/variable
2714 	 * indirection tables.
2715 	 */
2716 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2717 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2718 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2719 		config.ind_table_max_size);
2720 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2721 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2722 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2723 		(config.hw_vlan_strip ? "" : "not "));
2724 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2725 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2726 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2727 		(config.hw_fcs_strip ? "" : "not "));
2728 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2729 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2730 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2731 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2732 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2733 #endif
2734 	if (config.hw_padding && !hw_padding) {
2735 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2736 		config.hw_padding = 0;
2737 	} else if (config.hw_padding) {
2738 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2739 	}
2740 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2741 		      (sh->device_attr.tso_caps.supported_qpts &
2742 		       (1 << IBV_QPT_RAW_PACKET)));
2743 	if (config.tso)
2744 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2745 	/*
2746 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2747 	 * by default.
2748 	 */
2749 	if (config.mps == MLX5_ARG_UNSET)
2750 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2751 							  MLX5_MPW_DISABLED;
2752 	else
2753 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2754 	DRV_LOG(INFO, "%sMPS is %s",
2755 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2756 		config.mps == MLX5_MPW ? "legacy " : "",
2757 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2758 	if (config.cqe_comp && !cqe_comp) {
2759 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2760 		config.cqe_comp = 0;
2761 	}
2762 	if (config.cqe_pad && !cqe_pad) {
2763 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2764 		config.cqe_pad = 0;
2765 	} else if (config.cqe_pad) {
2766 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2767 	}
2768 	if (config.devx) {
2769 		priv->counter_fallback = 0;
2770 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2771 		if (err) {
2772 			err = -err;
2773 			goto error;
2774 		}
2775 		if (!config.hca_attr.flow_counters_dump)
2776 			priv->counter_fallback = 1;
2777 #ifndef HAVE_IBV_DEVX_ASYNC
2778 		priv->counter_fallback = 1;
2779 #endif
2780 		if (priv->counter_fallback)
2781 			DRV_LOG(INFO, "Use fall-back DV counter management");
2782 		/* Check for LRO support. */
2783 		if (config.dest_tir && config.hca_attr.lro_cap &&
2784 		    config.dv_flow_en) {
2785 			/* TBD check tunnel lro caps. */
2786 			config.lro.supported = config.hca_attr.lro_cap;
2787 			DRV_LOG(DEBUG, "Device supports LRO");
2788 			/*
2789 			 * If LRO timeout is not configured by application,
2790 			 * use the minimal supported value.
2791 			 */
2792 			if (!config.lro.timeout)
2793 				config.lro.timeout =
2794 				config.hca_attr.lro_timer_supported_periods[0];
2795 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2796 				config.lro.timeout);
2797 		}
2798 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2799 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2800 		    config.dv_flow_en) {
2801 			uint8_t reg_c_mask =
2802 				config.hca_attr.qos.flow_meter_reg_c_ids;
2803 			/*
2804 			 * Meter needs two REG_C's for color match and pre-sfx
2805 			 * flow match. Here get the REG_C for color match.
2806 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2807 			 */
2808 			reg_c_mask &= 0xfc;
2809 			if (__builtin_popcount(reg_c_mask) < 1) {
2810 				priv->mtr_en = 0;
2811 				DRV_LOG(WARNING, "No available register for"
2812 					" meter.");
2813 			} else {
2814 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2815 						      REG_C_0;
2816 				priv->mtr_en = 1;
2817 				priv->mtr_reg_share =
2818 				      config.hca_attr.qos.flow_meter_reg_share;
2819 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2820 					priv->mtr_color_reg);
2821 			}
2822 		}
2823 #endif
2824 	}
2825 	if (config.mprq.enabled && mprq) {
2826 		if (config.mprq.stride_num_n &&
2827 		    (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2828 		     config.mprq.stride_num_n < mprq_min_stride_num_n)) {
2829 			config.mprq.stride_num_n =
2830 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2831 						mprq_min_stride_num_n),
2832 					mprq_max_stride_num_n);
2833 			DRV_LOG(WARNING,
2834 				"the number of strides"
2835 				" for Multi-Packet RQ is out of range,"
2836 				" setting default value (%u)",
2837 				1 << config.mprq.stride_num_n);
2838 		}
2839 		if (config.mprq.stride_size_n &&
2840 		    (config.mprq.stride_size_n > mprq_max_stride_size_n ||
2841 		     config.mprq.stride_size_n < mprq_min_stride_size_n)) {
2842 			config.mprq.stride_size_n =
2843 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
2844 						mprq_min_stride_size_n),
2845 					mprq_max_stride_size_n);
2846 			DRV_LOG(WARNING,
2847 				"the size of a stride"
2848 				" for Multi-Packet RQ is out of range,"
2849 				" setting default value (%u)",
2850 				1 << config.mprq.stride_size_n);
2851 		}
2852 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2853 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2854 	} else if (config.mprq.enabled && !mprq) {
2855 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2856 		config.mprq.enabled = 0;
2857 	}
2858 	if (config.max_dump_files_num == 0)
2859 		config.max_dump_files_num = 128;
2860 	eth_dev = rte_eth_dev_allocate(name);
2861 	if (eth_dev == NULL) {
2862 		DRV_LOG(ERR, "can not allocate rte ethdev");
2863 		err = ENOMEM;
2864 		goto error;
2865 	}
2866 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2867 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2868 	if (priv->representor) {
2869 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2870 		eth_dev->data->representor_id = priv->representor_id;
2871 	}
2872 	/*
2873 	 * Store associated network device interface index. This index
2874 	 * is permanent throughout the lifetime of device. So, we may store
2875 	 * the ifindex here and use the cached value further.
2876 	 */
2877 	MLX5_ASSERT(spawn->ifindex);
2878 	priv->if_index = spawn->ifindex;
2879 	eth_dev->data->dev_private = priv;
2880 	priv->dev_data = eth_dev->data;
2881 	eth_dev->data->mac_addrs = priv->mac;
2882 	eth_dev->device = dpdk_dev;
2883 	/* Configure the first MAC address by default. */
2884 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2885 		DRV_LOG(ERR,
2886 			"port %u cannot get MAC address, is mlx5_en"
2887 			" loaded? (errno: %s)",
2888 			eth_dev->data->port_id, strerror(rte_errno));
2889 		err = ENODEV;
2890 		goto error;
2891 	}
2892 	DRV_LOG(INFO,
2893 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2894 		eth_dev->data->port_id,
2895 		mac.addr_bytes[0], mac.addr_bytes[1],
2896 		mac.addr_bytes[2], mac.addr_bytes[3],
2897 		mac.addr_bytes[4], mac.addr_bytes[5]);
2898 #ifdef RTE_LIBRTE_MLX5_DEBUG
2899 	{
2900 		char ifname[IF_NAMESIZE];
2901 
2902 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2903 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2904 				eth_dev->data->port_id, ifname);
2905 		else
2906 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2907 				eth_dev->data->port_id);
2908 	}
2909 #endif
2910 	/* Get actual MTU if possible. */
2911 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2912 	if (err) {
2913 		err = rte_errno;
2914 		goto error;
2915 	}
2916 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2917 		priv->mtu);
2918 	/* Initialize burst functions to prevent crashes before link-up. */
2919 	eth_dev->rx_pkt_burst = removed_rx_burst;
2920 	eth_dev->tx_pkt_burst = removed_tx_burst;
2921 	eth_dev->dev_ops = &mlx5_dev_ops;
2922 	/* Register MAC address. */
2923 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2924 	if (config.vf && config.vf_nl_en)
2925 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2926 				      mlx5_ifindex(eth_dev),
2927 				      eth_dev->data->mac_addrs,
2928 				      MLX5_MAX_MAC_ADDRESSES);
2929 	priv->flows = 0;
2930 	priv->ctrl_flows = 0;
2931 	TAILQ_INIT(&priv->flow_meters);
2932 	TAILQ_INIT(&priv->flow_meter_profiles);
2933 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2934 	struct mlx5dv_ctx_allocators alctr = {
2935 		.alloc = &mlx5_alloc_verbs_buf,
2936 		.free = &mlx5_free_verbs_buf,
2937 		.data = priv,
2938 	};
2939 	mlx5_glue->dv_set_context_attr(sh->ctx,
2940 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2941 				       (void *)((uintptr_t)&alctr));
2942 	/* Bring Ethernet device up. */
2943 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2944 		eth_dev->data->port_id);
2945 	mlx5_set_link_up(eth_dev);
2946 	/*
2947 	 * Even though the interrupt handler is not installed yet,
2948 	 * interrupts will still trigger on the async_fd from
2949 	 * Verbs context returned by ibv_open_device().
2950 	 */
2951 	mlx5_link_update(eth_dev, 0);
2952 #ifdef HAVE_MLX5DV_DR_ESWITCH
2953 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2954 	      (switch_info->representor || switch_info->master)))
2955 		config.dv_esw_en = 0;
2956 #else
2957 	config.dv_esw_en = 0;
2958 #endif
2959 	/* Detect minimal data bytes to inline. */
2960 	mlx5_set_min_inline(spawn, &config);
2961 	/* Store device configuration on private structure. */
2962 	priv->config = config;
2963 	/* Create context for virtual machine VLAN workaround. */
2964 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2965 	if (config.dv_flow_en) {
2966 		err = mlx5_alloc_shared_dr(priv);
2967 		if (err)
2968 			goto error;
2969 		/*
2970 		 * RSS id is shared with meter flow id. Meter flow id can only
2971 		 * use the 24 MSB of the register.
2972 		 */
2973 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2974 				     MLX5_MTR_COLOR_BITS);
2975 		if (!priv->qrss_id_pool) {
2976 			DRV_LOG(ERR, "can't create flow id pool");
2977 			err = ENOMEM;
2978 			goto error;
2979 		}
2980 	}
2981 	/* Supported Verbs flow priority number detection. */
2982 	err = mlx5_flow_discover_priorities(eth_dev);
2983 	if (err < 0) {
2984 		err = -err;
2985 		goto error;
2986 	}
2987 	priv->config.flow_prio = err;
2988 	if (!priv->config.dv_esw_en &&
2989 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2990 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2991 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2992 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2993 	}
2994 	mlx5_set_metadata_mask(eth_dev);
2995 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2996 	    !priv->sh->dv_regc0_mask) {
2997 		DRV_LOG(ERR, "metadata mode %u is not supported "
2998 			     "(no metadata reg_c[0] is available)",
2999 			     priv->config.dv_xmeta_en);
3000 			err = ENOTSUP;
3001 			goto error;
3002 	}
3003 	/*
3004 	 * Allocate the buffer for flow creating, just once.
3005 	 * The allocation must be done before any flow creating.
3006 	 */
3007 	mlx5_flow_alloc_intermediate(eth_dev);
3008 	/* Query availibility of metadata reg_c's. */
3009 	err = mlx5_flow_discover_mreg_c(eth_dev);
3010 	if (err < 0) {
3011 		err = -err;
3012 		goto error;
3013 	}
3014 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
3015 		DRV_LOG(DEBUG,
3016 			"port %u extensive metadata register is not supported",
3017 			eth_dev->data->port_id);
3018 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
3019 			DRV_LOG(ERR, "metadata mode %u is not supported "
3020 				     "(no metadata registers available)",
3021 				     priv->config.dv_xmeta_en);
3022 			err = ENOTSUP;
3023 			goto error;
3024 		}
3025 	}
3026 	if (priv->config.dv_flow_en &&
3027 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
3028 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
3029 	    priv->sh->dv_regc0_mask) {
3030 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
3031 						      MLX5_FLOW_MREG_HTABLE_SZ);
3032 		if (!priv->mreg_cp_tbl) {
3033 			err = ENOMEM;
3034 			goto error;
3035 		}
3036 	}
3037 	return eth_dev;
3038 error:
3039 	if (priv) {
3040 		if (priv->mreg_cp_tbl)
3041 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
3042 		if (priv->sh)
3043 			mlx5_free_shared_dr(priv);
3044 		if (priv->nl_socket_route >= 0)
3045 			close(priv->nl_socket_route);
3046 		if (priv->nl_socket_rdma >= 0)
3047 			close(priv->nl_socket_rdma);
3048 		if (priv->vmwa_context)
3049 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
3050 		if (priv->qrss_id_pool)
3051 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
3052 		if (own_domain_id)
3053 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
3054 		rte_free(priv);
3055 		if (eth_dev != NULL)
3056 			eth_dev->data->dev_private = NULL;
3057 	}
3058 	if (eth_dev != NULL) {
3059 		/* mac_addrs must not be freed alone because part of dev_private */
3060 		eth_dev->data->mac_addrs = NULL;
3061 		rte_eth_dev_release_port(eth_dev);
3062 	}
3063 	if (sh)
3064 		mlx5_free_shared_ibctx(sh);
3065 	MLX5_ASSERT(err > 0);
3066 	rte_errno = err;
3067 	return NULL;
3068 }
3069 
3070 /**
3071  * Comparison callback to sort device data.
3072  *
3073  * This is meant to be used with qsort().
3074  *
3075  * @param a[in]
3076  *   Pointer to pointer to first data object.
3077  * @param b[in]
3078  *   Pointer to pointer to second data object.
3079  *
3080  * @return
3081  *   0 if both objects are equal, less than 0 if the first argument is less
3082  *   than the second, greater than 0 otherwise.
3083  */
3084 static int
3085 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
3086 {
3087 	const struct mlx5_switch_info *si_a =
3088 		&((const struct mlx5_dev_spawn_data *)a)->info;
3089 	const struct mlx5_switch_info *si_b =
3090 		&((const struct mlx5_dev_spawn_data *)b)->info;
3091 	int ret;
3092 
3093 	/* Master device first. */
3094 	ret = si_b->master - si_a->master;
3095 	if (ret)
3096 		return ret;
3097 	/* Then representor devices. */
3098 	ret = si_b->representor - si_a->representor;
3099 	if (ret)
3100 		return ret;
3101 	/* Unidentified devices come last in no specific order. */
3102 	if (!si_a->representor)
3103 		return 0;
3104 	/* Order representors by name. */
3105 	return si_a->port_name - si_b->port_name;
3106 }
3107 
3108 /**
3109  * Match PCI information for possible slaves of bonding device.
3110  *
3111  * @param[in] ibv_dev
3112  *   Pointer to Infiniband device structure.
3113  * @param[in] pci_dev
3114  *   Pointer to PCI device structure to match PCI address.
3115  * @param[in] nl_rdma
3116  *   Netlink RDMA group socket handle.
3117  *
3118  * @return
3119  *   negative value if no bonding device found, otherwise
3120  *   positive index of slave PF in bonding.
3121  */
3122 static int
3123 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
3124 			   const struct rte_pci_device *pci_dev,
3125 			   int nl_rdma)
3126 {
3127 	char ifname[IF_NAMESIZE + 1];
3128 	unsigned int ifindex;
3129 	unsigned int np, i;
3130 	FILE *file = NULL;
3131 	int pf = -1;
3132 
3133 	/*
3134 	 * Try to get master device name. If something goes
3135 	 * wrong suppose the lack of kernel support and no
3136 	 * bonding devices.
3137 	 */
3138 	if (nl_rdma < 0)
3139 		return -1;
3140 	if (!strstr(ibv_dev->name, "bond"))
3141 		return -1;
3142 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
3143 	if (!np)
3144 		return -1;
3145 	/*
3146 	 * The Master device might not be on the predefined
3147 	 * port (not on port index 1, it is not garanted),
3148 	 * we have to scan all Infiniband device port and
3149 	 * find master.
3150 	 */
3151 	for (i = 1; i <= np; ++i) {
3152 		/* Check whether Infiniband port is populated. */
3153 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
3154 		if (!ifindex)
3155 			continue;
3156 		if (!if_indextoname(ifindex, ifname))
3157 			continue;
3158 		/* Try to read bonding slave names from sysfs. */
3159 		MKSTR(slaves,
3160 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
3161 		file = fopen(slaves, "r");
3162 		if (file)
3163 			break;
3164 	}
3165 	if (!file)
3166 		return -1;
3167 	/* Use safe format to check maximal buffer length. */
3168 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
3169 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
3170 		char tmp_str[IF_NAMESIZE + 32];
3171 		struct rte_pci_addr pci_addr;
3172 		struct mlx5_switch_info	info;
3173 
3174 		/* Process slave interface names in the loop. */
3175 		snprintf(tmp_str, sizeof(tmp_str),
3176 			 "/sys/class/net/%s", ifname);
3177 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
3178 			DRV_LOG(WARNING, "can not get PCI address"
3179 					 " for netdev \"%s\"", ifname);
3180 			continue;
3181 		}
3182 		if (pci_dev->addr.domain != pci_addr.domain ||
3183 		    pci_dev->addr.bus != pci_addr.bus ||
3184 		    pci_dev->addr.devid != pci_addr.devid ||
3185 		    pci_dev->addr.function != pci_addr.function)
3186 			continue;
3187 		/* Slave interface PCI address match found. */
3188 		fclose(file);
3189 		snprintf(tmp_str, sizeof(tmp_str),
3190 			 "/sys/class/net/%s/phys_port_name", ifname);
3191 		file = fopen(tmp_str, "rb");
3192 		if (!file)
3193 			break;
3194 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
3195 		if (fscanf(file, "%32s", tmp_str) == 1)
3196 			mlx5_translate_port_name(tmp_str, &info);
3197 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
3198 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
3199 			pf = info.port_name;
3200 		break;
3201 	}
3202 	if (file)
3203 		fclose(file);
3204 	return pf;
3205 }
3206 
3207 /**
3208  * DPDK callback to register a PCI device.
3209  *
3210  * This function spawns Ethernet devices out of a given PCI device.
3211  *
3212  * @param[in] pci_drv
3213  *   PCI driver structure (mlx5_driver).
3214  * @param[in] pci_dev
3215  *   PCI device information.
3216  *
3217  * @return
3218  *   0 on success, a negative errno value otherwise and rte_errno is set.
3219  */
3220 static int
3221 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
3222 	       struct rte_pci_device *pci_dev)
3223 {
3224 	struct ibv_device **ibv_list;
3225 	/*
3226 	 * Number of found IB Devices matching with requested PCI BDF.
3227 	 * nd != 1 means there are multiple IB devices over the same
3228 	 * PCI device and we have representors and master.
3229 	 */
3230 	unsigned int nd = 0;
3231 	/*
3232 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
3233 	 * we have the single multiport IB device, and there may be
3234 	 * representors attached to some of found ports.
3235 	 */
3236 	unsigned int np = 0;
3237 	/*
3238 	 * Number of DPDK ethernet devices to Spawn - either over
3239 	 * multiple IB devices or multiple ports of single IB device.
3240 	 * Actually this is the number of iterations to spawn.
3241 	 */
3242 	unsigned int ns = 0;
3243 	/*
3244 	 * Bonding device
3245 	 *   < 0 - no bonding device (single one)
3246 	 *  >= 0 - bonding device (value is slave PF index)
3247 	 */
3248 	int bd = -1;
3249 	struct mlx5_dev_spawn_data *list = NULL;
3250 	struct mlx5_dev_config dev_config;
3251 	int ret;
3252 
3253 	if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
3254 		DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
3255 			" driver.");
3256 		return 1;
3257 	}
3258 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
3259 		mlx5_pmd_socket_init();
3260 	ret = mlx5_init_once();
3261 	if (ret) {
3262 		DRV_LOG(ERR, "unable to init PMD global data: %s",
3263 			strerror(rte_errno));
3264 		return -rte_errno;
3265 	}
3266 	MLX5_ASSERT(pci_drv == &mlx5_driver);
3267 	errno = 0;
3268 	ibv_list = mlx5_glue->get_device_list(&ret);
3269 	if (!ibv_list) {
3270 		rte_errno = errno ? errno : ENOSYS;
3271 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
3272 		return -rte_errno;
3273 	}
3274 	/*
3275 	 * First scan the list of all Infiniband devices to find
3276 	 * matching ones, gathering into the list.
3277 	 */
3278 	struct ibv_device *ibv_match[ret + 1];
3279 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3280 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3281 	unsigned int i;
3282 
3283 	while (ret-- > 0) {
3284 		struct rte_pci_addr pci_addr;
3285 
3286 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3287 		bd = mlx5_device_bond_pci_match
3288 				(ibv_list[ret], pci_dev, nl_rdma);
3289 		if (bd >= 0) {
3290 			/*
3291 			 * Bonding device detected. Only one match is allowed,
3292 			 * the bonding is supported over multi-port IB device,
3293 			 * there should be no matches on representor PCI
3294 			 * functions or non VF LAG bonding devices with
3295 			 * specified address.
3296 			 */
3297 			if (nd) {
3298 				DRV_LOG(ERR,
3299 					"multiple PCI match on bonding device"
3300 					"\"%s\" found", ibv_list[ret]->name);
3301 				rte_errno = ENOENT;
3302 				ret = -rte_errno;
3303 				goto exit;
3304 			}
3305 			DRV_LOG(INFO, "PCI information matches for"
3306 				      " slave %d bonding device \"%s\"",
3307 				      bd, ibv_list[ret]->name);
3308 			ibv_match[nd++] = ibv_list[ret];
3309 			break;
3310 		}
3311 		if (mlx5_dev_to_pci_addr
3312 			(ibv_list[ret]->ibdev_path, &pci_addr))
3313 			continue;
3314 		if (pci_dev->addr.domain != pci_addr.domain ||
3315 		    pci_dev->addr.bus != pci_addr.bus ||
3316 		    pci_dev->addr.devid != pci_addr.devid ||
3317 		    pci_dev->addr.function != pci_addr.function)
3318 			continue;
3319 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3320 			ibv_list[ret]->name);
3321 		ibv_match[nd++] = ibv_list[ret];
3322 	}
3323 	ibv_match[nd] = NULL;
3324 	if (!nd) {
3325 		/* No device matches, just complain and bail out. */
3326 		DRV_LOG(WARNING,
3327 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
3328 			" are kernel drivers loaded?",
3329 			pci_dev->addr.domain, pci_dev->addr.bus,
3330 			pci_dev->addr.devid, pci_dev->addr.function);
3331 		rte_errno = ENOENT;
3332 		ret = -rte_errno;
3333 		goto exit;
3334 	}
3335 	if (nd == 1) {
3336 		/*
3337 		 * Found single matching device may have multiple ports.
3338 		 * Each port may be representor, we have to check the port
3339 		 * number and check the representors existence.
3340 		 */
3341 		if (nl_rdma >= 0)
3342 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3343 		if (!np)
3344 			DRV_LOG(WARNING, "can not get IB device \"%s\""
3345 					 " ports number", ibv_match[0]->name);
3346 		if (bd >= 0 && !np) {
3347 			DRV_LOG(ERR, "can not get ports"
3348 				     " for bonding device");
3349 			rte_errno = ENOENT;
3350 			ret = -rte_errno;
3351 			goto exit;
3352 		}
3353 	}
3354 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3355 	if (bd >= 0) {
3356 		/*
3357 		 * This may happen if there is VF LAG kernel support and
3358 		 * application is compiled with older rdma_core library.
3359 		 */
3360 		DRV_LOG(ERR,
3361 			"No kernel/verbs support for VF LAG bonding found.");
3362 		rte_errno = ENOTSUP;
3363 		ret = -rte_errno;
3364 		goto exit;
3365 	}
3366 #endif
3367 	/*
3368 	 * Now we can determine the maximal
3369 	 * amount of devices to be spawned.
3370 	 */
3371 	list = rte_zmalloc("device spawn data",
3372 			 sizeof(struct mlx5_dev_spawn_data) *
3373 			 (np ? np : nd),
3374 			 RTE_CACHE_LINE_SIZE);
3375 	if (!list) {
3376 		DRV_LOG(ERR, "spawn data array allocation failure");
3377 		rte_errno = ENOMEM;
3378 		ret = -rte_errno;
3379 		goto exit;
3380 	}
3381 	if (bd >= 0 || np > 1) {
3382 		/*
3383 		 * Single IB device with multiple ports found,
3384 		 * it may be E-Switch master device and representors.
3385 		 * We have to perform identification trough the ports.
3386 		 */
3387 		MLX5_ASSERT(nl_rdma >= 0);
3388 		MLX5_ASSERT(ns == 0);
3389 		MLX5_ASSERT(nd == 1);
3390 		MLX5_ASSERT(np);
3391 		for (i = 1; i <= np; ++i) {
3392 			list[ns].max_port = np;
3393 			list[ns].ibv_port = i;
3394 			list[ns].ibv_dev = ibv_match[0];
3395 			list[ns].eth_dev = NULL;
3396 			list[ns].pci_dev = pci_dev;
3397 			list[ns].pf_bond = bd;
3398 			list[ns].ifindex = mlx5_nl_ifindex
3399 					(nl_rdma, list[ns].ibv_dev->name, i);
3400 			if (!list[ns].ifindex) {
3401 				/*
3402 				 * No network interface index found for the
3403 				 * specified port, it means there is no
3404 				 * representor on this port. It's OK,
3405 				 * there can be disabled ports, for example
3406 				 * if sriov_numvfs < sriov_totalvfs.
3407 				 */
3408 				continue;
3409 			}
3410 			ret = -1;
3411 			if (nl_route >= 0)
3412 				ret = mlx5_nl_switch_info
3413 					       (nl_route,
3414 						list[ns].ifindex,
3415 						&list[ns].info);
3416 			if (ret || (!list[ns].info.representor &&
3417 				    !list[ns].info.master)) {
3418 				/*
3419 				 * We failed to recognize representors with
3420 				 * Netlink, let's try to perform the task
3421 				 * with sysfs.
3422 				 */
3423 				ret =  mlx5_sysfs_switch_info
3424 						(list[ns].ifindex,
3425 						 &list[ns].info);
3426 			}
3427 			if (!ret && bd >= 0) {
3428 				switch (list[ns].info.name_type) {
3429 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3430 					if (list[ns].info.port_name == bd)
3431 						ns++;
3432 					break;
3433 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3434 					if (list[ns].info.pf_num == bd)
3435 						ns++;
3436 					break;
3437 				default:
3438 					break;
3439 				}
3440 				continue;
3441 			}
3442 			if (!ret && (list[ns].info.representor ^
3443 				     list[ns].info.master))
3444 				ns++;
3445 		}
3446 		if (!ns) {
3447 			DRV_LOG(ERR,
3448 				"unable to recognize master/representors"
3449 				" on the IB device with multiple ports");
3450 			rte_errno = ENOENT;
3451 			ret = -rte_errno;
3452 			goto exit;
3453 		}
3454 	} else {
3455 		/*
3456 		 * The existence of several matching entries (nd > 1) means
3457 		 * port representors have been instantiated. No existing Verbs
3458 		 * call nor sysfs entries can tell them apart, this can only
3459 		 * be done through Netlink calls assuming kernel drivers are
3460 		 * recent enough to support them.
3461 		 *
3462 		 * In the event of identification failure through Netlink,
3463 		 * try again through sysfs, then:
3464 		 *
3465 		 * 1. A single IB device matches (nd == 1) with single
3466 		 *    port (np=0/1) and is not a representor, assume
3467 		 *    no switch support.
3468 		 *
3469 		 * 2. Otherwise no safe assumptions can be made;
3470 		 *    complain louder and bail out.
3471 		 */
3472 		np = 1;
3473 		for (i = 0; i != nd; ++i) {
3474 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3475 			list[ns].max_port = 1;
3476 			list[ns].ibv_port = 1;
3477 			list[ns].ibv_dev = ibv_match[i];
3478 			list[ns].eth_dev = NULL;
3479 			list[ns].pci_dev = pci_dev;
3480 			list[ns].pf_bond = -1;
3481 			list[ns].ifindex = 0;
3482 			if (nl_rdma >= 0)
3483 				list[ns].ifindex = mlx5_nl_ifindex
3484 					(nl_rdma, list[ns].ibv_dev->name, 1);
3485 			if (!list[ns].ifindex) {
3486 				char ifname[IF_NAMESIZE];
3487 
3488 				/*
3489 				 * Netlink failed, it may happen with old
3490 				 * ib_core kernel driver (before 4.16).
3491 				 * We can assume there is old driver because
3492 				 * here we are processing single ports IB
3493 				 * devices. Let's try sysfs to retrieve
3494 				 * the ifindex. The method works for
3495 				 * master device only.
3496 				 */
3497 				if (nd > 1) {
3498 					/*
3499 					 * Multiple devices found, assume
3500 					 * representors, can not distinguish
3501 					 * master/representor and retrieve
3502 					 * ifindex via sysfs.
3503 					 */
3504 					continue;
3505 				}
3506 				ret = mlx5_get_master_ifname
3507 					(ibv_match[i]->ibdev_path, &ifname);
3508 				if (!ret)
3509 					list[ns].ifindex =
3510 						if_nametoindex(ifname);
3511 				if (!list[ns].ifindex) {
3512 					/*
3513 					 * No network interface index found
3514 					 * for the specified device, it means
3515 					 * there it is neither representor
3516 					 * nor master.
3517 					 */
3518 					continue;
3519 				}
3520 			}
3521 			ret = -1;
3522 			if (nl_route >= 0)
3523 				ret = mlx5_nl_switch_info
3524 					       (nl_route,
3525 						list[ns].ifindex,
3526 						&list[ns].info);
3527 			if (ret || (!list[ns].info.representor &&
3528 				    !list[ns].info.master)) {
3529 				/*
3530 				 * We failed to recognize representors with
3531 				 * Netlink, let's try to perform the task
3532 				 * with sysfs.
3533 				 */
3534 				ret =  mlx5_sysfs_switch_info
3535 						(list[ns].ifindex,
3536 						 &list[ns].info);
3537 			}
3538 			if (!ret && (list[ns].info.representor ^
3539 				     list[ns].info.master)) {
3540 				ns++;
3541 			} else if ((nd == 1) &&
3542 				   !list[ns].info.representor &&
3543 				   !list[ns].info.master) {
3544 				/*
3545 				 * Single IB device with
3546 				 * one physical port and
3547 				 * attached network device.
3548 				 * May be SRIOV is not enabled
3549 				 * or there is no representors.
3550 				 */
3551 				DRV_LOG(INFO, "no E-Switch support detected");
3552 				ns++;
3553 				break;
3554 			}
3555 		}
3556 		if (!ns) {
3557 			DRV_LOG(ERR,
3558 				"unable to recognize master/representors"
3559 				" on the multiple IB devices");
3560 			rte_errno = ENOENT;
3561 			ret = -rte_errno;
3562 			goto exit;
3563 		}
3564 	}
3565 	MLX5_ASSERT(ns);
3566 	/*
3567 	 * Sort list to probe devices in natural order for users convenience
3568 	 * (i.e. master first, then representors from lowest to highest ID).
3569 	 */
3570 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3571 	/* Default configuration. */
3572 	dev_config = (struct mlx5_dev_config){
3573 		.hw_padding = 0,
3574 		.mps = MLX5_ARG_UNSET,
3575 		.dbnc = MLX5_ARG_UNSET,
3576 		.rx_vec_en = 1,
3577 		.txq_inline_max = MLX5_ARG_UNSET,
3578 		.txq_inline_min = MLX5_ARG_UNSET,
3579 		.txq_inline_mpw = MLX5_ARG_UNSET,
3580 		.txqs_inline = MLX5_ARG_UNSET,
3581 		.vf_nl_en = 1,
3582 		.mr_ext_memseg_en = 1,
3583 		.mprq = {
3584 			.enabled = 0, /* Disabled by default. */
3585 			.stride_num_n = 0,
3586 			.stride_size_n = 0,
3587 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3588 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3589 		},
3590 		.dv_esw_en = 1,
3591 		.dv_flow_en = 1,
3592 		.log_hp_size = MLX5_ARG_UNSET,
3593 	};
3594 	/* Device specific configuration. */
3595 	switch (pci_dev->id.device_id) {
3596 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3597 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3598 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3599 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3600 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3601 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3602 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3603 		dev_config.vf = 1;
3604 		break;
3605 	default:
3606 		break;
3607 	}
3608 	for (i = 0; i != ns; ++i) {
3609 		uint32_t restore;
3610 
3611 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3612 						 &list[i],
3613 						 dev_config);
3614 		if (!list[i].eth_dev) {
3615 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3616 				break;
3617 			/* Device is disabled or already spawned. Ignore it. */
3618 			continue;
3619 		}
3620 		restore = list[i].eth_dev->data->dev_flags;
3621 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3622 		/* Restore non-PCI flags cleared by the above call. */
3623 		list[i].eth_dev->data->dev_flags |= restore;
3624 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3625 		rte_eth_dev_probing_finish(list[i].eth_dev);
3626 	}
3627 	if (i != ns) {
3628 		DRV_LOG(ERR,
3629 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3630 			" encountering an error: %s",
3631 			pci_dev->addr.domain, pci_dev->addr.bus,
3632 			pci_dev->addr.devid, pci_dev->addr.function,
3633 			strerror(rte_errno));
3634 		ret = -rte_errno;
3635 		/* Roll back. */
3636 		while (i--) {
3637 			if (!list[i].eth_dev)
3638 				continue;
3639 			mlx5_dev_close(list[i].eth_dev);
3640 			/* mac_addrs must not be freed because in dev_private */
3641 			list[i].eth_dev->data->mac_addrs = NULL;
3642 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3643 		}
3644 		/* Restore original error. */
3645 		rte_errno = -ret;
3646 	} else {
3647 		ret = 0;
3648 	}
3649 exit:
3650 	/*
3651 	 * Do the routine cleanup:
3652 	 * - close opened Netlink sockets
3653 	 * - free allocated spawn data array
3654 	 * - free the Infiniband device list
3655 	 */
3656 	if (nl_rdma >= 0)
3657 		close(nl_rdma);
3658 	if (nl_route >= 0)
3659 		close(nl_route);
3660 	if (list)
3661 		rte_free(list);
3662 	MLX5_ASSERT(ibv_list);
3663 	mlx5_glue->free_device_list(ibv_list);
3664 	return ret;
3665 }
3666 
3667 /**
3668  * Look for the ethernet device belonging to mlx5 driver.
3669  *
3670  * @param[in] port_id
3671  *   port_id to start looking for device.
3672  * @param[in] pci_dev
3673  *   Pointer to the hint PCI device. When device is being probed
3674  *   the its siblings (master and preceding representors might
3675  *   not have assigned driver yet (because the mlx5_pci_probe()
3676  *   is not completed yet, for this case match on hint PCI
3677  *   device may be used to detect sibling device.
3678  *
3679  * @return
3680  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3681  */
3682 uint16_t
3683 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3684 {
3685 	while (port_id < RTE_MAX_ETHPORTS) {
3686 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3687 
3688 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3689 		    dev->device &&
3690 		    (dev->device == &pci_dev->device ||
3691 		     (dev->device->driver &&
3692 		     dev->device->driver->name &&
3693 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3694 			break;
3695 		port_id++;
3696 	}
3697 	if (port_id >= RTE_MAX_ETHPORTS)
3698 		return RTE_MAX_ETHPORTS;
3699 	return port_id;
3700 }
3701 
3702 /**
3703  * DPDK callback to remove a PCI device.
3704  *
3705  * This function removes all Ethernet devices belong to a given PCI device.
3706  *
3707  * @param[in] pci_dev
3708  *   Pointer to the PCI device.
3709  *
3710  * @return
3711  *   0 on success, the function cannot fail.
3712  */
3713 static int
3714 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3715 {
3716 	uint16_t port_id;
3717 
3718 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3719 		rte_eth_dev_close(port_id);
3720 	return 0;
3721 }
3722 
3723 static const struct rte_pci_id mlx5_pci_id_map[] = {
3724 	{
3725 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3726 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3727 	},
3728 	{
3729 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3730 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3731 	},
3732 	{
3733 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3734 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3735 	},
3736 	{
3737 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3738 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3739 	},
3740 	{
3741 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3742 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3743 	},
3744 	{
3745 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3746 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3747 	},
3748 	{
3749 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3750 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3751 	},
3752 	{
3753 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3754 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3755 	},
3756 	{
3757 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3758 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3759 	},
3760 	{
3761 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3762 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3763 	},
3764 	{
3765 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3766 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3767 	},
3768 	{
3769 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3770 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3771 	},
3772 	{
3773 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3774 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3775 	},
3776 	{
3777 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3778 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3779 	},
3780 	{
3781 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3782 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
3783 	},
3784 	{
3785 		.vendor_id = 0
3786 	}
3787 };
3788 
3789 static struct rte_pci_driver mlx5_driver = {
3790 	.driver = {
3791 		.name = MLX5_DRIVER_NAME
3792 	},
3793 	.id_table = mlx5_pci_id_map,
3794 	.probe = mlx5_pci_probe,
3795 	.remove = mlx5_pci_remove,
3796 	.dma_map = mlx5_dma_map,
3797 	.dma_unmap = mlx5_dma_unmap,
3798 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3799 		     RTE_PCI_DRV_PROBE_AGAIN,
3800 };
3801 
3802 /**
3803  * Driver initialization routine.
3804  */
3805 RTE_INIT(rte_mlx5_pmd_init)
3806 {
3807 	/* Initialize driver log type. */
3808 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3809 	if (mlx5_logtype >= 0)
3810 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3811 
3812 	/* Build the static tables for Verbs conversion. */
3813 	mlx5_set_ptype_table();
3814 	mlx5_set_cksum_table();
3815 	mlx5_set_swp_types_table();
3816 	if (mlx5_glue)
3817 		rte_pci_register(&mlx5_driver);
3818 }
3819 
3820 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3821 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3822 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3823