xref: /dpdk/drivers/net/mlx5/mlx5.c (revision dede694cd58f6ee9c53f1e7d8d5960835bdc8fff)
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 	int i;
471 
472 	memset(&sh->cmng, 0, sizeof(sh->cmng));
473 	TAILQ_INIT(&sh->cmng.flow_counters);
474 	for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
475 		TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
476 		rte_spinlock_init(&sh->cmng.ccont[i].resize_sl);
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 	int i;
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 	for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
519 		struct mlx5_flow_counter_pool *pool;
520 		uint32_t batch = !!(i > 1);
521 
522 		if (!sh->cmng.ccont[i].pools)
523 			continue;
524 		pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
525 		while (pool) {
526 			if (batch && pool->min_dcs)
527 				claim_zero(mlx5_devx_cmd_destroy
528 							       (pool->min_dcs));
529 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
530 				if (MLX5_POOL_GET_CNT(pool, j)->action)
531 					claim_zero
532 					 (mlx5_glue->destroy_flow_action
533 					  (MLX5_POOL_GET_CNT
534 					  (pool, j)->action));
535 				if (!batch && MLX5_GET_POOL_CNT_EXT
536 				    (pool, j)->dcs)
537 					claim_zero(mlx5_devx_cmd_destroy
538 						   (MLX5_GET_POOL_CNT_EXT
539 						    (pool, j)->dcs));
540 			}
541 			TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool, next);
542 			rte_free(pool);
543 			pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
544 		}
545 		rte_free(sh->cmng.ccont[i].pools);
546 	}
547 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
548 	while (mng) {
549 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
550 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
551 	}
552 	memset(&sh->cmng, 0, sizeof(sh->cmng));
553 }
554 
555 /**
556  * Initialize the flow resources' indexed mempool.
557  *
558  * @param[in] sh
559  *   Pointer to mlx5_ibv_shared object.
560  * @param[in] sh
561  *   Pointer to user dev config.
562  */
563 static void
564 mlx5_flow_ipool_create(struct mlx5_ibv_shared *sh,
565 		       const struct mlx5_dev_config *config __rte_unused)
566 {
567 	uint8_t i;
568 
569 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
570 	/*
571 	 * While DV is supported, user chooses the verbs mode,
572 	 * the mlx5 flow handle size is different with the
573 	 * MLX5_FLOW_HANDLE_VERBS_SIZE.
574 	 */
575 	if (!config->dv_flow_en)
576 		mlx5_ipool_cfg[MLX5_IPOOL_MLX5_FLOW].size =
577 					MLX5_FLOW_HANDLE_VERBS_SIZE;
578 #endif
579 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
580 		sh->ipool[i] = mlx5_ipool_create(&mlx5_ipool_cfg[i]);
581 }
582 
583 /**
584  * Release the flow resources' indexed mempool.
585  *
586  * @param[in] sh
587  *   Pointer to mlx5_ibv_shared object.
588  */
589 static void
590 mlx5_flow_ipool_destroy(struct mlx5_ibv_shared *sh)
591 {
592 	uint8_t i;
593 
594 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
595 		mlx5_ipool_destroy(sh->ipool[i]);
596 }
597 
598 /**
599  * Extract pdn of PD object using DV API.
600  *
601  * @param[in] pd
602  *   Pointer to the verbs PD object.
603  * @param[out] pdn
604  *   Pointer to the PD object number variable.
605  *
606  * @return
607  *   0 on success, error value otherwise.
608  */
609 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
610 static int
611 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
612 {
613 	struct mlx5dv_obj obj;
614 	struct mlx5dv_pd pd_info;
615 	int ret = 0;
616 
617 	obj.pd.in = pd;
618 	obj.pd.out = &pd_info;
619 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
620 	if (ret) {
621 		DRV_LOG(DEBUG, "Fail to get PD object info");
622 		return ret;
623 	}
624 	*pdn = pd_info.pdn;
625 	return 0;
626 }
627 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
628 
629 static int
630 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
631 {
632 	char *env;
633 	int value;
634 
635 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
636 	/* Get environment variable to store. */
637 	env = getenv(MLX5_SHUT_UP_BF);
638 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
639 	if (config->dbnc == MLX5_ARG_UNSET)
640 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
641 	else
642 		setenv(MLX5_SHUT_UP_BF,
643 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
644 	return value;
645 }
646 
647 static void
648 mlx5_restore_doorbell_mapping_env(int value)
649 {
650 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
651 	/* Restore the original environment variable state. */
652 	if (value == MLX5_ARG_UNSET)
653 		unsetenv(MLX5_SHUT_UP_BF);
654 	else
655 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
656 }
657 
658 /**
659  * Allocate shared IB device context. If there is multiport device the
660  * master and representors will share this context, if there is single
661  * port dedicated IB device, the context will be used by only given
662  * port due to unification.
663  *
664  * Routine first searches the context for the specified IB device name,
665  * if found the shared context assumed and reference counter is incremented.
666  * If no context found the new one is created and initialized with specified
667  * IB device context and parameters.
668  *
669  * @param[in] spawn
670  *   Pointer to the IB device attributes (name, port, etc).
671  * @param[in] config
672  *   Pointer to device configuration structure.
673  *
674  * @return
675  *   Pointer to mlx5_ibv_shared object on success,
676  *   otherwise NULL and rte_errno is set.
677  */
678 static struct mlx5_ibv_shared *
679 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
680 			const struct mlx5_dev_config *config)
681 {
682 	struct mlx5_ibv_shared *sh;
683 	int dbmap_env;
684 	int err = 0;
685 	uint32_t i;
686 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
687 	struct mlx5_devx_tis_attr tis_attr = { 0 };
688 #endif
689 
690 	MLX5_ASSERT(spawn);
691 	/* Secondary process should not create the shared context. */
692 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
693 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
694 	/* Search for IB context by device name. */
695 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
696 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
697 			sh->refcnt++;
698 			goto exit;
699 		}
700 	}
701 	/* No device found, we have to create new shared context. */
702 	MLX5_ASSERT(spawn->max_port);
703 	sh = rte_zmalloc("ethdev shared ib context",
704 			 sizeof(struct mlx5_ibv_shared) +
705 			 spawn->max_port *
706 			 sizeof(struct mlx5_ibv_shared_port),
707 			 RTE_CACHE_LINE_SIZE);
708 	if (!sh) {
709 		DRV_LOG(ERR, "shared context allocation failure");
710 		rte_errno  = ENOMEM;
711 		goto exit;
712 	}
713 	/*
714 	 * Configure environment variable "MLX5_BF_SHUT_UP"
715 	 * before the device creation. The rdma_core library
716 	 * checks the variable at device creation and
717 	 * stores the result internally.
718 	 */
719 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
720 	/* Try to open IB device with DV first, then usual Verbs. */
721 	errno = 0;
722 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
723 	if (sh->ctx) {
724 		sh->devx = 1;
725 		DRV_LOG(DEBUG, "DevX is supported");
726 		/* The device is created, no need for environment. */
727 		mlx5_restore_doorbell_mapping_env(dbmap_env);
728 	} else {
729 		/* The environment variable is still configured. */
730 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
731 		err = errno ? errno : ENODEV;
732 		/*
733 		 * The environment variable is not needed anymore,
734 		 * all device creation attempts are completed.
735 		 */
736 		mlx5_restore_doorbell_mapping_env(dbmap_env);
737 		if (!sh->ctx)
738 			goto error;
739 		DRV_LOG(DEBUG, "DevX is NOT supported");
740 	}
741 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
742 	if (err) {
743 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
744 		goto error;
745 	}
746 	sh->refcnt = 1;
747 	sh->max_port = spawn->max_port;
748 	strncpy(sh->ibdev_name, sh->ctx->device->name,
749 		sizeof(sh->ibdev_name));
750 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
751 		sizeof(sh->ibdev_path));
752 	pthread_mutex_init(&sh->intr_mutex, NULL);
753 	/*
754 	 * Setting port_id to max unallowed value means
755 	 * there is no interrupt subhandler installed for
756 	 * the given port index i.
757 	 */
758 	for (i = 0; i < sh->max_port; i++) {
759 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
760 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
761 	}
762 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
763 	if (sh->pd == NULL) {
764 		DRV_LOG(ERR, "PD allocation failure");
765 		err = ENOMEM;
766 		goto error;
767 	}
768 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
769 	if (sh->devx) {
770 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
771 		if (err) {
772 			DRV_LOG(ERR, "Fail to extract pdn from PD");
773 			goto error;
774 		}
775 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
776 		if (!sh->td) {
777 			DRV_LOG(ERR, "TD allocation failure");
778 			err = ENOMEM;
779 			goto error;
780 		}
781 		tis_attr.transport_domain = sh->td->id;
782 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
783 		if (!sh->tis) {
784 			DRV_LOG(ERR, "TIS allocation failure");
785 			err = ENOMEM;
786 			goto error;
787 		}
788 	}
789 	sh->flow_id_pool = mlx5_flow_id_pool_alloc
790 					((1 << HAIRPIN_FLOW_ID_BITS) - 1);
791 	if (!sh->flow_id_pool) {
792 		DRV_LOG(ERR, "can't create flow id pool");
793 		err = ENOMEM;
794 		goto error;
795 	}
796 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
797 	/*
798 	 * Once the device is added to the list of memory event
799 	 * callback, its global MR cache table cannot be expanded
800 	 * on the fly because of deadlock. If it overflows, lookup
801 	 * should be done by searching MR list linearly, which is slow.
802 	 *
803 	 * At this point the device is not added to the memory
804 	 * event list yet, context is just being created.
805 	 */
806 	err = mlx5_mr_btree_init(&sh->share_cache.cache,
807 				 MLX5_MR_BTREE_CACHE_N * 2,
808 				 spawn->pci_dev->device.numa_node);
809 	if (err) {
810 		err = rte_errno;
811 		goto error;
812 	}
813 	mlx5_flow_aging_init(sh);
814 	mlx5_flow_counters_mng_init(sh);
815 	mlx5_flow_ipool_create(sh, config);
816 	/* Add device to memory callback list. */
817 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
818 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
819 			 sh, mem_event_cb);
820 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
821 	/* Add context to the global device list. */
822 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
823 exit:
824 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
825 	return sh;
826 error:
827 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
828 	MLX5_ASSERT(sh);
829 	if (sh->tis)
830 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
831 	if (sh->td)
832 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
833 	if (sh->pd)
834 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
835 	if (sh->ctx)
836 		claim_zero(mlx5_glue->close_device(sh->ctx));
837 	if (sh->flow_id_pool)
838 		mlx5_flow_id_pool_release(sh->flow_id_pool);
839 	rte_free(sh);
840 	MLX5_ASSERT(err > 0);
841 	rte_errno = err;
842 	return NULL;
843 }
844 
845 /**
846  * Free shared IB device context. Decrement counter and if zero free
847  * all allocated resources and close handles.
848  *
849  * @param[in] sh
850  *   Pointer to mlx5_ibv_shared object to free
851  */
852 static void
853 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
854 {
855 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
856 #ifdef RTE_LIBRTE_MLX5_DEBUG
857 	/* Check the object presence in the list. */
858 	struct mlx5_ibv_shared *lctx;
859 
860 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
861 		if (lctx == sh)
862 			break;
863 	MLX5_ASSERT(lctx);
864 	if (lctx != sh) {
865 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
866 		goto exit;
867 	}
868 #endif
869 	MLX5_ASSERT(sh);
870 	MLX5_ASSERT(sh->refcnt);
871 	/* Secondary process should not free the shared context. */
872 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
873 	if (--sh->refcnt)
874 		goto exit;
875 	/* Remove from memory callback device list. */
876 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
877 	LIST_REMOVE(sh, mem_event_cb);
878 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
879 	/* Release created Memory Regions. */
880 	mlx5_mr_release_cache(&sh->share_cache);
881 	/* Remove context from the global device list. */
882 	LIST_REMOVE(sh, next);
883 	/*
884 	 *  Ensure there is no async event handler installed.
885 	 *  Only primary process handles async device events.
886 	 **/
887 	mlx5_flow_counters_mng_close(sh);
888 	mlx5_flow_ipool_destroy(sh);
889 	MLX5_ASSERT(!sh->intr_cnt);
890 	if (sh->intr_cnt)
891 		mlx5_intr_callback_unregister
892 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
893 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
894 	if (sh->devx_intr_cnt) {
895 		if (sh->intr_handle_devx.fd)
896 			rte_intr_callback_unregister(&sh->intr_handle_devx,
897 					  mlx5_dev_interrupt_handler_devx, sh);
898 		if (sh->devx_comp)
899 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
900 	}
901 #endif
902 	pthread_mutex_destroy(&sh->intr_mutex);
903 	if (sh->pd)
904 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
905 	if (sh->tis)
906 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
907 	if (sh->td)
908 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
909 	if (sh->ctx)
910 		claim_zero(mlx5_glue->close_device(sh->ctx));
911 	if (sh->flow_id_pool)
912 		mlx5_flow_id_pool_release(sh->flow_id_pool);
913 	rte_free(sh);
914 exit:
915 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
916 }
917 
918 /**
919  * Destroy table hash list and all the root entries per domain.
920  *
921  * @param[in] priv
922  *   Pointer to the private device data structure.
923  */
924 static void
925 mlx5_free_table_hash_list(struct mlx5_priv *priv)
926 {
927 	struct mlx5_ibv_shared *sh = priv->sh;
928 	struct mlx5_flow_tbl_data_entry *tbl_data;
929 	union mlx5_flow_tbl_key table_key = {
930 		{
931 			.table_id = 0,
932 			.reserved = 0,
933 			.domain = 0,
934 			.direction = 0,
935 		}
936 	};
937 	struct mlx5_hlist_entry *pos;
938 
939 	if (!sh->flow_tbls)
940 		return;
941 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
942 	if (pos) {
943 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
944 					entry);
945 		MLX5_ASSERT(tbl_data);
946 		mlx5_hlist_remove(sh->flow_tbls, pos);
947 		rte_free(tbl_data);
948 	}
949 	table_key.direction = 1;
950 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
951 	if (pos) {
952 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
953 					entry);
954 		MLX5_ASSERT(tbl_data);
955 		mlx5_hlist_remove(sh->flow_tbls, pos);
956 		rte_free(tbl_data);
957 	}
958 	table_key.direction = 0;
959 	table_key.domain = 1;
960 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
961 	if (pos) {
962 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
963 					entry);
964 		MLX5_ASSERT(tbl_data);
965 		mlx5_hlist_remove(sh->flow_tbls, pos);
966 		rte_free(tbl_data);
967 	}
968 	mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
969 }
970 
971 /**
972  * Initialize flow table hash list and create the root tables entry
973  * for each domain.
974  *
975  * @param[in] priv
976  *   Pointer to the private device data structure.
977  *
978  * @return
979  *   Zero on success, positive error code otherwise.
980  */
981 static int
982 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
983 {
984 	struct mlx5_ibv_shared *sh = priv->sh;
985 	char s[MLX5_HLIST_NAMESIZE];
986 	int err = 0;
987 
988 	MLX5_ASSERT(sh);
989 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
990 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
991 	if (!sh->flow_tbls) {
992 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
993 		err = ENOMEM;
994 		return err;
995 	}
996 #ifndef HAVE_MLX5DV_DR
997 	/*
998 	 * In case we have not DR support, the zero tables should be created
999 	 * because DV expect to see them even if they cannot be created by
1000 	 * RDMA-CORE.
1001 	 */
1002 	union mlx5_flow_tbl_key table_key = {
1003 		{
1004 			.table_id = 0,
1005 			.reserved = 0,
1006 			.domain = 0,
1007 			.direction = 0,
1008 		}
1009 	};
1010 	struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
1011 							  sizeof(*tbl_data), 0);
1012 
1013 	if (!tbl_data) {
1014 		err = ENOMEM;
1015 		goto error;
1016 	}
1017 	tbl_data->entry.key = table_key.v64;
1018 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1019 	if (err)
1020 		goto error;
1021 	rte_atomic32_init(&tbl_data->tbl.refcnt);
1022 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
1023 	table_key.direction = 1;
1024 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
1025 	if (!tbl_data) {
1026 		err = ENOMEM;
1027 		goto error;
1028 	}
1029 	tbl_data->entry.key = table_key.v64;
1030 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1031 	if (err)
1032 		goto error;
1033 	rte_atomic32_init(&tbl_data->tbl.refcnt);
1034 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
1035 	table_key.direction = 0;
1036 	table_key.domain = 1;
1037 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
1038 	if (!tbl_data) {
1039 		err = ENOMEM;
1040 		goto error;
1041 	}
1042 	tbl_data->entry.key = table_key.v64;
1043 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1044 	if (err)
1045 		goto error;
1046 	rte_atomic32_init(&tbl_data->tbl.refcnt);
1047 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
1048 	return err;
1049 error:
1050 	mlx5_free_table_hash_list(priv);
1051 #endif /* HAVE_MLX5DV_DR */
1052 	return err;
1053 }
1054 
1055 /**
1056  * Initialize DR related data within private structure.
1057  * Routine checks the reference counter and does actual
1058  * resources creation/initialization only if counter is zero.
1059  *
1060  * @param[in] priv
1061  *   Pointer to the private device data structure.
1062  *
1063  * @return
1064  *   Zero on success, positive error code otherwise.
1065  */
1066 static int
1067 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
1068 {
1069 	struct mlx5_ibv_shared *sh = priv->sh;
1070 	char s[MLX5_HLIST_NAMESIZE];
1071 	int err = 0;
1072 
1073 	if (!sh->flow_tbls)
1074 		err = mlx5_alloc_table_hash_list(priv);
1075 	else
1076 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
1077 			(void *)sh->flow_tbls);
1078 	if (err)
1079 		return err;
1080 	/* Create tags hash list table. */
1081 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
1082 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
1083 	if (!sh->tag_table) {
1084 		DRV_LOG(ERR, "tags with hash creation failed.\n");
1085 		err = ENOMEM;
1086 		goto error;
1087 	}
1088 #ifdef HAVE_MLX5DV_DR
1089 	void *domain;
1090 
1091 	if (sh->dv_refcnt) {
1092 		/* Shared DV/DR structures is already initialized. */
1093 		sh->dv_refcnt++;
1094 		priv->dr_shared = 1;
1095 		return 0;
1096 	}
1097 	/* Reference counter is zero, we should initialize structures. */
1098 	domain = mlx5_glue->dr_create_domain(sh->ctx,
1099 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
1100 	if (!domain) {
1101 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
1102 		err = errno;
1103 		goto error;
1104 	}
1105 	sh->rx_domain = domain;
1106 	domain = mlx5_glue->dr_create_domain(sh->ctx,
1107 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
1108 	if (!domain) {
1109 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
1110 		err = errno;
1111 		goto error;
1112 	}
1113 	pthread_mutex_init(&sh->dv_mutex, NULL);
1114 	sh->tx_domain = domain;
1115 #ifdef HAVE_MLX5DV_DR_ESWITCH
1116 	if (priv->config.dv_esw_en) {
1117 		domain  = mlx5_glue->dr_create_domain
1118 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
1119 		if (!domain) {
1120 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
1121 			err = errno;
1122 			goto error;
1123 		}
1124 		sh->fdb_domain = domain;
1125 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
1126 	}
1127 #endif
1128 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
1129 #endif /* HAVE_MLX5DV_DR */
1130 	sh->dv_refcnt++;
1131 	priv->dr_shared = 1;
1132 	return 0;
1133 error:
1134 	/* Rollback the created objects. */
1135 	if (sh->rx_domain) {
1136 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
1137 		sh->rx_domain = NULL;
1138 	}
1139 	if (sh->tx_domain) {
1140 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
1141 		sh->tx_domain = NULL;
1142 	}
1143 	if (sh->fdb_domain) {
1144 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1145 		sh->fdb_domain = NULL;
1146 	}
1147 	if (sh->esw_drop_action) {
1148 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1149 		sh->esw_drop_action = NULL;
1150 	}
1151 	if (sh->pop_vlan_action) {
1152 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1153 		sh->pop_vlan_action = NULL;
1154 	}
1155 	if (sh->tag_table) {
1156 		/* tags should be destroyed with flow before. */
1157 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1158 		sh->tag_table = NULL;
1159 	}
1160 	mlx5_free_table_hash_list(priv);
1161 	return err;
1162 }
1163 
1164 /**
1165  * Destroy DR related data within private structure.
1166  *
1167  * @param[in] priv
1168  *   Pointer to the private device data structure.
1169  */
1170 static void
1171 mlx5_free_shared_dr(struct mlx5_priv *priv)
1172 {
1173 	struct mlx5_ibv_shared *sh;
1174 
1175 	if (!priv->dr_shared)
1176 		return;
1177 	priv->dr_shared = 0;
1178 	sh = priv->sh;
1179 	MLX5_ASSERT(sh);
1180 #ifdef HAVE_MLX5DV_DR
1181 	MLX5_ASSERT(sh->dv_refcnt);
1182 	if (sh->dv_refcnt && --sh->dv_refcnt)
1183 		return;
1184 	if (sh->rx_domain) {
1185 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
1186 		sh->rx_domain = NULL;
1187 	}
1188 	if (sh->tx_domain) {
1189 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
1190 		sh->tx_domain = NULL;
1191 	}
1192 #ifdef HAVE_MLX5DV_DR_ESWITCH
1193 	if (sh->fdb_domain) {
1194 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1195 		sh->fdb_domain = NULL;
1196 	}
1197 	if (sh->esw_drop_action) {
1198 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1199 		sh->esw_drop_action = NULL;
1200 	}
1201 #endif
1202 	if (sh->pop_vlan_action) {
1203 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1204 		sh->pop_vlan_action = NULL;
1205 	}
1206 	pthread_mutex_destroy(&sh->dv_mutex);
1207 #endif /* HAVE_MLX5DV_DR */
1208 	if (sh->tag_table) {
1209 		/* tags should be destroyed with flow before. */
1210 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1211 		sh->tag_table = NULL;
1212 	}
1213 	mlx5_free_table_hash_list(priv);
1214 }
1215 
1216 /**
1217  * Initialize shared data between primary and secondary process.
1218  *
1219  * A memzone is reserved by primary process and secondary processes attach to
1220  * the memzone.
1221  *
1222  * @return
1223  *   0 on success, a negative errno value otherwise and rte_errno is set.
1224  */
1225 static int
1226 mlx5_init_shared_data(void)
1227 {
1228 	const struct rte_memzone *mz;
1229 	int ret = 0;
1230 
1231 	rte_spinlock_lock(&mlx5_shared_data_lock);
1232 	if (mlx5_shared_data == NULL) {
1233 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1234 			/* Allocate shared memory. */
1235 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1236 						 sizeof(*mlx5_shared_data),
1237 						 SOCKET_ID_ANY, 0);
1238 			if (mz == NULL) {
1239 				DRV_LOG(ERR,
1240 					"Cannot allocate mlx5 shared data");
1241 				ret = -rte_errno;
1242 				goto error;
1243 			}
1244 			mlx5_shared_data = mz->addr;
1245 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1246 			rte_spinlock_init(&mlx5_shared_data->lock);
1247 		} else {
1248 			/* Lookup allocated shared memory. */
1249 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1250 			if (mz == NULL) {
1251 				DRV_LOG(ERR,
1252 					"Cannot attach mlx5 shared data");
1253 				ret = -rte_errno;
1254 				goto error;
1255 			}
1256 			mlx5_shared_data = mz->addr;
1257 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1258 		}
1259 	}
1260 error:
1261 	rte_spinlock_unlock(&mlx5_shared_data_lock);
1262 	return ret;
1263 }
1264 
1265 /**
1266  * Retrieve integer value from environment variable.
1267  *
1268  * @param[in] name
1269  *   Environment variable name.
1270  *
1271  * @return
1272  *   Integer value, 0 if the variable is not set.
1273  */
1274 int
1275 mlx5_getenv_int(const char *name)
1276 {
1277 	const char *val = getenv(name);
1278 
1279 	if (val == NULL)
1280 		return 0;
1281 	return atoi(val);
1282 }
1283 
1284 /**
1285  * Verbs callback to allocate a memory. This function should allocate the space
1286  * according to the size provided residing inside a huge page.
1287  * Please note that all allocation must respect the alignment from libmlx5
1288  * (i.e. currently sysconf(_SC_PAGESIZE)).
1289  *
1290  * @param[in] size
1291  *   The size in bytes of the memory to allocate.
1292  * @param[in] data
1293  *   A pointer to the callback data.
1294  *
1295  * @return
1296  *   Allocated buffer, NULL otherwise and rte_errno is set.
1297  */
1298 static void *
1299 mlx5_alloc_verbs_buf(size_t size, void *data)
1300 {
1301 	struct mlx5_priv *priv = data;
1302 	void *ret;
1303 	size_t alignment = sysconf(_SC_PAGESIZE);
1304 	unsigned int socket = SOCKET_ID_ANY;
1305 
1306 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1307 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1308 
1309 		socket = ctrl->socket;
1310 	} else if (priv->verbs_alloc_ctx.type ==
1311 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1312 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1313 
1314 		socket = ctrl->socket;
1315 	}
1316 	MLX5_ASSERT(data != NULL);
1317 	ret = rte_malloc_socket(__func__, size, alignment, socket);
1318 	if (!ret && size)
1319 		rte_errno = ENOMEM;
1320 	return ret;
1321 }
1322 
1323 /**
1324  * Verbs callback to free a memory.
1325  *
1326  * @param[in] ptr
1327  *   A pointer to the memory to free.
1328  * @param[in] data
1329  *   A pointer to the callback data.
1330  */
1331 static void
1332 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1333 {
1334 	MLX5_ASSERT(data != NULL);
1335 	rte_free(ptr);
1336 }
1337 
1338 /**
1339  * DPDK callback to add udp tunnel port
1340  *
1341  * @param[in] dev
1342  *   A pointer to eth_dev
1343  * @param[in] udp_tunnel
1344  *   A pointer to udp tunnel
1345  *
1346  * @return
1347  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1348  */
1349 int
1350 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1351 			 struct rte_eth_udp_tunnel *udp_tunnel)
1352 {
1353 	MLX5_ASSERT(udp_tunnel != NULL);
1354 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1355 	    udp_tunnel->udp_port == 4789)
1356 		return 0;
1357 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1358 	    udp_tunnel->udp_port == 4790)
1359 		return 0;
1360 	return -ENOTSUP;
1361 }
1362 
1363 /**
1364  * Initialize process private data structure.
1365  *
1366  * @param dev
1367  *   Pointer to Ethernet device structure.
1368  *
1369  * @return
1370  *   0 on success, a negative errno value otherwise and rte_errno is set.
1371  */
1372 int
1373 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1374 {
1375 	struct mlx5_priv *priv = dev->data->dev_private;
1376 	struct mlx5_proc_priv *ppriv;
1377 	size_t ppriv_size;
1378 
1379 	/*
1380 	 * UAR register table follows the process private structure. BlueFlame
1381 	 * registers for Tx queues are stored in the table.
1382 	 */
1383 	ppriv_size =
1384 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1385 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1386 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1387 	if (!ppriv) {
1388 		rte_errno = ENOMEM;
1389 		return -rte_errno;
1390 	}
1391 	ppriv->uar_table_sz = ppriv_size;
1392 	dev->process_private = ppriv;
1393 	return 0;
1394 }
1395 
1396 /**
1397  * Un-initialize process private data structure.
1398  *
1399  * @param dev
1400  *   Pointer to Ethernet device structure.
1401  */
1402 static void
1403 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1404 {
1405 	if (!dev->process_private)
1406 		return;
1407 	rte_free(dev->process_private);
1408 	dev->process_private = NULL;
1409 }
1410 
1411 /**
1412  * DPDK callback to close the device.
1413  *
1414  * Destroy all queues and objects, free memory.
1415  *
1416  * @param dev
1417  *   Pointer to Ethernet device structure.
1418  */
1419 static void
1420 mlx5_dev_close(struct rte_eth_dev *dev)
1421 {
1422 	struct mlx5_priv *priv = dev->data->dev_private;
1423 	unsigned int i;
1424 	int ret;
1425 
1426 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1427 		dev->data->port_id,
1428 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1429 	/* In case mlx5_dev_stop() has not been called. */
1430 	mlx5_dev_interrupt_handler_uninstall(dev);
1431 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1432 	/*
1433 	 * If default mreg copy action is removed at the stop stage,
1434 	 * the search will return none and nothing will be done anymore.
1435 	 */
1436 	mlx5_flow_stop_default(dev);
1437 	mlx5_traffic_disable(dev);
1438 	/*
1439 	 * If all the flows are already flushed in the device stop stage,
1440 	 * then this will return directly without any action.
1441 	 */
1442 	mlx5_flow_list_flush(dev, &priv->flows, true);
1443 	mlx5_flow_meter_flush(dev, NULL);
1444 	/* Free the intermediate buffers for flow creation. */
1445 	mlx5_flow_free_intermediate(dev);
1446 	/* Prevent crashes when queues are still in use. */
1447 	dev->rx_pkt_burst = removed_rx_burst;
1448 	dev->tx_pkt_burst = removed_tx_burst;
1449 	rte_wmb();
1450 	/* Disable datapath on secondary process. */
1451 	mlx5_mp_req_stop_rxtx(dev);
1452 	if (priv->rxqs != NULL) {
1453 		/* XXX race condition if mlx5_rx_burst() is still running. */
1454 		usleep(1000);
1455 		for (i = 0; (i != priv->rxqs_n); ++i)
1456 			mlx5_rxq_release(dev, i);
1457 		priv->rxqs_n = 0;
1458 		priv->rxqs = NULL;
1459 	}
1460 	if (priv->txqs != NULL) {
1461 		/* XXX race condition if mlx5_tx_burst() is still running. */
1462 		usleep(1000);
1463 		for (i = 0; (i != priv->txqs_n); ++i)
1464 			mlx5_txq_release(dev, i);
1465 		priv->txqs_n = 0;
1466 		priv->txqs = NULL;
1467 	}
1468 	mlx5_proc_priv_uninit(dev);
1469 	if (priv->mreg_cp_tbl)
1470 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1471 	mlx5_mprq_free_mp(dev);
1472 	mlx5_free_shared_dr(priv);
1473 	if (priv->rss_conf.rss_key != NULL)
1474 		rte_free(priv->rss_conf.rss_key);
1475 	if (priv->reta_idx != NULL)
1476 		rte_free(priv->reta_idx);
1477 	if (priv->config.vf)
1478 		mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1479 				       dev->data->mac_addrs,
1480 				       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1481 	if (priv->nl_socket_route >= 0)
1482 		close(priv->nl_socket_route);
1483 	if (priv->nl_socket_rdma >= 0)
1484 		close(priv->nl_socket_rdma);
1485 	if (priv->vmwa_context)
1486 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1487 	ret = mlx5_hrxq_verify(dev);
1488 	if (ret)
1489 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1490 			dev->data->port_id);
1491 	ret = mlx5_ind_table_obj_verify(dev);
1492 	if (ret)
1493 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1494 			dev->data->port_id);
1495 	ret = mlx5_rxq_obj_verify(dev);
1496 	if (ret)
1497 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1498 			dev->data->port_id);
1499 	ret = mlx5_rxq_verify(dev);
1500 	if (ret)
1501 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1502 			dev->data->port_id);
1503 	ret = mlx5_txq_obj_verify(dev);
1504 	if (ret)
1505 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1506 			dev->data->port_id);
1507 	ret = mlx5_txq_verify(dev);
1508 	if (ret)
1509 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1510 			dev->data->port_id);
1511 	ret = mlx5_flow_verify(dev);
1512 	if (ret)
1513 		DRV_LOG(WARNING, "port %u some flows still remain",
1514 			dev->data->port_id);
1515 	if (priv->sh) {
1516 		/*
1517 		 * Free the shared context in last turn, because the cleanup
1518 		 * routines above may use some shared fields, like
1519 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1520 		 * ifindex if Netlink fails.
1521 		 */
1522 		mlx5_free_shared_ibctx(priv->sh);
1523 		priv->sh = NULL;
1524 	}
1525 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1526 		unsigned int c = 0;
1527 		uint16_t port_id;
1528 
1529 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1530 			struct mlx5_priv *opriv =
1531 				rte_eth_devices[port_id].data->dev_private;
1532 
1533 			if (!opriv ||
1534 			    opriv->domain_id != priv->domain_id ||
1535 			    &rte_eth_devices[port_id] == dev)
1536 				continue;
1537 			++c;
1538 			break;
1539 		}
1540 		if (!c)
1541 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1542 	}
1543 	memset(priv, 0, sizeof(*priv));
1544 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1545 	/*
1546 	 * Reset mac_addrs to NULL such that it is not freed as part of
1547 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1548 	 * it is freed when dev_private is freed.
1549 	 */
1550 	dev->data->mac_addrs = NULL;
1551 }
1552 
1553 const struct eth_dev_ops mlx5_dev_ops = {
1554 	.dev_configure = mlx5_dev_configure,
1555 	.dev_start = mlx5_dev_start,
1556 	.dev_stop = mlx5_dev_stop,
1557 	.dev_set_link_down = mlx5_set_link_down,
1558 	.dev_set_link_up = mlx5_set_link_up,
1559 	.dev_close = mlx5_dev_close,
1560 	.promiscuous_enable = mlx5_promiscuous_enable,
1561 	.promiscuous_disable = mlx5_promiscuous_disable,
1562 	.allmulticast_enable = mlx5_allmulticast_enable,
1563 	.allmulticast_disable = mlx5_allmulticast_disable,
1564 	.link_update = mlx5_link_update,
1565 	.stats_get = mlx5_stats_get,
1566 	.stats_reset = mlx5_stats_reset,
1567 	.xstats_get = mlx5_xstats_get,
1568 	.xstats_reset = mlx5_xstats_reset,
1569 	.xstats_get_names = mlx5_xstats_get_names,
1570 	.fw_version_get = mlx5_fw_version_get,
1571 	.dev_infos_get = mlx5_dev_infos_get,
1572 	.read_clock = mlx5_read_clock,
1573 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1574 	.vlan_filter_set = mlx5_vlan_filter_set,
1575 	.rx_queue_setup = mlx5_rx_queue_setup,
1576 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1577 	.tx_queue_setup = mlx5_tx_queue_setup,
1578 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1579 	.rx_queue_release = mlx5_rx_queue_release,
1580 	.tx_queue_release = mlx5_tx_queue_release,
1581 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1582 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1583 	.mac_addr_remove = mlx5_mac_addr_remove,
1584 	.mac_addr_add = mlx5_mac_addr_add,
1585 	.mac_addr_set = mlx5_mac_addr_set,
1586 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1587 	.mtu_set = mlx5_dev_set_mtu,
1588 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1589 	.vlan_offload_set = mlx5_vlan_offload_set,
1590 	.reta_update = mlx5_dev_rss_reta_update,
1591 	.reta_query = mlx5_dev_rss_reta_query,
1592 	.rss_hash_update = mlx5_rss_hash_update,
1593 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1594 	.filter_ctrl = mlx5_dev_filter_ctrl,
1595 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1596 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1597 	.rxq_info_get = mlx5_rxq_info_get,
1598 	.txq_info_get = mlx5_txq_info_get,
1599 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1600 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1601 	.rx_queue_count = mlx5_rx_queue_count,
1602 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1603 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1604 	.is_removed = mlx5_is_removed,
1605 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1606 	.get_module_info = mlx5_get_module_info,
1607 	.get_module_eeprom = mlx5_get_module_eeprom,
1608 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1609 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1610 };
1611 
1612 /* Available operations from secondary process. */
1613 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1614 	.stats_get = mlx5_stats_get,
1615 	.stats_reset = mlx5_stats_reset,
1616 	.xstats_get = mlx5_xstats_get,
1617 	.xstats_reset = mlx5_xstats_reset,
1618 	.xstats_get_names = mlx5_xstats_get_names,
1619 	.fw_version_get = mlx5_fw_version_get,
1620 	.dev_infos_get = mlx5_dev_infos_get,
1621 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1622 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1623 	.rxq_info_get = mlx5_rxq_info_get,
1624 	.txq_info_get = mlx5_txq_info_get,
1625 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1626 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1627 	.get_module_info = mlx5_get_module_info,
1628 	.get_module_eeprom = mlx5_get_module_eeprom,
1629 };
1630 
1631 /* Available operations in flow isolated mode. */
1632 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1633 	.dev_configure = mlx5_dev_configure,
1634 	.dev_start = mlx5_dev_start,
1635 	.dev_stop = mlx5_dev_stop,
1636 	.dev_set_link_down = mlx5_set_link_down,
1637 	.dev_set_link_up = mlx5_set_link_up,
1638 	.dev_close = mlx5_dev_close,
1639 	.promiscuous_enable = mlx5_promiscuous_enable,
1640 	.promiscuous_disable = mlx5_promiscuous_disable,
1641 	.allmulticast_enable = mlx5_allmulticast_enable,
1642 	.allmulticast_disable = mlx5_allmulticast_disable,
1643 	.link_update = mlx5_link_update,
1644 	.stats_get = mlx5_stats_get,
1645 	.stats_reset = mlx5_stats_reset,
1646 	.xstats_get = mlx5_xstats_get,
1647 	.xstats_reset = mlx5_xstats_reset,
1648 	.xstats_get_names = mlx5_xstats_get_names,
1649 	.fw_version_get = mlx5_fw_version_get,
1650 	.dev_infos_get = mlx5_dev_infos_get,
1651 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1652 	.vlan_filter_set = mlx5_vlan_filter_set,
1653 	.rx_queue_setup = mlx5_rx_queue_setup,
1654 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1655 	.tx_queue_setup = mlx5_tx_queue_setup,
1656 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1657 	.rx_queue_release = mlx5_rx_queue_release,
1658 	.tx_queue_release = mlx5_tx_queue_release,
1659 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1660 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1661 	.mac_addr_remove = mlx5_mac_addr_remove,
1662 	.mac_addr_add = mlx5_mac_addr_add,
1663 	.mac_addr_set = mlx5_mac_addr_set,
1664 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1665 	.mtu_set = mlx5_dev_set_mtu,
1666 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1667 	.vlan_offload_set = mlx5_vlan_offload_set,
1668 	.filter_ctrl = mlx5_dev_filter_ctrl,
1669 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1670 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1671 	.rxq_info_get = mlx5_rxq_info_get,
1672 	.txq_info_get = mlx5_txq_info_get,
1673 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1674 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1675 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1676 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1677 	.is_removed = mlx5_is_removed,
1678 	.get_module_info = mlx5_get_module_info,
1679 	.get_module_eeprom = mlx5_get_module_eeprom,
1680 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1681 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1682 };
1683 
1684 /**
1685  * Verify and store value for device argument.
1686  *
1687  * @param[in] key
1688  *   Key argument to verify.
1689  * @param[in] val
1690  *   Value associated with key.
1691  * @param opaque
1692  *   User data.
1693  *
1694  * @return
1695  *   0 on success, a negative errno value otherwise and rte_errno is set.
1696  */
1697 static int
1698 mlx5_args_check(const char *key, const char *val, void *opaque)
1699 {
1700 	struct mlx5_dev_config *config = opaque;
1701 	unsigned long tmp;
1702 
1703 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1704 	if (!strcmp(MLX5_REPRESENTOR, key))
1705 		return 0;
1706 	errno = 0;
1707 	tmp = strtoul(val, NULL, 0);
1708 	if (errno) {
1709 		rte_errno = errno;
1710 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1711 		return -rte_errno;
1712 	}
1713 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1714 		config->cqe_comp = !!tmp;
1715 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1716 		config->cqe_pad = !!tmp;
1717 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1718 		config->hw_padding = !!tmp;
1719 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1720 		config->mprq.enabled = !!tmp;
1721 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1722 		config->mprq.stride_num_n = tmp;
1723 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1724 		config->mprq.stride_size_n = tmp;
1725 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1726 		config->mprq.max_memcpy_len = tmp;
1727 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1728 		config->mprq.min_rxqs_num = tmp;
1729 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1730 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1731 				 " converted to txq_inline_max", key);
1732 		config->txq_inline_max = tmp;
1733 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1734 		config->txq_inline_max = tmp;
1735 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1736 		config->txq_inline_min = tmp;
1737 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1738 		config->txq_inline_mpw = tmp;
1739 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1740 		config->txqs_inline = tmp;
1741 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1742 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1743 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1744 		config->mps = !!tmp;
1745 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1746 		if (tmp != MLX5_TXDB_CACHED &&
1747 		    tmp != MLX5_TXDB_NCACHED &&
1748 		    tmp != MLX5_TXDB_HEURISTIC) {
1749 			DRV_LOG(ERR, "invalid Tx doorbell "
1750 				     "mapping parameter");
1751 			rte_errno = EINVAL;
1752 			return -rte_errno;
1753 		}
1754 		config->dbnc = tmp;
1755 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1756 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1757 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1758 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1759 				 " converted to txq_inline_mpw", key);
1760 		config->txq_inline_mpw = tmp;
1761 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1762 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1763 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1764 		config->rx_vec_en = !!tmp;
1765 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1766 		config->l3_vxlan_en = !!tmp;
1767 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1768 		config->vf_nl_en = !!tmp;
1769 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1770 		config->dv_esw_en = !!tmp;
1771 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1772 		config->dv_flow_en = !!tmp;
1773 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1774 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1775 		    tmp != MLX5_XMETA_MODE_META16 &&
1776 		    tmp != MLX5_XMETA_MODE_META32) {
1777 			DRV_LOG(ERR, "invalid extensive "
1778 				     "metadata parameter");
1779 			rte_errno = EINVAL;
1780 			return -rte_errno;
1781 		}
1782 		config->dv_xmeta_en = tmp;
1783 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1784 		config->mr_ext_memseg_en = !!tmp;
1785 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1786 		config->max_dump_files_num = tmp;
1787 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1788 		config->lro.timeout = tmp;
1789 	} else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1790 		DRV_LOG(DEBUG, "class argument is %s.", val);
1791 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1792 		config->log_hp_size = tmp;
1793 	} else {
1794 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1795 		rte_errno = EINVAL;
1796 		return -rte_errno;
1797 	}
1798 	return 0;
1799 }
1800 
1801 /**
1802  * Parse device parameters.
1803  *
1804  * @param config
1805  *   Pointer to device configuration structure.
1806  * @param devargs
1807  *   Device arguments structure.
1808  *
1809  * @return
1810  *   0 on success, a negative errno value otherwise and rte_errno is set.
1811  */
1812 static int
1813 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1814 {
1815 	const char **params = (const char *[]){
1816 		MLX5_RXQ_CQE_COMP_EN,
1817 		MLX5_RXQ_CQE_PAD_EN,
1818 		MLX5_RXQ_PKT_PAD_EN,
1819 		MLX5_RX_MPRQ_EN,
1820 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1821 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1822 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1823 		MLX5_RXQS_MIN_MPRQ,
1824 		MLX5_TXQ_INLINE,
1825 		MLX5_TXQ_INLINE_MIN,
1826 		MLX5_TXQ_INLINE_MAX,
1827 		MLX5_TXQ_INLINE_MPW,
1828 		MLX5_TXQS_MIN_INLINE,
1829 		MLX5_TXQS_MAX_VEC,
1830 		MLX5_TXQ_MPW_EN,
1831 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1832 		MLX5_TXQ_MAX_INLINE_LEN,
1833 		MLX5_TX_DB_NC,
1834 		MLX5_TX_VEC_EN,
1835 		MLX5_RX_VEC_EN,
1836 		MLX5_L3_VXLAN_EN,
1837 		MLX5_VF_NL_EN,
1838 		MLX5_DV_ESW_EN,
1839 		MLX5_DV_FLOW_EN,
1840 		MLX5_DV_XMETA_EN,
1841 		MLX5_MR_EXT_MEMSEG_EN,
1842 		MLX5_REPRESENTOR,
1843 		MLX5_MAX_DUMP_FILES_NUM,
1844 		MLX5_LRO_TIMEOUT_USEC,
1845 		MLX5_CLASS_ARG_NAME,
1846 		MLX5_HP_BUF_SIZE,
1847 		NULL,
1848 	};
1849 	struct rte_kvargs *kvlist;
1850 	int ret = 0;
1851 	int i;
1852 
1853 	if (devargs == NULL)
1854 		return 0;
1855 	/* Following UGLY cast is done to pass checkpatch. */
1856 	kvlist = rte_kvargs_parse(devargs->args, params);
1857 	if (kvlist == NULL) {
1858 		rte_errno = EINVAL;
1859 		return -rte_errno;
1860 	}
1861 	/* Process parameters. */
1862 	for (i = 0; (params[i] != NULL); ++i) {
1863 		if (rte_kvargs_count(kvlist, params[i])) {
1864 			ret = rte_kvargs_process(kvlist, params[i],
1865 						 mlx5_args_check, config);
1866 			if (ret) {
1867 				rte_errno = EINVAL;
1868 				rte_kvargs_free(kvlist);
1869 				return -rte_errno;
1870 			}
1871 		}
1872 	}
1873 	rte_kvargs_free(kvlist);
1874 	return 0;
1875 }
1876 
1877 static struct rte_pci_driver mlx5_driver;
1878 
1879 /**
1880  * PMD global initialization.
1881  *
1882  * Independent from individual device, this function initializes global
1883  * per-PMD data structures distinguishing primary and secondary processes.
1884  * Hence, each initialization is called once per a process.
1885  *
1886  * @return
1887  *   0 on success, a negative errno value otherwise and rte_errno is set.
1888  */
1889 static int
1890 mlx5_init_once(void)
1891 {
1892 	struct mlx5_shared_data *sd;
1893 	struct mlx5_local_data *ld = &mlx5_local_data;
1894 	int ret = 0;
1895 
1896 	if (mlx5_init_shared_data())
1897 		return -rte_errno;
1898 	sd = mlx5_shared_data;
1899 	MLX5_ASSERT(sd);
1900 	rte_spinlock_lock(&sd->lock);
1901 	switch (rte_eal_process_type()) {
1902 	case RTE_PROC_PRIMARY:
1903 		if (sd->init_done)
1904 			break;
1905 		LIST_INIT(&sd->mem_event_cb_list);
1906 		rte_rwlock_init(&sd->mem_event_rwlock);
1907 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1908 						mlx5_mr_mem_event_cb, NULL);
1909 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
1910 					   mlx5_mp_primary_handle);
1911 		if (ret)
1912 			goto out;
1913 		sd->init_done = true;
1914 		break;
1915 	case RTE_PROC_SECONDARY:
1916 		if (ld->init_done)
1917 			break;
1918 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
1919 					     mlx5_mp_secondary_handle);
1920 		if (ret)
1921 			goto out;
1922 		++sd->secondary_cnt;
1923 		ld->init_done = true;
1924 		break;
1925 	default:
1926 		break;
1927 	}
1928 out:
1929 	rte_spinlock_unlock(&sd->lock);
1930 	return ret;
1931 }
1932 
1933 /**
1934  * Configures the minimal amount of data to inline into WQE
1935  * while sending packets.
1936  *
1937  * - the txq_inline_min has the maximal priority, if this
1938  *   key is specified in devargs
1939  * - if DevX is enabled the inline mode is queried from the
1940  *   device (HCA attributes and NIC vport context if needed).
1941  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1942  *   and none (0 bytes) for other NICs
1943  *
1944  * @param spawn
1945  *   Verbs device parameters (name, port, switch_info) to spawn.
1946  * @param config
1947  *   Device configuration parameters.
1948  */
1949 static void
1950 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1951 		    struct mlx5_dev_config *config)
1952 {
1953 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1954 		/* Application defines size of inlined data explicitly. */
1955 		switch (spawn->pci_dev->id.device_id) {
1956 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1957 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1958 			if (config->txq_inline_min <
1959 				       (int)MLX5_INLINE_HSIZE_L2) {
1960 				DRV_LOG(DEBUG,
1961 					"txq_inline_mix aligned to minimal"
1962 					" ConnectX-4 required value %d",
1963 					(int)MLX5_INLINE_HSIZE_L2);
1964 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1965 			}
1966 			break;
1967 		}
1968 		goto exit;
1969 	}
1970 	if (config->hca_attr.eth_net_offloads) {
1971 		/* We have DevX enabled, inline mode queried successfully. */
1972 		switch (config->hca_attr.wqe_inline_mode) {
1973 		case MLX5_CAP_INLINE_MODE_L2:
1974 			/* outer L2 header must be inlined. */
1975 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1976 			goto exit;
1977 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1978 			/* No inline data are required by NIC. */
1979 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1980 			config->hw_vlan_insert =
1981 				config->hca_attr.wqe_vlan_insert;
1982 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1983 			goto exit;
1984 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1985 			/* inline mode is defined by NIC vport context. */
1986 			if (!config->hca_attr.eth_virt)
1987 				break;
1988 			switch (config->hca_attr.vport_inline_mode) {
1989 			case MLX5_INLINE_MODE_NONE:
1990 				config->txq_inline_min =
1991 					MLX5_INLINE_HSIZE_NONE;
1992 				goto exit;
1993 			case MLX5_INLINE_MODE_L2:
1994 				config->txq_inline_min =
1995 					MLX5_INLINE_HSIZE_L2;
1996 				goto exit;
1997 			case MLX5_INLINE_MODE_IP:
1998 				config->txq_inline_min =
1999 					MLX5_INLINE_HSIZE_L3;
2000 				goto exit;
2001 			case MLX5_INLINE_MODE_TCP_UDP:
2002 				config->txq_inline_min =
2003 					MLX5_INLINE_HSIZE_L4;
2004 				goto exit;
2005 			case MLX5_INLINE_MODE_INNER_L2:
2006 				config->txq_inline_min =
2007 					MLX5_INLINE_HSIZE_INNER_L2;
2008 				goto exit;
2009 			case MLX5_INLINE_MODE_INNER_IP:
2010 				config->txq_inline_min =
2011 					MLX5_INLINE_HSIZE_INNER_L3;
2012 				goto exit;
2013 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
2014 				config->txq_inline_min =
2015 					MLX5_INLINE_HSIZE_INNER_L4;
2016 				goto exit;
2017 			}
2018 		}
2019 	}
2020 	/*
2021 	 * We get here if we are unable to deduce
2022 	 * inline data size with DevX. Try PCI ID
2023 	 * to determine old NICs.
2024 	 */
2025 	switch (spawn->pci_dev->id.device_id) {
2026 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2027 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2028 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2029 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2030 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2031 		config->hw_vlan_insert = 0;
2032 		break;
2033 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2034 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2035 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2036 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2037 		/*
2038 		 * These NICs support VLAN insertion from WQE and
2039 		 * report the wqe_vlan_insert flag. But there is the bug
2040 		 * and PFC control may be broken, so disable feature.
2041 		 */
2042 		config->hw_vlan_insert = 0;
2043 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2044 		break;
2045 	default:
2046 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2047 		break;
2048 	}
2049 exit:
2050 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2051 }
2052 
2053 /**
2054  * Configures the metadata mask fields in the shared context.
2055  *
2056  * @param [in] dev
2057  *   Pointer to Ethernet device.
2058  */
2059 static void
2060 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2061 {
2062 	struct mlx5_priv *priv = dev->data->dev_private;
2063 	struct mlx5_ibv_shared *sh = priv->sh;
2064 	uint32_t meta, mark, reg_c0;
2065 
2066 	reg_c0 = ~priv->vport_meta_mask;
2067 	switch (priv->config.dv_xmeta_en) {
2068 	case MLX5_XMETA_MODE_LEGACY:
2069 		meta = UINT32_MAX;
2070 		mark = MLX5_FLOW_MARK_MASK;
2071 		break;
2072 	case MLX5_XMETA_MODE_META16:
2073 		meta = reg_c0 >> rte_bsf32(reg_c0);
2074 		mark = MLX5_FLOW_MARK_MASK;
2075 		break;
2076 	case MLX5_XMETA_MODE_META32:
2077 		meta = UINT32_MAX;
2078 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2079 		break;
2080 	default:
2081 		meta = 0;
2082 		mark = 0;
2083 		MLX5_ASSERT(false);
2084 		break;
2085 	}
2086 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2087 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2088 				 sh->dv_mark_mask, mark);
2089 	else
2090 		sh->dv_mark_mask = mark;
2091 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2092 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2093 				 sh->dv_meta_mask, meta);
2094 	else
2095 		sh->dv_meta_mask = meta;
2096 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2097 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2098 				 sh->dv_meta_mask, reg_c0);
2099 	else
2100 		sh->dv_regc0_mask = reg_c0;
2101 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2102 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2103 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2104 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2105 }
2106 
2107 /**
2108  * Allocate page of door-bells and register it using DevX API.
2109  *
2110  * @param [in] dev
2111  *   Pointer to Ethernet device.
2112  *
2113  * @return
2114  *   Pointer to new page on success, NULL otherwise.
2115  */
2116 static struct mlx5_devx_dbr_page *
2117 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
2118 {
2119 	struct mlx5_priv *priv = dev->data->dev_private;
2120 	struct mlx5_devx_dbr_page *page;
2121 
2122 	/* Allocate space for door-bell page and management data. */
2123 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
2124 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
2125 	if (!page) {
2126 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
2127 			dev->data->port_id);
2128 		return NULL;
2129 	}
2130 	/* Register allocated memory. */
2131 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
2132 					      MLX5_DBR_PAGE_SIZE, 0);
2133 	if (!page->umem) {
2134 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
2135 			dev->data->port_id);
2136 		rte_free(page);
2137 		return NULL;
2138 	}
2139 	return page;
2140 }
2141 
2142 /**
2143  * Find the next available door-bell, allocate new page if needed.
2144  *
2145  * @param [in] dev
2146  *   Pointer to Ethernet device.
2147  * @param [out] dbr_page
2148  *   Door-bell page containing the page data.
2149  *
2150  * @return
2151  *   Door-bell address offset on success, a negative error value otherwise.
2152  */
2153 int64_t
2154 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
2155 {
2156 	struct mlx5_priv *priv = dev->data->dev_private;
2157 	struct mlx5_devx_dbr_page *page = NULL;
2158 	uint32_t i, j;
2159 
2160 	LIST_FOREACH(page, &priv->dbrpgs, next)
2161 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
2162 			break;
2163 	if (!page) { /* No page with free door-bell exists. */
2164 		page = mlx5_alloc_dbr_page(dev);
2165 		if (!page) /* Failed to allocate new page. */
2166 			return (-1);
2167 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
2168 	}
2169 	/* Loop to find bitmap part with clear bit. */
2170 	for (i = 0;
2171 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
2172 	     i++)
2173 		; /* Empty. */
2174 	/* Find the first clear bit. */
2175 	MLX5_ASSERT(i < MLX5_DBR_BITMAP_SIZE);
2176 	j = rte_bsf64(~page->dbr_bitmap[i]);
2177 	page->dbr_bitmap[i] |= (UINT64_C(1) << j);
2178 	page->dbr_count++;
2179 	*dbr_page = page;
2180 	return (((i * 64) + j) * sizeof(uint64_t));
2181 }
2182 
2183 /**
2184  * Release a door-bell record.
2185  *
2186  * @param [in] dev
2187  *   Pointer to Ethernet device.
2188  * @param [in] umem_id
2189  *   UMEM ID of page containing the door-bell record to release.
2190  * @param [in] offset
2191  *   Offset of door-bell record in page.
2192  *
2193  * @return
2194  *   0 on success, a negative error value otherwise.
2195  */
2196 int32_t
2197 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
2198 {
2199 	struct mlx5_priv *priv = dev->data->dev_private;
2200 	struct mlx5_devx_dbr_page *page = NULL;
2201 	int ret = 0;
2202 
2203 	LIST_FOREACH(page, &priv->dbrpgs, next)
2204 		/* Find the page this address belongs to. */
2205 		if (page->umem->umem_id == umem_id)
2206 			break;
2207 	if (!page)
2208 		return -EINVAL;
2209 	page->dbr_count--;
2210 	if (!page->dbr_count) {
2211 		/* Page not used, free it and remove from list. */
2212 		LIST_REMOVE(page, next);
2213 		if (page->umem)
2214 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
2215 		rte_free(page);
2216 	} else {
2217 		/* Mark in bitmap that this door-bell is not in use. */
2218 		offset /= MLX5_DBR_SIZE;
2219 		int i = offset / 64;
2220 		int j = offset % 64;
2221 
2222 		page->dbr_bitmap[i] &= ~(UINT64_C(1) << j);
2223 	}
2224 	return ret;
2225 }
2226 
2227 int
2228 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2229 {
2230 	static const char *const dynf_names[] = {
2231 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2232 		RTE_MBUF_DYNFLAG_METADATA_NAME
2233 	};
2234 	unsigned int i;
2235 
2236 	if (n < RTE_DIM(dynf_names))
2237 		return -ENOMEM;
2238 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2239 		if (names[i] == NULL)
2240 			return -EINVAL;
2241 		strcpy(names[i], dynf_names[i]);
2242 	}
2243 	return RTE_DIM(dynf_names);
2244 }
2245 
2246 /**
2247  * Check sibling device configurations.
2248  *
2249  * Sibling devices sharing the Infiniband device context
2250  * should have compatible configurations. This regards
2251  * representors and bonding slaves.
2252  *
2253  * @param priv
2254  *   Private device descriptor.
2255  * @param config
2256  *   Configuration of the device is going to be created.
2257  *
2258  * @return
2259  *   0 on success, EINVAL otherwise
2260  */
2261 static int
2262 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2263 			      struct mlx5_dev_config *config)
2264 {
2265 	struct mlx5_ibv_shared *sh = priv->sh;
2266 	struct mlx5_dev_config *sh_conf = NULL;
2267 	uint16_t port_id;
2268 
2269 	MLX5_ASSERT(sh);
2270 	/* Nothing to compare for the single/first device. */
2271 	if (sh->refcnt == 1)
2272 		return 0;
2273 	/* Find the device with shared context. */
2274 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2275 		struct mlx5_priv *opriv =
2276 			rte_eth_devices[port_id].data->dev_private;
2277 
2278 		if (opriv && opriv != priv && opriv->sh == sh) {
2279 			sh_conf = &opriv->config;
2280 			break;
2281 		}
2282 	}
2283 	if (!sh_conf)
2284 		return 0;
2285 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2286 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2287 			     " for shared %s context", sh->ibdev_name);
2288 		rte_errno = EINVAL;
2289 		return rte_errno;
2290 	}
2291 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2292 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2293 			     " for shared %s context", sh->ibdev_name);
2294 		rte_errno = EINVAL;
2295 		return rte_errno;
2296 	}
2297 	return 0;
2298 }
2299 /**
2300  * Spawn an Ethernet device from Verbs information.
2301  *
2302  * @param dpdk_dev
2303  *   Backing DPDK device.
2304  * @param spawn
2305  *   Verbs device parameters (name, port, switch_info) to spawn.
2306  * @param config
2307  *   Device configuration parameters.
2308  *
2309  * @return
2310  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2311  *   is set. The following errors are defined:
2312  *
2313  *   EBUSY: device is not supposed to be spawned.
2314  *   EEXIST: device is already spawned
2315  */
2316 static struct rte_eth_dev *
2317 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2318 	       struct mlx5_dev_spawn_data *spawn,
2319 	       struct mlx5_dev_config config)
2320 {
2321 	const struct mlx5_switch_info *switch_info = &spawn->info;
2322 	struct mlx5_ibv_shared *sh = NULL;
2323 	struct ibv_port_attr port_attr;
2324 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2325 	struct rte_eth_dev *eth_dev = NULL;
2326 	struct mlx5_priv *priv = NULL;
2327 	int err = 0;
2328 	unsigned int hw_padding = 0;
2329 	unsigned int mps;
2330 	unsigned int cqe_comp;
2331 	unsigned int cqe_pad = 0;
2332 	unsigned int tunnel_en = 0;
2333 	unsigned int mpls_en = 0;
2334 	unsigned int swp = 0;
2335 	unsigned int mprq = 0;
2336 	unsigned int mprq_min_stride_size_n = 0;
2337 	unsigned int mprq_max_stride_size_n = 0;
2338 	unsigned int mprq_min_stride_num_n = 0;
2339 	unsigned int mprq_max_stride_num_n = 0;
2340 	struct rte_ether_addr mac;
2341 	char name[RTE_ETH_NAME_MAX_LEN];
2342 	int own_domain_id = 0;
2343 	uint16_t port_id;
2344 	unsigned int i;
2345 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2346 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2347 #endif
2348 
2349 	/* Determine if this port representor is supposed to be spawned. */
2350 	if (switch_info->representor && dpdk_dev->devargs) {
2351 		struct rte_eth_devargs eth_da;
2352 
2353 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2354 		if (err) {
2355 			rte_errno = -err;
2356 			DRV_LOG(ERR, "failed to process device arguments: %s",
2357 				strerror(rte_errno));
2358 			return NULL;
2359 		}
2360 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
2361 			if (eth_da.representor_ports[i] ==
2362 			    (uint16_t)switch_info->port_name)
2363 				break;
2364 		if (i == eth_da.nb_representor_ports) {
2365 			rte_errno = EBUSY;
2366 			return NULL;
2367 		}
2368 	}
2369 	/* Build device name. */
2370 	if (spawn->pf_bond <  0) {
2371 		/* Single device. */
2372 		if (!switch_info->representor)
2373 			strlcpy(name, dpdk_dev->name, sizeof(name));
2374 		else
2375 			snprintf(name, sizeof(name), "%s_representor_%u",
2376 				 dpdk_dev->name, switch_info->port_name);
2377 	} else {
2378 		/* Bonding device. */
2379 		if (!switch_info->representor)
2380 			snprintf(name, sizeof(name), "%s_%s",
2381 				 dpdk_dev->name, spawn->ibv_dev->name);
2382 		else
2383 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
2384 				 dpdk_dev->name, spawn->ibv_dev->name,
2385 				 switch_info->port_name);
2386 	}
2387 	/* check if the device is already spawned */
2388 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2389 		rte_errno = EEXIST;
2390 		return NULL;
2391 	}
2392 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2393 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2394 		struct mlx5_mp_id mp_id;
2395 
2396 		eth_dev = rte_eth_dev_attach_secondary(name);
2397 		if (eth_dev == NULL) {
2398 			DRV_LOG(ERR, "can not attach rte ethdev");
2399 			rte_errno = ENOMEM;
2400 			return NULL;
2401 		}
2402 		eth_dev->device = dpdk_dev;
2403 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2404 		err = mlx5_proc_priv_init(eth_dev);
2405 		if (err)
2406 			return NULL;
2407 		mp_id.port_id = eth_dev->data->port_id;
2408 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
2409 		/* Receive command fd from primary process */
2410 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
2411 		if (err < 0)
2412 			return NULL;
2413 		/* Remap UAR for Tx queues. */
2414 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2415 		if (err)
2416 			return NULL;
2417 		/*
2418 		 * Ethdev pointer is still required as input since
2419 		 * the primary device is not accessible from the
2420 		 * secondary process.
2421 		 */
2422 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2423 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2424 		return eth_dev;
2425 	}
2426 	/*
2427 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2428 	 * advance to create dv/verbs device context. We proceed the
2429 	 * devargs here to get ones, and later proceed devargs again
2430 	 * to override some hardware settings.
2431 	 */
2432 	err = mlx5_args(&config, dpdk_dev->devargs);
2433 	if (err) {
2434 		err = rte_errno;
2435 		DRV_LOG(ERR, "failed to process device arguments: %s",
2436 			strerror(rte_errno));
2437 		goto error;
2438 	}
2439 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2440 	if (!sh)
2441 		return NULL;
2442 	config.devx = sh->devx;
2443 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2444 	config.dest_tir = 1;
2445 #endif
2446 #ifdef HAVE_IBV_MLX5_MOD_SWP
2447 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2448 #endif
2449 	/*
2450 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2451 	 * as all ConnectX-5 devices.
2452 	 */
2453 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2454 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2455 #endif
2456 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2457 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2458 #endif
2459 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2460 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2461 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2462 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2463 			mps = MLX5_MPW_ENHANCED;
2464 		} else {
2465 			DRV_LOG(DEBUG, "MPW is supported");
2466 			mps = MLX5_MPW;
2467 		}
2468 	} else {
2469 		DRV_LOG(DEBUG, "MPW isn't supported");
2470 		mps = MLX5_MPW_DISABLED;
2471 	}
2472 #ifdef HAVE_IBV_MLX5_MOD_SWP
2473 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2474 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2475 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2476 #endif
2477 	config.swp = !!swp;
2478 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2479 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2480 		struct mlx5dv_striding_rq_caps mprq_caps =
2481 			dv_attr.striding_rq_caps;
2482 
2483 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2484 			mprq_caps.min_single_stride_log_num_of_bytes);
2485 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2486 			mprq_caps.max_single_stride_log_num_of_bytes);
2487 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2488 			mprq_caps.min_single_wqe_log_num_of_strides);
2489 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2490 			mprq_caps.max_single_wqe_log_num_of_strides);
2491 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2492 			mprq_caps.supported_qpts);
2493 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2494 		mprq = 1;
2495 		mprq_min_stride_size_n =
2496 			mprq_caps.min_single_stride_log_num_of_bytes;
2497 		mprq_max_stride_size_n =
2498 			mprq_caps.max_single_stride_log_num_of_bytes;
2499 		mprq_min_stride_num_n =
2500 			mprq_caps.min_single_wqe_log_num_of_strides;
2501 		mprq_max_stride_num_n =
2502 			mprq_caps.max_single_wqe_log_num_of_strides;
2503 	}
2504 #endif
2505 	if (RTE_CACHE_LINE_SIZE == 128 &&
2506 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2507 		cqe_comp = 0;
2508 	else
2509 		cqe_comp = 1;
2510 	config.cqe_comp = cqe_comp;
2511 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2512 	/* Whether device supports 128B Rx CQE padding. */
2513 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2514 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2515 #endif
2516 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2517 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2518 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2519 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2520 			     (dv_attr.tunnel_offloads_caps &
2521 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2522 			     (dv_attr.tunnel_offloads_caps &
2523 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2524 	}
2525 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2526 		tunnel_en ? "" : "not ");
2527 #else
2528 	DRV_LOG(WARNING,
2529 		"tunnel offloading disabled due to old OFED/rdma-core version");
2530 #endif
2531 	config.tunnel_en = tunnel_en;
2532 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2533 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2534 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2535 		   (dv_attr.tunnel_offloads_caps &
2536 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2537 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2538 		mpls_en ? "" : "not ");
2539 #else
2540 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2541 		" old OFED/rdma-core version or firmware configuration");
2542 #endif
2543 	config.mpls_en = mpls_en;
2544 	/* Check port status. */
2545 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2546 	if (err) {
2547 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2548 		goto error;
2549 	}
2550 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2551 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2552 		err = EINVAL;
2553 		goto error;
2554 	}
2555 	if (port_attr.state != IBV_PORT_ACTIVE)
2556 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2557 			mlx5_glue->port_state_str(port_attr.state),
2558 			port_attr.state);
2559 	/* Allocate private eth device data. */
2560 	priv = rte_zmalloc("ethdev private structure",
2561 			   sizeof(*priv),
2562 			   RTE_CACHE_LINE_SIZE);
2563 	if (priv == NULL) {
2564 		DRV_LOG(ERR, "priv allocation failure");
2565 		err = ENOMEM;
2566 		goto error;
2567 	}
2568 	priv->sh = sh;
2569 	priv->ibv_port = spawn->ibv_port;
2570 	priv->pci_dev = spawn->pci_dev;
2571 	priv->mtu = RTE_ETHER_MTU;
2572 	priv->mp_id.port_id = port_id;
2573 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
2574 #ifndef RTE_ARCH_64
2575 	/* Initialize UAR access locks for 32bit implementations. */
2576 	rte_spinlock_init(&priv->uar_lock_cq);
2577 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2578 		rte_spinlock_init(&priv->uar_lock[i]);
2579 #endif
2580 	/* Some internal functions rely on Netlink sockets, open them now. */
2581 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2582 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2583 	priv->representor = !!switch_info->representor;
2584 	priv->master = !!switch_info->master;
2585 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2586 	priv->vport_meta_tag = 0;
2587 	priv->vport_meta_mask = 0;
2588 	priv->pf_bond = spawn->pf_bond;
2589 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2590 	/*
2591 	 * The DevX port query API is implemented. E-Switch may use
2592 	 * either vport or reg_c[0] metadata register to match on
2593 	 * vport index. The engaged part of metadata register is
2594 	 * defined by mask.
2595 	 */
2596 	if (switch_info->representor || switch_info->master) {
2597 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2598 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2599 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2600 						 &devx_port);
2601 		if (err) {
2602 			DRV_LOG(WARNING,
2603 				"can't query devx port %d on device %s",
2604 				spawn->ibv_port, spawn->ibv_dev->name);
2605 			devx_port.comp_mask = 0;
2606 		}
2607 	}
2608 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2609 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2610 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2611 		if (!priv->vport_meta_mask) {
2612 			DRV_LOG(ERR, "vport zero mask for port %d"
2613 				     " on bonding device %s",
2614 				     spawn->ibv_port, spawn->ibv_dev->name);
2615 			err = ENOTSUP;
2616 			goto error;
2617 		}
2618 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2619 			DRV_LOG(ERR, "invalid vport tag for port %d"
2620 				     " on bonding device %s",
2621 				     spawn->ibv_port, spawn->ibv_dev->name);
2622 			err = ENOTSUP;
2623 			goto error;
2624 		}
2625 	}
2626 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2627 		priv->vport_id = devx_port.vport_num;
2628 	} else if (spawn->pf_bond >= 0) {
2629 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2630 			     " on bonding device %s",
2631 			     spawn->ibv_port, spawn->ibv_dev->name);
2632 		err = ENOTSUP;
2633 		goto error;
2634 	} else {
2635 		/* Suppose vport index in compatible way. */
2636 		priv->vport_id = switch_info->representor ?
2637 				 switch_info->port_name + 1 : -1;
2638 	}
2639 #else
2640 	/*
2641 	 * Kernel/rdma_core support single E-Switch per PF configurations
2642 	 * only and vport_id field contains the vport index for
2643 	 * associated VF, which is deduced from representor port name.
2644 	 * For example, let's have the IB device port 10, it has
2645 	 * attached network device eth0, which has port name attribute
2646 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2647 	 * as 3 (2+1). This assigning schema should be changed if the
2648 	 * multiple E-Switch instances per PF configurations or/and PCI
2649 	 * subfunctions are added.
2650 	 */
2651 	priv->vport_id = switch_info->representor ?
2652 			 switch_info->port_name + 1 : -1;
2653 #endif
2654 	/* representor_id field keeps the unmodified VF index. */
2655 	priv->representor_id = switch_info->representor ?
2656 			       switch_info->port_name : -1;
2657 	/*
2658 	 * Look for sibling devices in order to reuse their switch domain
2659 	 * if any, otherwise allocate one.
2660 	 */
2661 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2662 		const struct mlx5_priv *opriv =
2663 			rte_eth_devices[port_id].data->dev_private;
2664 
2665 		if (!opriv ||
2666 		    opriv->sh != priv->sh ||
2667 			opriv->domain_id ==
2668 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2669 			continue;
2670 		priv->domain_id = opriv->domain_id;
2671 		break;
2672 	}
2673 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2674 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2675 		if (err) {
2676 			err = rte_errno;
2677 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2678 				strerror(rte_errno));
2679 			goto error;
2680 		}
2681 		own_domain_id = 1;
2682 	}
2683 	/* Override some values set by hardware configuration. */
2684 	mlx5_args(&config, dpdk_dev->devargs);
2685 	err = mlx5_dev_check_sibling_config(priv, &config);
2686 	if (err)
2687 		goto error;
2688 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2689 			    IBV_DEVICE_RAW_IP_CSUM);
2690 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2691 		(config.hw_csum ? "" : "not "));
2692 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2693 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2694 	DRV_LOG(DEBUG, "counters are not supported");
2695 #endif
2696 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2697 	if (config.dv_flow_en) {
2698 		DRV_LOG(WARNING, "DV flow is not supported");
2699 		config.dv_flow_en = 0;
2700 	}
2701 #endif
2702 	config.ind_table_max_size =
2703 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2704 	/*
2705 	 * Remove this check once DPDK supports larger/variable
2706 	 * indirection tables.
2707 	 */
2708 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2709 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2710 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2711 		config.ind_table_max_size);
2712 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2713 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2714 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2715 		(config.hw_vlan_strip ? "" : "not "));
2716 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2717 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2718 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2719 		(config.hw_fcs_strip ? "" : "not "));
2720 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2721 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2722 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2723 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2724 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2725 #endif
2726 	if (config.hw_padding && !hw_padding) {
2727 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2728 		config.hw_padding = 0;
2729 	} else if (config.hw_padding) {
2730 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2731 	}
2732 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2733 		      (sh->device_attr.tso_caps.supported_qpts &
2734 		       (1 << IBV_QPT_RAW_PACKET)));
2735 	if (config.tso)
2736 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2737 	/*
2738 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2739 	 * by default.
2740 	 */
2741 	if (config.mps == MLX5_ARG_UNSET)
2742 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2743 							  MLX5_MPW_DISABLED;
2744 	else
2745 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2746 	DRV_LOG(INFO, "%sMPS is %s",
2747 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2748 		config.mps == MLX5_MPW ? "legacy " : "",
2749 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2750 	if (config.cqe_comp && !cqe_comp) {
2751 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2752 		config.cqe_comp = 0;
2753 	}
2754 	if (config.cqe_pad && !cqe_pad) {
2755 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2756 		config.cqe_pad = 0;
2757 	} else if (config.cqe_pad) {
2758 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2759 	}
2760 	if (config.devx) {
2761 		priv->counter_fallback = 0;
2762 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2763 		if (err) {
2764 			err = -err;
2765 			goto error;
2766 		}
2767 		if (!config.hca_attr.flow_counters_dump)
2768 			priv->counter_fallback = 1;
2769 #ifndef HAVE_IBV_DEVX_ASYNC
2770 		priv->counter_fallback = 1;
2771 #endif
2772 		if (priv->counter_fallback)
2773 			DRV_LOG(INFO, "Use fall-back DV counter management");
2774 		/* Check for LRO support. */
2775 		if (config.dest_tir && config.hca_attr.lro_cap &&
2776 		    config.dv_flow_en) {
2777 			/* TBD check tunnel lro caps. */
2778 			config.lro.supported = config.hca_attr.lro_cap;
2779 			DRV_LOG(DEBUG, "Device supports LRO");
2780 			/*
2781 			 * If LRO timeout is not configured by application,
2782 			 * use the minimal supported value.
2783 			 */
2784 			if (!config.lro.timeout)
2785 				config.lro.timeout =
2786 				config.hca_attr.lro_timer_supported_periods[0];
2787 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2788 				config.lro.timeout);
2789 		}
2790 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2791 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2792 		    config.dv_flow_en) {
2793 			uint8_t reg_c_mask =
2794 				config.hca_attr.qos.flow_meter_reg_c_ids;
2795 			/*
2796 			 * Meter needs two REG_C's for color match and pre-sfx
2797 			 * flow match. Here get the REG_C for color match.
2798 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2799 			 */
2800 			reg_c_mask &= 0xfc;
2801 			if (__builtin_popcount(reg_c_mask) < 1) {
2802 				priv->mtr_en = 0;
2803 				DRV_LOG(WARNING, "No available register for"
2804 					" meter.");
2805 			} else {
2806 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2807 						      REG_C_0;
2808 				priv->mtr_en = 1;
2809 				priv->mtr_reg_share =
2810 				      config.hca_attr.qos.flow_meter_reg_share;
2811 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2812 					priv->mtr_color_reg);
2813 			}
2814 		}
2815 #endif
2816 	}
2817 	if (config.mprq.enabled && mprq) {
2818 		if (config.mprq.stride_num_n &&
2819 		    (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2820 		     config.mprq.stride_num_n < mprq_min_stride_num_n)) {
2821 			config.mprq.stride_num_n =
2822 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2823 						mprq_min_stride_num_n),
2824 					mprq_max_stride_num_n);
2825 			DRV_LOG(WARNING,
2826 				"the number of strides"
2827 				" for Multi-Packet RQ is out of range,"
2828 				" setting default value (%u)",
2829 				1 << config.mprq.stride_num_n);
2830 		}
2831 		if (config.mprq.stride_size_n &&
2832 		    (config.mprq.stride_size_n > mprq_max_stride_size_n ||
2833 		     config.mprq.stride_size_n < mprq_min_stride_size_n)) {
2834 			config.mprq.stride_size_n =
2835 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
2836 						mprq_min_stride_size_n),
2837 					mprq_max_stride_size_n);
2838 			DRV_LOG(WARNING,
2839 				"the size of a stride"
2840 				" for Multi-Packet RQ is out of range,"
2841 				" setting default value (%u)",
2842 				1 << config.mprq.stride_size_n);
2843 		}
2844 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2845 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2846 	} else if (config.mprq.enabled && !mprq) {
2847 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2848 		config.mprq.enabled = 0;
2849 	}
2850 	if (config.max_dump_files_num == 0)
2851 		config.max_dump_files_num = 128;
2852 	eth_dev = rte_eth_dev_allocate(name);
2853 	if (eth_dev == NULL) {
2854 		DRV_LOG(ERR, "can not allocate rte ethdev");
2855 		err = ENOMEM;
2856 		goto error;
2857 	}
2858 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2859 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2860 	if (priv->representor) {
2861 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2862 		eth_dev->data->representor_id = priv->representor_id;
2863 	}
2864 	/*
2865 	 * Store associated network device interface index. This index
2866 	 * is permanent throughout the lifetime of device. So, we may store
2867 	 * the ifindex here and use the cached value further.
2868 	 */
2869 	MLX5_ASSERT(spawn->ifindex);
2870 	priv->if_index = spawn->ifindex;
2871 	eth_dev->data->dev_private = priv;
2872 	priv->dev_data = eth_dev->data;
2873 	eth_dev->data->mac_addrs = priv->mac;
2874 	eth_dev->device = dpdk_dev;
2875 	/* Configure the first MAC address by default. */
2876 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2877 		DRV_LOG(ERR,
2878 			"port %u cannot get MAC address, is mlx5_en"
2879 			" loaded? (errno: %s)",
2880 			eth_dev->data->port_id, strerror(rte_errno));
2881 		err = ENODEV;
2882 		goto error;
2883 	}
2884 	DRV_LOG(INFO,
2885 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2886 		eth_dev->data->port_id,
2887 		mac.addr_bytes[0], mac.addr_bytes[1],
2888 		mac.addr_bytes[2], mac.addr_bytes[3],
2889 		mac.addr_bytes[4], mac.addr_bytes[5]);
2890 #ifdef RTE_LIBRTE_MLX5_DEBUG
2891 	{
2892 		char ifname[IF_NAMESIZE];
2893 
2894 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2895 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2896 				eth_dev->data->port_id, ifname);
2897 		else
2898 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2899 				eth_dev->data->port_id);
2900 	}
2901 #endif
2902 	/* Get actual MTU if possible. */
2903 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2904 	if (err) {
2905 		err = rte_errno;
2906 		goto error;
2907 	}
2908 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2909 		priv->mtu);
2910 	/* Initialize burst functions to prevent crashes before link-up. */
2911 	eth_dev->rx_pkt_burst = removed_rx_burst;
2912 	eth_dev->tx_pkt_burst = removed_tx_burst;
2913 	eth_dev->dev_ops = &mlx5_dev_ops;
2914 	/* Register MAC address. */
2915 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2916 	if (config.vf && config.vf_nl_en)
2917 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2918 				      mlx5_ifindex(eth_dev),
2919 				      eth_dev->data->mac_addrs,
2920 				      MLX5_MAX_MAC_ADDRESSES);
2921 	priv->flows = 0;
2922 	priv->ctrl_flows = 0;
2923 	TAILQ_INIT(&priv->flow_meters);
2924 	TAILQ_INIT(&priv->flow_meter_profiles);
2925 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2926 	struct mlx5dv_ctx_allocators alctr = {
2927 		.alloc = &mlx5_alloc_verbs_buf,
2928 		.free = &mlx5_free_verbs_buf,
2929 		.data = priv,
2930 	};
2931 	mlx5_glue->dv_set_context_attr(sh->ctx,
2932 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2933 				       (void *)((uintptr_t)&alctr));
2934 	/* Bring Ethernet device up. */
2935 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2936 		eth_dev->data->port_id);
2937 	mlx5_set_link_up(eth_dev);
2938 	/*
2939 	 * Even though the interrupt handler is not installed yet,
2940 	 * interrupts will still trigger on the async_fd from
2941 	 * Verbs context returned by ibv_open_device().
2942 	 */
2943 	mlx5_link_update(eth_dev, 0);
2944 #ifdef HAVE_MLX5DV_DR_ESWITCH
2945 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2946 	      (switch_info->representor || switch_info->master)))
2947 		config.dv_esw_en = 0;
2948 #else
2949 	config.dv_esw_en = 0;
2950 #endif
2951 	/* Detect minimal data bytes to inline. */
2952 	mlx5_set_min_inline(spawn, &config);
2953 	/* Store device configuration on private structure. */
2954 	priv->config = config;
2955 	/* Create context for virtual machine VLAN workaround. */
2956 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2957 	if (config.dv_flow_en) {
2958 		err = mlx5_alloc_shared_dr(priv);
2959 		if (err)
2960 			goto error;
2961 		/*
2962 		 * RSS id is shared with meter flow id. Meter flow id can only
2963 		 * use the 24 MSB of the register.
2964 		 */
2965 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2966 				     MLX5_MTR_COLOR_BITS);
2967 		if (!priv->qrss_id_pool) {
2968 			DRV_LOG(ERR, "can't create flow id pool");
2969 			err = ENOMEM;
2970 			goto error;
2971 		}
2972 	}
2973 	/* Supported Verbs flow priority number detection. */
2974 	err = mlx5_flow_discover_priorities(eth_dev);
2975 	if (err < 0) {
2976 		err = -err;
2977 		goto error;
2978 	}
2979 	priv->config.flow_prio = err;
2980 	if (!priv->config.dv_esw_en &&
2981 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2982 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2983 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2984 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2985 	}
2986 	mlx5_set_metadata_mask(eth_dev);
2987 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2988 	    !priv->sh->dv_regc0_mask) {
2989 		DRV_LOG(ERR, "metadata mode %u is not supported "
2990 			     "(no metadata reg_c[0] is available)",
2991 			     priv->config.dv_xmeta_en);
2992 			err = ENOTSUP;
2993 			goto error;
2994 	}
2995 	/*
2996 	 * Allocate the buffer for flow creating, just once.
2997 	 * The allocation must be done before any flow creating.
2998 	 */
2999 	mlx5_flow_alloc_intermediate(eth_dev);
3000 	/* Query availibility of metadata reg_c's. */
3001 	err = mlx5_flow_discover_mreg_c(eth_dev);
3002 	if (err < 0) {
3003 		err = -err;
3004 		goto error;
3005 	}
3006 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
3007 		DRV_LOG(DEBUG,
3008 			"port %u extensive metadata register is not supported",
3009 			eth_dev->data->port_id);
3010 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
3011 			DRV_LOG(ERR, "metadata mode %u is not supported "
3012 				     "(no metadata registers available)",
3013 				     priv->config.dv_xmeta_en);
3014 			err = ENOTSUP;
3015 			goto error;
3016 		}
3017 	}
3018 	if (priv->config.dv_flow_en &&
3019 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
3020 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
3021 	    priv->sh->dv_regc0_mask) {
3022 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
3023 						      MLX5_FLOW_MREG_HTABLE_SZ);
3024 		if (!priv->mreg_cp_tbl) {
3025 			err = ENOMEM;
3026 			goto error;
3027 		}
3028 	}
3029 	return eth_dev;
3030 error:
3031 	if (priv) {
3032 		if (priv->mreg_cp_tbl)
3033 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
3034 		if (priv->sh)
3035 			mlx5_free_shared_dr(priv);
3036 		if (priv->nl_socket_route >= 0)
3037 			close(priv->nl_socket_route);
3038 		if (priv->nl_socket_rdma >= 0)
3039 			close(priv->nl_socket_rdma);
3040 		if (priv->vmwa_context)
3041 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
3042 		if (priv->qrss_id_pool)
3043 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
3044 		if (own_domain_id)
3045 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
3046 		rte_free(priv);
3047 		if (eth_dev != NULL)
3048 			eth_dev->data->dev_private = NULL;
3049 	}
3050 	if (eth_dev != NULL) {
3051 		/* mac_addrs must not be freed alone because part of dev_private */
3052 		eth_dev->data->mac_addrs = NULL;
3053 		rte_eth_dev_release_port(eth_dev);
3054 	}
3055 	if (sh)
3056 		mlx5_free_shared_ibctx(sh);
3057 	MLX5_ASSERT(err > 0);
3058 	rte_errno = err;
3059 	return NULL;
3060 }
3061 
3062 /**
3063  * Comparison callback to sort device data.
3064  *
3065  * This is meant to be used with qsort().
3066  *
3067  * @param a[in]
3068  *   Pointer to pointer to first data object.
3069  * @param b[in]
3070  *   Pointer to pointer to second data object.
3071  *
3072  * @return
3073  *   0 if both objects are equal, less than 0 if the first argument is less
3074  *   than the second, greater than 0 otherwise.
3075  */
3076 static int
3077 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
3078 {
3079 	const struct mlx5_switch_info *si_a =
3080 		&((const struct mlx5_dev_spawn_data *)a)->info;
3081 	const struct mlx5_switch_info *si_b =
3082 		&((const struct mlx5_dev_spawn_data *)b)->info;
3083 	int ret;
3084 
3085 	/* Master device first. */
3086 	ret = si_b->master - si_a->master;
3087 	if (ret)
3088 		return ret;
3089 	/* Then representor devices. */
3090 	ret = si_b->representor - si_a->representor;
3091 	if (ret)
3092 		return ret;
3093 	/* Unidentified devices come last in no specific order. */
3094 	if (!si_a->representor)
3095 		return 0;
3096 	/* Order representors by name. */
3097 	return si_a->port_name - si_b->port_name;
3098 }
3099 
3100 /**
3101  * Match PCI information for possible slaves of bonding device.
3102  *
3103  * @param[in] ibv_dev
3104  *   Pointer to Infiniband device structure.
3105  * @param[in] pci_dev
3106  *   Pointer to PCI device structure to match PCI address.
3107  * @param[in] nl_rdma
3108  *   Netlink RDMA group socket handle.
3109  *
3110  * @return
3111  *   negative value if no bonding device found, otherwise
3112  *   positive index of slave PF in bonding.
3113  */
3114 static int
3115 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
3116 			   const struct rte_pci_device *pci_dev,
3117 			   int nl_rdma)
3118 {
3119 	char ifname[IF_NAMESIZE + 1];
3120 	unsigned int ifindex;
3121 	unsigned int np, i;
3122 	FILE *file = NULL;
3123 	int pf = -1;
3124 
3125 	/*
3126 	 * Try to get master device name. If something goes
3127 	 * wrong suppose the lack of kernel support and no
3128 	 * bonding devices.
3129 	 */
3130 	if (nl_rdma < 0)
3131 		return -1;
3132 	if (!strstr(ibv_dev->name, "bond"))
3133 		return -1;
3134 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
3135 	if (!np)
3136 		return -1;
3137 	/*
3138 	 * The Master device might not be on the predefined
3139 	 * port (not on port index 1, it is not garanted),
3140 	 * we have to scan all Infiniband device port and
3141 	 * find master.
3142 	 */
3143 	for (i = 1; i <= np; ++i) {
3144 		/* Check whether Infiniband port is populated. */
3145 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
3146 		if (!ifindex)
3147 			continue;
3148 		if (!if_indextoname(ifindex, ifname))
3149 			continue;
3150 		/* Try to read bonding slave names from sysfs. */
3151 		MKSTR(slaves,
3152 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
3153 		file = fopen(slaves, "r");
3154 		if (file)
3155 			break;
3156 	}
3157 	if (!file)
3158 		return -1;
3159 	/* Use safe format to check maximal buffer length. */
3160 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
3161 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
3162 		char tmp_str[IF_NAMESIZE + 32];
3163 		struct rte_pci_addr pci_addr;
3164 		struct mlx5_switch_info	info;
3165 
3166 		/* Process slave interface names in the loop. */
3167 		snprintf(tmp_str, sizeof(tmp_str),
3168 			 "/sys/class/net/%s", ifname);
3169 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
3170 			DRV_LOG(WARNING, "can not get PCI address"
3171 					 " for netdev \"%s\"", ifname);
3172 			continue;
3173 		}
3174 		if (pci_dev->addr.domain != pci_addr.domain ||
3175 		    pci_dev->addr.bus != pci_addr.bus ||
3176 		    pci_dev->addr.devid != pci_addr.devid ||
3177 		    pci_dev->addr.function != pci_addr.function)
3178 			continue;
3179 		/* Slave interface PCI address match found. */
3180 		fclose(file);
3181 		snprintf(tmp_str, sizeof(tmp_str),
3182 			 "/sys/class/net/%s/phys_port_name", ifname);
3183 		file = fopen(tmp_str, "rb");
3184 		if (!file)
3185 			break;
3186 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
3187 		if (fscanf(file, "%32s", tmp_str) == 1)
3188 			mlx5_translate_port_name(tmp_str, &info);
3189 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
3190 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
3191 			pf = info.port_name;
3192 		break;
3193 	}
3194 	if (file)
3195 		fclose(file);
3196 	return pf;
3197 }
3198 
3199 /**
3200  * DPDK callback to register a PCI device.
3201  *
3202  * This function spawns Ethernet devices out of a given PCI device.
3203  *
3204  * @param[in] pci_drv
3205  *   PCI driver structure (mlx5_driver).
3206  * @param[in] pci_dev
3207  *   PCI device information.
3208  *
3209  * @return
3210  *   0 on success, a negative errno value otherwise and rte_errno is set.
3211  */
3212 static int
3213 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
3214 	       struct rte_pci_device *pci_dev)
3215 {
3216 	struct ibv_device **ibv_list;
3217 	/*
3218 	 * Number of found IB Devices matching with requested PCI BDF.
3219 	 * nd != 1 means there are multiple IB devices over the same
3220 	 * PCI device and we have representors and master.
3221 	 */
3222 	unsigned int nd = 0;
3223 	/*
3224 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
3225 	 * we have the single multiport IB device, and there may be
3226 	 * representors attached to some of found ports.
3227 	 */
3228 	unsigned int np = 0;
3229 	/*
3230 	 * Number of DPDK ethernet devices to Spawn - either over
3231 	 * multiple IB devices or multiple ports of single IB device.
3232 	 * Actually this is the number of iterations to spawn.
3233 	 */
3234 	unsigned int ns = 0;
3235 	/*
3236 	 * Bonding device
3237 	 *   < 0 - no bonding device (single one)
3238 	 *  >= 0 - bonding device (value is slave PF index)
3239 	 */
3240 	int bd = -1;
3241 	struct mlx5_dev_spawn_data *list = NULL;
3242 	struct mlx5_dev_config dev_config;
3243 	int ret;
3244 
3245 	if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
3246 		DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
3247 			" driver.");
3248 		return 1;
3249 	}
3250 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
3251 		mlx5_pmd_socket_init();
3252 	ret = mlx5_init_once();
3253 	if (ret) {
3254 		DRV_LOG(ERR, "unable to init PMD global data: %s",
3255 			strerror(rte_errno));
3256 		return -rte_errno;
3257 	}
3258 	MLX5_ASSERT(pci_drv == &mlx5_driver);
3259 	errno = 0;
3260 	ibv_list = mlx5_glue->get_device_list(&ret);
3261 	if (!ibv_list) {
3262 		rte_errno = errno ? errno : ENOSYS;
3263 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
3264 		return -rte_errno;
3265 	}
3266 	/*
3267 	 * First scan the list of all Infiniband devices to find
3268 	 * matching ones, gathering into the list.
3269 	 */
3270 	struct ibv_device *ibv_match[ret + 1];
3271 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3272 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3273 	unsigned int i;
3274 
3275 	while (ret-- > 0) {
3276 		struct rte_pci_addr pci_addr;
3277 
3278 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3279 		bd = mlx5_device_bond_pci_match
3280 				(ibv_list[ret], pci_dev, nl_rdma);
3281 		if (bd >= 0) {
3282 			/*
3283 			 * Bonding device detected. Only one match is allowed,
3284 			 * the bonding is supported over multi-port IB device,
3285 			 * there should be no matches on representor PCI
3286 			 * functions or non VF LAG bonding devices with
3287 			 * specified address.
3288 			 */
3289 			if (nd) {
3290 				DRV_LOG(ERR,
3291 					"multiple PCI match on bonding device"
3292 					"\"%s\" found", ibv_list[ret]->name);
3293 				rte_errno = ENOENT;
3294 				ret = -rte_errno;
3295 				goto exit;
3296 			}
3297 			DRV_LOG(INFO, "PCI information matches for"
3298 				      " slave %d bonding device \"%s\"",
3299 				      bd, ibv_list[ret]->name);
3300 			ibv_match[nd++] = ibv_list[ret];
3301 			break;
3302 		}
3303 		if (mlx5_dev_to_pci_addr
3304 			(ibv_list[ret]->ibdev_path, &pci_addr))
3305 			continue;
3306 		if (pci_dev->addr.domain != pci_addr.domain ||
3307 		    pci_dev->addr.bus != pci_addr.bus ||
3308 		    pci_dev->addr.devid != pci_addr.devid ||
3309 		    pci_dev->addr.function != pci_addr.function)
3310 			continue;
3311 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3312 			ibv_list[ret]->name);
3313 		ibv_match[nd++] = ibv_list[ret];
3314 	}
3315 	ibv_match[nd] = NULL;
3316 	if (!nd) {
3317 		/* No device matches, just complain and bail out. */
3318 		DRV_LOG(WARNING,
3319 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
3320 			" are kernel drivers loaded?",
3321 			pci_dev->addr.domain, pci_dev->addr.bus,
3322 			pci_dev->addr.devid, pci_dev->addr.function);
3323 		rte_errno = ENOENT;
3324 		ret = -rte_errno;
3325 		goto exit;
3326 	}
3327 	if (nd == 1) {
3328 		/*
3329 		 * Found single matching device may have multiple ports.
3330 		 * Each port may be representor, we have to check the port
3331 		 * number and check the representors existence.
3332 		 */
3333 		if (nl_rdma >= 0)
3334 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3335 		if (!np)
3336 			DRV_LOG(WARNING, "can not get IB device \"%s\""
3337 					 " ports number", ibv_match[0]->name);
3338 		if (bd >= 0 && !np) {
3339 			DRV_LOG(ERR, "can not get ports"
3340 				     " for bonding device");
3341 			rte_errno = ENOENT;
3342 			ret = -rte_errno;
3343 			goto exit;
3344 		}
3345 	}
3346 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3347 	if (bd >= 0) {
3348 		/*
3349 		 * This may happen if there is VF LAG kernel support and
3350 		 * application is compiled with older rdma_core library.
3351 		 */
3352 		DRV_LOG(ERR,
3353 			"No kernel/verbs support for VF LAG bonding found.");
3354 		rte_errno = ENOTSUP;
3355 		ret = -rte_errno;
3356 		goto exit;
3357 	}
3358 #endif
3359 	/*
3360 	 * Now we can determine the maximal
3361 	 * amount of devices to be spawned.
3362 	 */
3363 	list = rte_zmalloc("device spawn data",
3364 			 sizeof(struct mlx5_dev_spawn_data) *
3365 			 (np ? np : nd),
3366 			 RTE_CACHE_LINE_SIZE);
3367 	if (!list) {
3368 		DRV_LOG(ERR, "spawn data array allocation failure");
3369 		rte_errno = ENOMEM;
3370 		ret = -rte_errno;
3371 		goto exit;
3372 	}
3373 	if (bd >= 0 || np > 1) {
3374 		/*
3375 		 * Single IB device with multiple ports found,
3376 		 * it may be E-Switch master device and representors.
3377 		 * We have to perform identification through the ports.
3378 		 */
3379 		MLX5_ASSERT(nl_rdma >= 0);
3380 		MLX5_ASSERT(ns == 0);
3381 		MLX5_ASSERT(nd == 1);
3382 		MLX5_ASSERT(np);
3383 		for (i = 1; i <= np; ++i) {
3384 			list[ns].max_port = np;
3385 			list[ns].ibv_port = i;
3386 			list[ns].ibv_dev = ibv_match[0];
3387 			list[ns].eth_dev = NULL;
3388 			list[ns].pci_dev = pci_dev;
3389 			list[ns].pf_bond = bd;
3390 			list[ns].ifindex = mlx5_nl_ifindex
3391 					(nl_rdma, list[ns].ibv_dev->name, i);
3392 			if (!list[ns].ifindex) {
3393 				/*
3394 				 * No network interface index found for the
3395 				 * specified port, it means there is no
3396 				 * representor on this port. It's OK,
3397 				 * there can be disabled ports, for example
3398 				 * if sriov_numvfs < sriov_totalvfs.
3399 				 */
3400 				continue;
3401 			}
3402 			ret = -1;
3403 			if (nl_route >= 0)
3404 				ret = mlx5_nl_switch_info
3405 					       (nl_route,
3406 						list[ns].ifindex,
3407 						&list[ns].info);
3408 			if (ret || (!list[ns].info.representor &&
3409 				    !list[ns].info.master)) {
3410 				/*
3411 				 * We failed to recognize representors with
3412 				 * Netlink, let's try to perform the task
3413 				 * with sysfs.
3414 				 */
3415 				ret =  mlx5_sysfs_switch_info
3416 						(list[ns].ifindex,
3417 						 &list[ns].info);
3418 			}
3419 			if (!ret && bd >= 0) {
3420 				switch (list[ns].info.name_type) {
3421 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3422 					if (list[ns].info.port_name == bd)
3423 						ns++;
3424 					break;
3425 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3426 					if (list[ns].info.pf_num == bd)
3427 						ns++;
3428 					break;
3429 				default:
3430 					break;
3431 				}
3432 				continue;
3433 			}
3434 			if (!ret && (list[ns].info.representor ^
3435 				     list[ns].info.master))
3436 				ns++;
3437 		}
3438 		if (!ns) {
3439 			DRV_LOG(ERR,
3440 				"unable to recognize master/representors"
3441 				" on the IB device with multiple ports");
3442 			rte_errno = ENOENT;
3443 			ret = -rte_errno;
3444 			goto exit;
3445 		}
3446 	} else {
3447 		/*
3448 		 * The existence of several matching entries (nd > 1) means
3449 		 * port representors have been instantiated. No existing Verbs
3450 		 * call nor sysfs entries can tell them apart, this can only
3451 		 * be done through Netlink calls assuming kernel drivers are
3452 		 * recent enough to support them.
3453 		 *
3454 		 * In the event of identification failure through Netlink,
3455 		 * try again through sysfs, then:
3456 		 *
3457 		 * 1. A single IB device matches (nd == 1) with single
3458 		 *    port (np=0/1) and is not a representor, assume
3459 		 *    no switch support.
3460 		 *
3461 		 * 2. Otherwise no safe assumptions can be made;
3462 		 *    complain louder and bail out.
3463 		 */
3464 		np = 1;
3465 		for (i = 0; i != nd; ++i) {
3466 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3467 			list[ns].max_port = 1;
3468 			list[ns].ibv_port = 1;
3469 			list[ns].ibv_dev = ibv_match[i];
3470 			list[ns].eth_dev = NULL;
3471 			list[ns].pci_dev = pci_dev;
3472 			list[ns].pf_bond = -1;
3473 			list[ns].ifindex = 0;
3474 			if (nl_rdma >= 0)
3475 				list[ns].ifindex = mlx5_nl_ifindex
3476 					(nl_rdma, list[ns].ibv_dev->name, 1);
3477 			if (!list[ns].ifindex) {
3478 				char ifname[IF_NAMESIZE];
3479 
3480 				/*
3481 				 * Netlink failed, it may happen with old
3482 				 * ib_core kernel driver (before 4.16).
3483 				 * We can assume there is old driver because
3484 				 * here we are processing single ports IB
3485 				 * devices. Let's try sysfs to retrieve
3486 				 * the ifindex. The method works for
3487 				 * master device only.
3488 				 */
3489 				if (nd > 1) {
3490 					/*
3491 					 * Multiple devices found, assume
3492 					 * representors, can not distinguish
3493 					 * master/representor and retrieve
3494 					 * ifindex via sysfs.
3495 					 */
3496 					continue;
3497 				}
3498 				ret = mlx5_get_master_ifname
3499 					(ibv_match[i]->ibdev_path, &ifname);
3500 				if (!ret)
3501 					list[ns].ifindex =
3502 						if_nametoindex(ifname);
3503 				if (!list[ns].ifindex) {
3504 					/*
3505 					 * No network interface index found
3506 					 * for the specified device, it means
3507 					 * there it is neither representor
3508 					 * nor master.
3509 					 */
3510 					continue;
3511 				}
3512 			}
3513 			ret = -1;
3514 			if (nl_route >= 0)
3515 				ret = mlx5_nl_switch_info
3516 					       (nl_route,
3517 						list[ns].ifindex,
3518 						&list[ns].info);
3519 			if (ret || (!list[ns].info.representor &&
3520 				    !list[ns].info.master)) {
3521 				/*
3522 				 * We failed to recognize representors with
3523 				 * Netlink, let's try to perform the task
3524 				 * with sysfs.
3525 				 */
3526 				ret =  mlx5_sysfs_switch_info
3527 						(list[ns].ifindex,
3528 						 &list[ns].info);
3529 			}
3530 			if (!ret && (list[ns].info.representor ^
3531 				     list[ns].info.master)) {
3532 				ns++;
3533 			} else if ((nd == 1) &&
3534 				   !list[ns].info.representor &&
3535 				   !list[ns].info.master) {
3536 				/*
3537 				 * Single IB device with
3538 				 * one physical port and
3539 				 * attached network device.
3540 				 * May be SRIOV is not enabled
3541 				 * or there is no representors.
3542 				 */
3543 				DRV_LOG(INFO, "no E-Switch support detected");
3544 				ns++;
3545 				break;
3546 			}
3547 		}
3548 		if (!ns) {
3549 			DRV_LOG(ERR,
3550 				"unable to recognize master/representors"
3551 				" on the multiple IB devices");
3552 			rte_errno = ENOENT;
3553 			ret = -rte_errno;
3554 			goto exit;
3555 		}
3556 	}
3557 	MLX5_ASSERT(ns);
3558 	/*
3559 	 * Sort list to probe devices in natural order for users convenience
3560 	 * (i.e. master first, then representors from lowest to highest ID).
3561 	 */
3562 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3563 	/* Default configuration. */
3564 	dev_config = (struct mlx5_dev_config){
3565 		.hw_padding = 0,
3566 		.mps = MLX5_ARG_UNSET,
3567 		.dbnc = MLX5_ARG_UNSET,
3568 		.rx_vec_en = 1,
3569 		.txq_inline_max = MLX5_ARG_UNSET,
3570 		.txq_inline_min = MLX5_ARG_UNSET,
3571 		.txq_inline_mpw = MLX5_ARG_UNSET,
3572 		.txqs_inline = MLX5_ARG_UNSET,
3573 		.vf_nl_en = 1,
3574 		.mr_ext_memseg_en = 1,
3575 		.mprq = {
3576 			.enabled = 0, /* Disabled by default. */
3577 			.stride_num_n = 0,
3578 			.stride_size_n = 0,
3579 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3580 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3581 		},
3582 		.dv_esw_en = 1,
3583 		.dv_flow_en = 1,
3584 		.log_hp_size = MLX5_ARG_UNSET,
3585 	};
3586 	/* Device specific configuration. */
3587 	switch (pci_dev->id.device_id) {
3588 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3589 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3590 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3591 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3592 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3593 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3594 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3595 		dev_config.vf = 1;
3596 		break;
3597 	default:
3598 		break;
3599 	}
3600 	for (i = 0; i != ns; ++i) {
3601 		uint32_t restore;
3602 
3603 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3604 						 &list[i],
3605 						 dev_config);
3606 		if (!list[i].eth_dev) {
3607 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3608 				break;
3609 			/* Device is disabled or already spawned. Ignore it. */
3610 			continue;
3611 		}
3612 		restore = list[i].eth_dev->data->dev_flags;
3613 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3614 		/* Restore non-PCI flags cleared by the above call. */
3615 		list[i].eth_dev->data->dev_flags |= restore;
3616 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3617 		rte_eth_dev_probing_finish(list[i].eth_dev);
3618 	}
3619 	if (i != ns) {
3620 		DRV_LOG(ERR,
3621 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3622 			" encountering an error: %s",
3623 			pci_dev->addr.domain, pci_dev->addr.bus,
3624 			pci_dev->addr.devid, pci_dev->addr.function,
3625 			strerror(rte_errno));
3626 		ret = -rte_errno;
3627 		/* Roll back. */
3628 		while (i--) {
3629 			if (!list[i].eth_dev)
3630 				continue;
3631 			mlx5_dev_close(list[i].eth_dev);
3632 			/* mac_addrs must not be freed because in dev_private */
3633 			list[i].eth_dev->data->mac_addrs = NULL;
3634 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3635 		}
3636 		/* Restore original error. */
3637 		rte_errno = -ret;
3638 	} else {
3639 		ret = 0;
3640 	}
3641 exit:
3642 	/*
3643 	 * Do the routine cleanup:
3644 	 * - close opened Netlink sockets
3645 	 * - free allocated spawn data array
3646 	 * - free the Infiniband device list
3647 	 */
3648 	if (nl_rdma >= 0)
3649 		close(nl_rdma);
3650 	if (nl_route >= 0)
3651 		close(nl_route);
3652 	if (list)
3653 		rte_free(list);
3654 	MLX5_ASSERT(ibv_list);
3655 	mlx5_glue->free_device_list(ibv_list);
3656 	return ret;
3657 }
3658 
3659 /**
3660  * Look for the ethernet device belonging to mlx5 driver.
3661  *
3662  * @param[in] port_id
3663  *   port_id to start looking for device.
3664  * @param[in] pci_dev
3665  *   Pointer to the hint PCI device. When device is being probed
3666  *   the its siblings (master and preceding representors might
3667  *   not have assigned driver yet (because the mlx5_pci_probe()
3668  *   is not completed yet, for this case match on hint PCI
3669  *   device may be used to detect sibling device.
3670  *
3671  * @return
3672  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3673  */
3674 uint16_t
3675 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3676 {
3677 	while (port_id < RTE_MAX_ETHPORTS) {
3678 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3679 
3680 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3681 		    dev->device &&
3682 		    (dev->device == &pci_dev->device ||
3683 		     (dev->device->driver &&
3684 		     dev->device->driver->name &&
3685 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3686 			break;
3687 		port_id++;
3688 	}
3689 	if (port_id >= RTE_MAX_ETHPORTS)
3690 		return RTE_MAX_ETHPORTS;
3691 	return port_id;
3692 }
3693 
3694 /**
3695  * DPDK callback to remove a PCI device.
3696  *
3697  * This function removes all Ethernet devices belong to a given PCI device.
3698  *
3699  * @param[in] pci_dev
3700  *   Pointer to the PCI device.
3701  *
3702  * @return
3703  *   0 on success, the function cannot fail.
3704  */
3705 static int
3706 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3707 {
3708 	uint16_t port_id;
3709 
3710 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3711 		rte_eth_dev_close(port_id);
3712 	return 0;
3713 }
3714 
3715 static const struct rte_pci_id mlx5_pci_id_map[] = {
3716 	{
3717 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3718 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3719 	},
3720 	{
3721 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3722 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3723 	},
3724 	{
3725 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3726 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3727 	},
3728 	{
3729 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3730 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3731 	},
3732 	{
3733 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3734 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3735 	},
3736 	{
3737 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3738 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3739 	},
3740 	{
3741 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3742 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3743 	},
3744 	{
3745 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3746 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3747 	},
3748 	{
3749 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3750 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3751 	},
3752 	{
3753 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3754 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3755 	},
3756 	{
3757 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3758 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3759 	},
3760 	{
3761 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3762 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3763 	},
3764 	{
3765 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3766 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3767 	},
3768 	{
3769 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3770 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3771 	},
3772 	{
3773 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3774 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
3775 	},
3776 	{
3777 		.vendor_id = 0
3778 	}
3779 };
3780 
3781 static struct rte_pci_driver mlx5_driver = {
3782 	.driver = {
3783 		.name = MLX5_DRIVER_NAME
3784 	},
3785 	.id_table = mlx5_pci_id_map,
3786 	.probe = mlx5_pci_probe,
3787 	.remove = mlx5_pci_remove,
3788 	.dma_map = mlx5_dma_map,
3789 	.dma_unmap = mlx5_dma_unmap,
3790 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3791 		     RTE_PCI_DRV_PROBE_AGAIN,
3792 };
3793 
3794 /**
3795  * Driver initialization routine.
3796  */
3797 RTE_INIT(rte_mlx5_pmd_init)
3798 {
3799 	/* Initialize driver log type. */
3800 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3801 	if (mlx5_logtype >= 0)
3802 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3803 
3804 	/* Build the static tables for Verbs conversion. */
3805 	mlx5_set_ptype_table();
3806 	mlx5_set_cksum_table();
3807 	mlx5_set_swp_types_table();
3808 	if (mlx5_glue)
3809 		rte_pci_register(&mlx5_driver);
3810 }
3811 
3812 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3813 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3814 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3815