xref: /dpdk/drivers/net/mlx5/mlx5.c (revision e4373bf1b3f51715bf66e87c0134e2c217e4612c)
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 <assert.h>
10 #include <dlfcn.h>
11 #include <stdint.h>
12 #include <stdlib.h>
13 #include <errno.h>
14 #include <net/if.h>
15 #include <sys/mman.h>
16 #include <linux/rtnetlink.h>
17 
18 /* Verbs header. */
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27 
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
31 #include <rte_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_kvargs.h>
36 #include <rte_rwlock.h>
37 #include <rte_spinlock.h>
38 #include <rte_string_fns.h>
39 #include <rte_alarm.h>
40 
41 #include "mlx5.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_autoconf.h"
45 #include "mlx5_defs.h"
46 #include "mlx5_glue.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49 
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52 
53 /* Device parameter to enable RX completion entry padding to 128B. */
54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
55 
56 /* Device parameter to enable padding Rx packet to cacheline size. */
57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
58 
59 /* Device parameter to enable Multi-Packet Rx queue. */
60 #define MLX5_RX_MPRQ_EN "mprq_en"
61 
62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
64 
65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
67 
68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
70 
71 /* Device parameter to configure inline send. Deprecated, ignored.*/
72 #define MLX5_TXQ_INLINE "txq_inline"
73 
74 /* Device parameter to limit packet size to inline with ordinary SEND. */
75 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
76 
77 /* Device parameter to configure minimal data size to inline. */
78 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
79 
80 /* Device parameter to limit packet size to inline with Enhanced MPW. */
81 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
82 
83 /*
84  * Device parameter to configure the number of TX queues threshold for
85  * enabling inline send.
86  */
87 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
88 
89 /*
90  * Device parameter to configure the number of TX queues threshold for
91  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
92  */
93 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
94 
95 /* Device parameter to enable multi-packet send WQEs. */
96 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
97 
98 /*
99  * Device parameter to force doorbell register mapping
100  * to non-cahed region eliminating the extra write memory barrier.
101  */
102 #define MLX5_TX_DB_NC "tx_db_nc"
103 
104 /*
105  * Device parameter to include 2 dsegs in the title WQEBB.
106  * Deprecated, ignored.
107  */
108 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
109 
110 /*
111  * Device parameter to limit the size of inlining packet.
112  * Deprecated, ignored.
113  */
114 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
115 
116 /*
117  * Device parameter to enable hardware Tx vector.
118  * Deprecated, ignored (no vectorized Tx routines anymore).
119  */
120 #define MLX5_TX_VEC_EN "tx_vec_en"
121 
122 /* Device parameter to enable hardware Rx vector. */
123 #define MLX5_RX_VEC_EN "rx_vec_en"
124 
125 /* Allow L3 VXLAN flow creation. */
126 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
127 
128 /* Activate DV E-Switch flow steering. */
129 #define MLX5_DV_ESW_EN "dv_esw_en"
130 
131 /* Activate DV flow steering. */
132 #define MLX5_DV_FLOW_EN "dv_flow_en"
133 
134 /* Enable extensive flow metadata support. */
135 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
136 
137 /* Activate Netlink support in VF mode. */
138 #define MLX5_VF_NL_EN "vf_nl_en"
139 
140 /* Enable extending memsegs when creating a MR. */
141 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
142 
143 /* Select port representors to instantiate. */
144 #define MLX5_REPRESENTOR "representor"
145 
146 /* Device parameter to configure the maximum number of dump files per queue. */
147 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
148 
149 /* Configure timeout of LRO session (in microseconds). */
150 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
151 
152 #ifndef HAVE_IBV_MLX5_MOD_MPW
153 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
154 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
155 #endif
156 
157 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
158 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
159 #endif
160 
161 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
162 
163 /* Shared memory between primary and secondary processes. */
164 struct mlx5_shared_data *mlx5_shared_data;
165 
166 /* Spinlock for mlx5_shared_data allocation. */
167 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
168 
169 /* Process local data for secondary processes. */
170 static struct mlx5_local_data mlx5_local_data;
171 
172 /** Driver-specific log messages type. */
173 int mlx5_logtype;
174 
175 /** Data associated with devices to spawn. */
176 struct mlx5_dev_spawn_data {
177 	uint32_t ifindex; /**< Network interface index. */
178 	uint32_t max_port; /**< IB device maximal port index. */
179 	uint32_t ibv_port; /**< IB device physical port index. */
180 	int pf_bond; /**< bonding device PF index. < 0 - no bonding */
181 	struct mlx5_switch_info info; /**< Switch information. */
182 	struct ibv_device *ibv_dev; /**< Associated IB device. */
183 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
184 	struct rte_pci_device *pci_dev; /**< Backend PCI device. */
185 };
186 
187 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
188 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
189 
190 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
191 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
192 
193 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
194 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
195 
196 /**
197  * Allocate ID pool structure.
198  *
199  * @return
200  *   Pointer to pool object, NULL value otherwise.
201  */
202 struct mlx5_flow_id_pool *
203 mlx5_flow_id_pool_alloc(void)
204 {
205 	struct mlx5_flow_id_pool *pool;
206 	void *mem;
207 
208 	pool = rte_zmalloc("id pool allocation", sizeof(*pool),
209 			   RTE_CACHE_LINE_SIZE);
210 	if (!pool) {
211 		DRV_LOG(ERR, "can't allocate id pool");
212 		rte_errno  = ENOMEM;
213 		return NULL;
214 	}
215 	mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
216 			  RTE_CACHE_LINE_SIZE);
217 	if (!mem) {
218 		DRV_LOG(ERR, "can't allocate mem for id pool");
219 		rte_errno  = ENOMEM;
220 		goto error;
221 	}
222 	pool->free_arr = mem;
223 	pool->curr = pool->free_arr;
224 	pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
225 	pool->base_index = 0;
226 	return pool;
227 error:
228 	rte_free(pool);
229 	return NULL;
230 }
231 
232 /**
233  * Release ID pool structure.
234  *
235  * @param[in] pool
236  *   Pointer to flow id pool object to free.
237  */
238 void
239 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
240 {
241 	rte_free(pool->free_arr);
242 	rte_free(pool);
243 }
244 
245 /**
246  * Generate ID.
247  *
248  * @param[in] pool
249  *   Pointer to flow id pool.
250  * @param[out] id
251  *   The generated ID.
252  *
253  * @return
254  *   0 on success, error value otherwise.
255  */
256 uint32_t
257 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
258 {
259 	if (pool->curr == pool->free_arr) {
260 		if (pool->base_index == UINT32_MAX) {
261 			rte_errno  = ENOMEM;
262 			DRV_LOG(ERR, "no free id");
263 			return -rte_errno;
264 		}
265 		*id = ++pool->base_index;
266 		return 0;
267 	}
268 	*id = *(--pool->curr);
269 	return 0;
270 }
271 
272 /**
273  * Release ID.
274  *
275  * @param[in] pool
276  *   Pointer to flow id pool.
277  * @param[out] id
278  *   The generated ID.
279  *
280  * @return
281  *   0 on success, error value otherwise.
282  */
283 uint32_t
284 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
285 {
286 	uint32_t size;
287 	uint32_t size2;
288 	void *mem;
289 
290 	if (pool->curr == pool->last) {
291 		size = pool->curr - pool->free_arr;
292 		size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
293 		assert(size2 > size);
294 		mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
295 		if (!mem) {
296 			DRV_LOG(ERR, "can't allocate mem for id pool");
297 			rte_errno  = ENOMEM;
298 			return -rte_errno;
299 		}
300 		memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
301 		rte_free(pool->free_arr);
302 		pool->free_arr = mem;
303 		pool->curr = pool->free_arr + size;
304 		pool->last = pool->free_arr + size2;
305 	}
306 	*pool->curr = id;
307 	pool->curr++;
308 	return 0;
309 }
310 
311 /**
312  * Initialize the counters management structure.
313  *
314  * @param[in] sh
315  *   Pointer to mlx5_ibv_shared object to free
316  */
317 static void
318 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
319 {
320 	uint8_t i;
321 
322 	TAILQ_INIT(&sh->cmng.flow_counters);
323 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
324 		TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
325 }
326 
327 /**
328  * Destroy all the resources allocated for a counter memory management.
329  *
330  * @param[in] mng
331  *   Pointer to the memory management structure.
332  */
333 static void
334 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
335 {
336 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
337 
338 	LIST_REMOVE(mng, next);
339 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
340 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
341 	rte_free(mem);
342 }
343 
344 /**
345  * Close and release all the resources of the counters management.
346  *
347  * @param[in] sh
348  *   Pointer to mlx5_ibv_shared object to free.
349  */
350 static void
351 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
352 {
353 	struct mlx5_counter_stats_mem_mng *mng;
354 	uint8_t i;
355 	int j;
356 	int retries = 1024;
357 
358 	rte_errno = 0;
359 	while (--retries) {
360 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
361 		if (rte_errno != EINPROGRESS)
362 			break;
363 		rte_pause();
364 	}
365 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
366 		struct mlx5_flow_counter_pool *pool;
367 		uint32_t batch = !!(i % 2);
368 
369 		if (!sh->cmng.ccont[i].pools)
370 			continue;
371 		pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
372 		while (pool) {
373 			if (batch) {
374 				if (pool->min_dcs)
375 					claim_zero
376 					(mlx5_devx_cmd_destroy(pool->min_dcs));
377 			}
378 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
379 				if (pool->counters_raw[j].action)
380 					claim_zero
381 					(mlx5_glue->destroy_flow_action
382 					       (pool->counters_raw[j].action));
383 				if (!batch && pool->counters_raw[j].dcs)
384 					claim_zero(mlx5_devx_cmd_destroy
385 						  (pool->counters_raw[j].dcs));
386 			}
387 			TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
388 				     next);
389 			rte_free(pool);
390 			pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
391 		}
392 		rte_free(sh->cmng.ccont[i].pools);
393 	}
394 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
395 	while (mng) {
396 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
397 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
398 	}
399 	memset(&sh->cmng, 0, sizeof(sh->cmng));
400 }
401 
402 /**
403  * Extract pdn of PD object using DV API.
404  *
405  * @param[in] pd
406  *   Pointer to the verbs PD object.
407  * @param[out] pdn
408  *   Pointer to the PD object number variable.
409  *
410  * @return
411  *   0 on success, error value otherwise.
412  */
413 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
414 static int
415 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
416 {
417 	struct mlx5dv_obj obj;
418 	struct mlx5dv_pd pd_info;
419 	int ret = 0;
420 
421 	obj.pd.in = pd;
422 	obj.pd.out = &pd_info;
423 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
424 	if (ret) {
425 		DRV_LOG(DEBUG, "Fail to get PD object info");
426 		return ret;
427 	}
428 	*pdn = pd_info.pdn;
429 	return 0;
430 }
431 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
432 
433 static int
434 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
435 {
436 	char *env;
437 	int value;
438 
439 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
440 	/* Get environment variable to store. */
441 	env = getenv(MLX5_SHUT_UP_BF);
442 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
443 	if (config->dbnc == MLX5_ARG_UNSET)
444 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
445 	else
446 		setenv(MLX5_SHUT_UP_BF,
447 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
448 	return value;
449 }
450 
451 static void
452 mlx5_restore_doorbell_mapping_env(int value)
453 {
454 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
455 	/* Restore the original environment variable state. */
456 	if (value == MLX5_ARG_UNSET)
457 		unsetenv(MLX5_SHUT_UP_BF);
458 	else
459 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
460 }
461 
462 /**
463  * Allocate shared IB device context. If there is multiport device the
464  * master and representors will share this context, if there is single
465  * port dedicated IB device, the context will be used by only given
466  * port due to unification.
467  *
468  * Routine first searches the context for the specified IB device name,
469  * if found the shared context assumed and reference counter is incremented.
470  * If no context found the new one is created and initialized with specified
471  * IB device context and parameters.
472  *
473  * @param[in] spawn
474  *   Pointer to the IB device attributes (name, port, etc).
475  * @param[in] config
476  *   Pointer to device configuration structure.
477  *
478  * @return
479  *   Pointer to mlx5_ibv_shared object on success,
480  *   otherwise NULL and rte_errno is set.
481  */
482 static struct mlx5_ibv_shared *
483 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
484 			const struct mlx5_dev_config *config)
485 {
486 	struct mlx5_ibv_shared *sh;
487 	int dbmap_env;
488 	int err = 0;
489 	uint32_t i;
490 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
491 	struct mlx5_devx_tis_attr tis_attr = { 0 };
492 #endif
493 
494 	assert(spawn);
495 	/* Secondary process should not create the shared context. */
496 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
497 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
498 	/* Search for IB context by device name. */
499 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
500 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
501 			sh->refcnt++;
502 			goto exit;
503 		}
504 	}
505 	/* No device found, we have to create new shared context. */
506 	assert(spawn->max_port);
507 	sh = rte_zmalloc("ethdev shared ib context",
508 			 sizeof(struct mlx5_ibv_shared) +
509 			 spawn->max_port *
510 			 sizeof(struct mlx5_ibv_shared_port),
511 			 RTE_CACHE_LINE_SIZE);
512 	if (!sh) {
513 		DRV_LOG(ERR, "shared context allocation failure");
514 		rte_errno  = ENOMEM;
515 		goto exit;
516 	}
517 	/*
518 	 * Configure environment variable "MLX5_BF_SHUT_UP"
519 	 * before the device creation. The rdma_core library
520 	 * checks the variable at device creation and
521 	 * stores the result internally.
522 	 */
523 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
524 	/* Try to open IB device with DV first, then usual Verbs. */
525 	errno = 0;
526 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
527 	if (sh->ctx) {
528 		sh->devx = 1;
529 		DRV_LOG(DEBUG, "DevX is supported");
530 		/* The device is created, no need for environment. */
531 		mlx5_restore_doorbell_mapping_env(dbmap_env);
532 	} else {
533 		/* The environment variable is still configured. */
534 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
535 		err = errno ? errno : ENODEV;
536 		/*
537 		 * The environment variable is not needed anymore,
538 		 * all device creation attempts are completed.
539 		 */
540 		mlx5_restore_doorbell_mapping_env(dbmap_env);
541 		if (!sh->ctx)
542 			goto error;
543 		DRV_LOG(DEBUG, "DevX is NOT supported");
544 	}
545 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
546 	if (err) {
547 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
548 		goto error;
549 	}
550 	sh->refcnt = 1;
551 	sh->max_port = spawn->max_port;
552 	strncpy(sh->ibdev_name, sh->ctx->device->name,
553 		sizeof(sh->ibdev_name));
554 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
555 		sizeof(sh->ibdev_path));
556 	pthread_mutex_init(&sh->intr_mutex, NULL);
557 	/*
558 	 * Setting port_id to max unallowed value means
559 	 * there is no interrupt subhandler installed for
560 	 * the given port index i.
561 	 */
562 	for (i = 0; i < sh->max_port; i++) {
563 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
564 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
565 	}
566 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
567 	if (sh->pd == NULL) {
568 		DRV_LOG(ERR, "PD allocation failure");
569 		err = ENOMEM;
570 		goto error;
571 	}
572 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
573 	if (sh->devx) {
574 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
575 		if (err) {
576 			DRV_LOG(ERR, "Fail to extract pdn from PD");
577 			goto error;
578 		}
579 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
580 		if (!sh->td) {
581 			DRV_LOG(ERR, "TD allocation failure");
582 			err = ENOMEM;
583 			goto error;
584 		}
585 		tis_attr.transport_domain = sh->td->id;
586 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
587 		if (!sh->tis) {
588 			DRV_LOG(ERR, "TIS allocation failure");
589 			err = ENOMEM;
590 			goto error;
591 		}
592 	}
593 	sh->flow_id_pool = mlx5_flow_id_pool_alloc();
594 	if (!sh->flow_id_pool) {
595 		DRV_LOG(ERR, "can't create flow id pool");
596 		err = ENOMEM;
597 		goto error;
598 	}
599 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
600 	/*
601 	 * Once the device is added to the list of memory event
602 	 * callback, its global MR cache table cannot be expanded
603 	 * on the fly because of deadlock. If it overflows, lookup
604 	 * should be done by searching MR list linearly, which is slow.
605 	 *
606 	 * At this point the device is not added to the memory
607 	 * event list yet, context is just being created.
608 	 */
609 	err = mlx5_mr_btree_init(&sh->mr.cache,
610 				 MLX5_MR_BTREE_CACHE_N * 2,
611 				 spawn->pci_dev->device.numa_node);
612 	if (err) {
613 		err = rte_errno;
614 		goto error;
615 	}
616 	mlx5_flow_counters_mng_init(sh);
617 	/* Add device to memory callback list. */
618 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
619 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
620 			 sh, mem_event_cb);
621 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
622 	/* Add context to the global device list. */
623 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
624 exit:
625 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
626 	return sh;
627 error:
628 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
629 	assert(sh);
630 	if (sh->tis)
631 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
632 	if (sh->td)
633 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
634 	if (sh->pd)
635 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
636 	if (sh->ctx)
637 		claim_zero(mlx5_glue->close_device(sh->ctx));
638 	if (sh->flow_id_pool)
639 		mlx5_flow_id_pool_release(sh->flow_id_pool);
640 	rte_free(sh);
641 	assert(err > 0);
642 	rte_errno = err;
643 	return NULL;
644 }
645 
646 /**
647  * Free shared IB device context. Decrement counter and if zero free
648  * all allocated resources and close handles.
649  *
650  * @param[in] sh
651  *   Pointer to mlx5_ibv_shared object to free
652  */
653 static void
654 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
655 {
656 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
657 #ifndef NDEBUG
658 	/* Check the object presence in the list. */
659 	struct mlx5_ibv_shared *lctx;
660 
661 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
662 		if (lctx == sh)
663 			break;
664 	assert(lctx);
665 	if (lctx != sh) {
666 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
667 		goto exit;
668 	}
669 #endif
670 	assert(sh);
671 	assert(sh->refcnt);
672 	/* Secondary process should not free the shared context. */
673 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
674 	if (--sh->refcnt)
675 		goto exit;
676 	/* Release created Memory Regions. */
677 	mlx5_mr_release(sh);
678 	/* Remove from memory callback device list. */
679 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
680 	LIST_REMOVE(sh, mem_event_cb);
681 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
682 	/* Remove context from the global device list. */
683 	LIST_REMOVE(sh, next);
684 	/*
685 	 *  Ensure there is no async event handler installed.
686 	 *  Only primary process handles async device events.
687 	 **/
688 	mlx5_flow_counters_mng_close(sh);
689 	assert(!sh->intr_cnt);
690 	if (sh->intr_cnt)
691 		mlx5_intr_callback_unregister
692 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
693 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
694 	if (sh->devx_intr_cnt) {
695 		if (sh->intr_handle_devx.fd)
696 			rte_intr_callback_unregister(&sh->intr_handle_devx,
697 					  mlx5_dev_interrupt_handler_devx, sh);
698 		if (sh->devx_comp)
699 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
700 	}
701 #endif
702 	pthread_mutex_destroy(&sh->intr_mutex);
703 	if (sh->pd)
704 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
705 	if (sh->tis)
706 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
707 	if (sh->td)
708 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
709 	if (sh->ctx)
710 		claim_zero(mlx5_glue->close_device(sh->ctx));
711 	if (sh->flow_id_pool)
712 		mlx5_flow_id_pool_release(sh->flow_id_pool);
713 	rte_free(sh);
714 exit:
715 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
716 }
717 
718 /**
719  * Destroy table hash list and all the root entries per domain.
720  *
721  * @param[in] priv
722  *   Pointer to the private device data structure.
723  */
724 static void
725 mlx5_free_table_hash_list(struct mlx5_priv *priv)
726 {
727 	struct mlx5_ibv_shared *sh = priv->sh;
728 	struct mlx5_flow_tbl_data_entry *tbl_data;
729 	union mlx5_flow_tbl_key table_key = {
730 		{
731 			.table_id = 0,
732 			.reserved = 0,
733 			.domain = 0,
734 			.direction = 0,
735 		}
736 	};
737 	struct mlx5_hlist_entry *pos;
738 
739 	if (!sh->flow_tbls)
740 		return;
741 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
742 	if (pos) {
743 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
744 					entry);
745 		assert(tbl_data);
746 		mlx5_hlist_remove(sh->flow_tbls, pos);
747 		rte_free(tbl_data);
748 	}
749 	table_key.direction = 1;
750 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
751 	if (pos) {
752 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
753 					entry);
754 		assert(tbl_data);
755 		mlx5_hlist_remove(sh->flow_tbls, pos);
756 		rte_free(tbl_data);
757 	}
758 	table_key.direction = 0;
759 	table_key.domain = 1;
760 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
761 	if (pos) {
762 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
763 					entry);
764 		assert(tbl_data);
765 		mlx5_hlist_remove(sh->flow_tbls, pos);
766 		rte_free(tbl_data);
767 	}
768 	mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
769 }
770 
771 /**
772  * Initialize flow table hash list and create the root tables entry
773  * for each domain.
774  *
775  * @param[in] priv
776  *   Pointer to the private device data structure.
777  *
778  * @return
779  *   Zero on success, positive error code otherwise.
780  */
781 static int
782 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
783 {
784 	struct mlx5_ibv_shared *sh = priv->sh;
785 	char s[MLX5_HLIST_NAMESIZE];
786 	int err = 0;
787 
788 	assert(sh);
789 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
790 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
791 	if (!sh->flow_tbls) {
792 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
793 		err = ENOMEM;
794 		return err;
795 	}
796 #ifndef HAVE_MLX5DV_DR
797 	/*
798 	 * In case we have not DR support, the zero tables should be created
799 	 * because DV expect to see them even if they cannot be created by
800 	 * RDMA-CORE.
801 	 */
802 	union mlx5_flow_tbl_key table_key = {
803 		{
804 			.table_id = 0,
805 			.reserved = 0,
806 			.domain = 0,
807 			.direction = 0,
808 		}
809 	};
810 	struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
811 							  sizeof(*tbl_data), 0);
812 
813 	if (!tbl_data) {
814 		err = ENOMEM;
815 		goto error;
816 	}
817 	tbl_data->entry.key = table_key.v64;
818 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
819 	if (err)
820 		goto error;
821 	rte_atomic32_init(&tbl_data->tbl.refcnt);
822 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
823 	table_key.direction = 1;
824 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
825 	if (!tbl_data) {
826 		err = ENOMEM;
827 		goto error;
828 	}
829 	tbl_data->entry.key = table_key.v64;
830 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
831 	if (err)
832 		goto error;
833 	rte_atomic32_init(&tbl_data->tbl.refcnt);
834 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
835 	table_key.direction = 0;
836 	table_key.domain = 1;
837 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
838 	if (!tbl_data) {
839 		err = ENOMEM;
840 		goto error;
841 	}
842 	tbl_data->entry.key = table_key.v64;
843 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
844 	if (err)
845 		goto error;
846 	rte_atomic32_init(&tbl_data->tbl.refcnt);
847 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
848 	return err;
849 error:
850 	mlx5_free_table_hash_list(priv);
851 #endif /* HAVE_MLX5DV_DR */
852 	return err;
853 }
854 
855 /**
856  * Initialize DR related data within private structure.
857  * Routine checks the reference counter and does actual
858  * resources creation/initialization only if counter is zero.
859  *
860  * @param[in] priv
861  *   Pointer to the private device data structure.
862  *
863  * @return
864  *   Zero on success, positive error code otherwise.
865  */
866 static int
867 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
868 {
869 	struct mlx5_ibv_shared *sh = priv->sh;
870 	char s[MLX5_HLIST_NAMESIZE];
871 	int err = 0;
872 
873 	if (!sh->flow_tbls)
874 		err = mlx5_alloc_table_hash_list(priv);
875 	else
876 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
877 			(void *)sh->flow_tbls);
878 	if (err)
879 		return err;
880 	/* Create tags hash list table. */
881 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
882 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
883 	if (!sh->tag_table) {
884 		DRV_LOG(ERR, "tags with hash creation failed.\n");
885 		err = ENOMEM;
886 		goto error;
887 	}
888 #ifdef HAVE_MLX5DV_DR
889 	void *domain;
890 
891 	if (sh->dv_refcnt) {
892 		/* Shared DV/DR structures is already initialized. */
893 		sh->dv_refcnt++;
894 		priv->dr_shared = 1;
895 		return 0;
896 	}
897 	/* Reference counter is zero, we should initialize structures. */
898 	domain = mlx5_glue->dr_create_domain(sh->ctx,
899 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
900 	if (!domain) {
901 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
902 		err = errno;
903 		goto error;
904 	}
905 	sh->rx_domain = domain;
906 	domain = mlx5_glue->dr_create_domain(sh->ctx,
907 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
908 	if (!domain) {
909 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
910 		err = errno;
911 		goto error;
912 	}
913 	pthread_mutex_init(&sh->dv_mutex, NULL);
914 	sh->tx_domain = domain;
915 #ifdef HAVE_MLX5DV_DR_ESWITCH
916 	if (priv->config.dv_esw_en) {
917 		domain  = mlx5_glue->dr_create_domain
918 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
919 		if (!domain) {
920 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
921 			err = errno;
922 			goto error;
923 		}
924 		sh->fdb_domain = domain;
925 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
926 	}
927 #endif
928 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
929 #endif /* HAVE_MLX5DV_DR */
930 	sh->dv_refcnt++;
931 	priv->dr_shared = 1;
932 	return 0;
933 error:
934 	/* Rollback the created objects. */
935 	if (sh->rx_domain) {
936 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
937 		sh->rx_domain = NULL;
938 	}
939 	if (sh->tx_domain) {
940 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
941 		sh->tx_domain = NULL;
942 	}
943 	if (sh->fdb_domain) {
944 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
945 		sh->fdb_domain = NULL;
946 	}
947 	if (sh->esw_drop_action) {
948 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
949 		sh->esw_drop_action = NULL;
950 	}
951 	if (sh->pop_vlan_action) {
952 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
953 		sh->pop_vlan_action = NULL;
954 	}
955 	if (sh->tag_table) {
956 		/* tags should be destroyed with flow before. */
957 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
958 		sh->tag_table = NULL;
959 	}
960 	mlx5_free_table_hash_list(priv);
961 	return err;
962 }
963 
964 /**
965  * Destroy DR related data within private structure.
966  *
967  * @param[in] priv
968  *   Pointer to the private device data structure.
969  */
970 static void
971 mlx5_free_shared_dr(struct mlx5_priv *priv)
972 {
973 	struct mlx5_ibv_shared *sh;
974 
975 	if (!priv->dr_shared)
976 		return;
977 	priv->dr_shared = 0;
978 	sh = priv->sh;
979 	assert(sh);
980 #ifdef HAVE_MLX5DV_DR
981 	assert(sh->dv_refcnt);
982 	if (sh->dv_refcnt && --sh->dv_refcnt)
983 		return;
984 	if (sh->rx_domain) {
985 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
986 		sh->rx_domain = NULL;
987 	}
988 	if (sh->tx_domain) {
989 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
990 		sh->tx_domain = NULL;
991 	}
992 #ifdef HAVE_MLX5DV_DR_ESWITCH
993 	if (sh->fdb_domain) {
994 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
995 		sh->fdb_domain = NULL;
996 	}
997 	if (sh->esw_drop_action) {
998 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
999 		sh->esw_drop_action = NULL;
1000 	}
1001 #endif
1002 	if (sh->pop_vlan_action) {
1003 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1004 		sh->pop_vlan_action = NULL;
1005 	}
1006 	pthread_mutex_destroy(&sh->dv_mutex);
1007 #endif /* HAVE_MLX5DV_DR */
1008 	if (sh->tag_table) {
1009 		/* tags should be destroyed with flow before. */
1010 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1011 		sh->tag_table = NULL;
1012 	}
1013 	mlx5_free_table_hash_list(priv);
1014 }
1015 
1016 /**
1017  * Initialize shared data between primary and secondary process.
1018  *
1019  * A memzone is reserved by primary process and secondary processes attach to
1020  * the memzone.
1021  *
1022  * @return
1023  *   0 on success, a negative errno value otherwise and rte_errno is set.
1024  */
1025 static int
1026 mlx5_init_shared_data(void)
1027 {
1028 	const struct rte_memzone *mz;
1029 	int ret = 0;
1030 
1031 	rte_spinlock_lock(&mlx5_shared_data_lock);
1032 	if (mlx5_shared_data == NULL) {
1033 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1034 			/* Allocate shared memory. */
1035 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1036 						 sizeof(*mlx5_shared_data),
1037 						 SOCKET_ID_ANY, 0);
1038 			if (mz == NULL) {
1039 				DRV_LOG(ERR,
1040 					"Cannot allocate mlx5 shared data");
1041 				ret = -rte_errno;
1042 				goto error;
1043 			}
1044 			mlx5_shared_data = mz->addr;
1045 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1046 			rte_spinlock_init(&mlx5_shared_data->lock);
1047 		} else {
1048 			/* Lookup allocated shared memory. */
1049 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1050 			if (mz == NULL) {
1051 				DRV_LOG(ERR,
1052 					"Cannot attach mlx5 shared data");
1053 				ret = -rte_errno;
1054 				goto error;
1055 			}
1056 			mlx5_shared_data = mz->addr;
1057 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1058 		}
1059 	}
1060 error:
1061 	rte_spinlock_unlock(&mlx5_shared_data_lock);
1062 	return ret;
1063 }
1064 
1065 /**
1066  * Retrieve integer value from environment variable.
1067  *
1068  * @param[in] name
1069  *   Environment variable name.
1070  *
1071  * @return
1072  *   Integer value, 0 if the variable is not set.
1073  */
1074 int
1075 mlx5_getenv_int(const char *name)
1076 {
1077 	const char *val = getenv(name);
1078 
1079 	if (val == NULL)
1080 		return 0;
1081 	return atoi(val);
1082 }
1083 
1084 /**
1085  * Verbs callback to allocate a memory. This function should allocate the space
1086  * according to the size provided residing inside a huge page.
1087  * Please note that all allocation must respect the alignment from libmlx5
1088  * (i.e. currently sysconf(_SC_PAGESIZE)).
1089  *
1090  * @param[in] size
1091  *   The size in bytes of the memory to allocate.
1092  * @param[in] data
1093  *   A pointer to the callback data.
1094  *
1095  * @return
1096  *   Allocated buffer, NULL otherwise and rte_errno is set.
1097  */
1098 static void *
1099 mlx5_alloc_verbs_buf(size_t size, void *data)
1100 {
1101 	struct mlx5_priv *priv = data;
1102 	void *ret;
1103 	size_t alignment = sysconf(_SC_PAGESIZE);
1104 	unsigned int socket = SOCKET_ID_ANY;
1105 
1106 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1107 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1108 
1109 		socket = ctrl->socket;
1110 	} else if (priv->verbs_alloc_ctx.type ==
1111 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1112 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1113 
1114 		socket = ctrl->socket;
1115 	}
1116 	assert(data != NULL);
1117 	ret = rte_malloc_socket(__func__, size, alignment, socket);
1118 	if (!ret && size)
1119 		rte_errno = ENOMEM;
1120 	return ret;
1121 }
1122 
1123 /**
1124  * Verbs callback to free a memory.
1125  *
1126  * @param[in] ptr
1127  *   A pointer to the memory to free.
1128  * @param[in] data
1129  *   A pointer to the callback data.
1130  */
1131 static void
1132 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1133 {
1134 	assert(data != NULL);
1135 	rte_free(ptr);
1136 }
1137 
1138 /**
1139  * DPDK callback to add udp tunnel port
1140  *
1141  * @param[in] dev
1142  *   A pointer to eth_dev
1143  * @param[in] udp_tunnel
1144  *   A pointer to udp tunnel
1145  *
1146  * @return
1147  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1148  */
1149 int
1150 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1151 			 struct rte_eth_udp_tunnel *udp_tunnel)
1152 {
1153 	assert(udp_tunnel != NULL);
1154 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1155 	    udp_tunnel->udp_port == 4789)
1156 		return 0;
1157 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1158 	    udp_tunnel->udp_port == 4790)
1159 		return 0;
1160 	return -ENOTSUP;
1161 }
1162 
1163 /**
1164  * Initialize process private data structure.
1165  *
1166  * @param dev
1167  *   Pointer to Ethernet device structure.
1168  *
1169  * @return
1170  *   0 on success, a negative errno value otherwise and rte_errno is set.
1171  */
1172 int
1173 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1174 {
1175 	struct mlx5_priv *priv = dev->data->dev_private;
1176 	struct mlx5_proc_priv *ppriv;
1177 	size_t ppriv_size;
1178 
1179 	/*
1180 	 * UAR register table follows the process private structure. BlueFlame
1181 	 * registers for Tx queues are stored in the table.
1182 	 */
1183 	ppriv_size =
1184 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1185 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1186 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1187 	if (!ppriv) {
1188 		rte_errno = ENOMEM;
1189 		return -rte_errno;
1190 	}
1191 	ppriv->uar_table_sz = ppriv_size;
1192 	dev->process_private = ppriv;
1193 	return 0;
1194 }
1195 
1196 /**
1197  * Un-initialize process private data structure.
1198  *
1199  * @param dev
1200  *   Pointer to Ethernet device structure.
1201  */
1202 static void
1203 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1204 {
1205 	if (!dev->process_private)
1206 		return;
1207 	rte_free(dev->process_private);
1208 	dev->process_private = NULL;
1209 }
1210 
1211 /**
1212  * DPDK callback to close the device.
1213  *
1214  * Destroy all queues and objects, free memory.
1215  *
1216  * @param dev
1217  *   Pointer to Ethernet device structure.
1218  */
1219 static void
1220 mlx5_dev_close(struct rte_eth_dev *dev)
1221 {
1222 	struct mlx5_priv *priv = dev->data->dev_private;
1223 	unsigned int i;
1224 	int ret;
1225 
1226 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1227 		dev->data->port_id,
1228 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1229 	/* In case mlx5_dev_stop() has not been called. */
1230 	mlx5_dev_interrupt_handler_uninstall(dev);
1231 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1232 	mlx5_traffic_disable(dev);
1233 	mlx5_flow_flush(dev, NULL);
1234 	mlx5_flow_meter_flush(dev, NULL);
1235 	/* Prevent crashes when queues are still in use. */
1236 	dev->rx_pkt_burst = removed_rx_burst;
1237 	dev->tx_pkt_burst = removed_tx_burst;
1238 	rte_wmb();
1239 	/* Disable datapath on secondary process. */
1240 	mlx5_mp_req_stop_rxtx(dev);
1241 	if (priv->rxqs != NULL) {
1242 		/* XXX race condition if mlx5_rx_burst() is still running. */
1243 		usleep(1000);
1244 		for (i = 0; (i != priv->rxqs_n); ++i)
1245 			mlx5_rxq_release(dev, i);
1246 		priv->rxqs_n = 0;
1247 		priv->rxqs = NULL;
1248 	}
1249 	if (priv->txqs != NULL) {
1250 		/* XXX race condition if mlx5_tx_burst() is still running. */
1251 		usleep(1000);
1252 		for (i = 0; (i != priv->txqs_n); ++i)
1253 			mlx5_txq_release(dev, i);
1254 		priv->txqs_n = 0;
1255 		priv->txqs = NULL;
1256 	}
1257 	mlx5_proc_priv_uninit(dev);
1258 	if (priv->mreg_cp_tbl)
1259 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1260 	mlx5_mprq_free_mp(dev);
1261 	mlx5_free_shared_dr(priv);
1262 	if (priv->rss_conf.rss_key != NULL)
1263 		rte_free(priv->rss_conf.rss_key);
1264 	if (priv->reta_idx != NULL)
1265 		rte_free(priv->reta_idx);
1266 	if (priv->config.vf)
1267 		mlx5_nl_mac_addr_flush(dev);
1268 	if (priv->nl_socket_route >= 0)
1269 		close(priv->nl_socket_route);
1270 	if (priv->nl_socket_rdma >= 0)
1271 		close(priv->nl_socket_rdma);
1272 	if (priv->vmwa_context)
1273 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1274 	if (priv->sh) {
1275 		/*
1276 		 * Free the shared context in last turn, because the cleanup
1277 		 * routines above may use some shared fields, like
1278 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1279 		 * ifindex if Netlink fails.
1280 		 */
1281 		mlx5_free_shared_ibctx(priv->sh);
1282 		priv->sh = NULL;
1283 	}
1284 	ret = mlx5_hrxq_verify(dev);
1285 	if (ret)
1286 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1287 			dev->data->port_id);
1288 	ret = mlx5_ind_table_obj_verify(dev);
1289 	if (ret)
1290 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1291 			dev->data->port_id);
1292 	ret = mlx5_rxq_obj_verify(dev);
1293 	if (ret)
1294 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1295 			dev->data->port_id);
1296 	ret = mlx5_rxq_verify(dev);
1297 	if (ret)
1298 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1299 			dev->data->port_id);
1300 	ret = mlx5_txq_obj_verify(dev);
1301 	if (ret)
1302 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1303 			dev->data->port_id);
1304 	ret = mlx5_txq_verify(dev);
1305 	if (ret)
1306 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1307 			dev->data->port_id);
1308 	ret = mlx5_flow_verify(dev);
1309 	if (ret)
1310 		DRV_LOG(WARNING, "port %u some flows still remain",
1311 			dev->data->port_id);
1312 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1313 		unsigned int c = 0;
1314 		uint16_t port_id;
1315 
1316 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1317 			struct mlx5_priv *opriv =
1318 				rte_eth_devices[port_id].data->dev_private;
1319 
1320 			if (!opriv ||
1321 			    opriv->domain_id != priv->domain_id ||
1322 			    &rte_eth_devices[port_id] == dev)
1323 				continue;
1324 			++c;
1325 			break;
1326 		}
1327 		if (!c)
1328 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1329 	}
1330 	memset(priv, 0, sizeof(*priv));
1331 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1332 	/*
1333 	 * Reset mac_addrs to NULL such that it is not freed as part of
1334 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1335 	 * it is freed when dev_private is freed.
1336 	 */
1337 	dev->data->mac_addrs = NULL;
1338 }
1339 
1340 const struct eth_dev_ops mlx5_dev_ops = {
1341 	.dev_configure = mlx5_dev_configure,
1342 	.dev_start = mlx5_dev_start,
1343 	.dev_stop = mlx5_dev_stop,
1344 	.dev_set_link_down = mlx5_set_link_down,
1345 	.dev_set_link_up = mlx5_set_link_up,
1346 	.dev_close = mlx5_dev_close,
1347 	.promiscuous_enable = mlx5_promiscuous_enable,
1348 	.promiscuous_disable = mlx5_promiscuous_disable,
1349 	.allmulticast_enable = mlx5_allmulticast_enable,
1350 	.allmulticast_disable = mlx5_allmulticast_disable,
1351 	.link_update = mlx5_link_update,
1352 	.stats_get = mlx5_stats_get,
1353 	.stats_reset = mlx5_stats_reset,
1354 	.xstats_get = mlx5_xstats_get,
1355 	.xstats_reset = mlx5_xstats_reset,
1356 	.xstats_get_names = mlx5_xstats_get_names,
1357 	.fw_version_get = mlx5_fw_version_get,
1358 	.dev_infos_get = mlx5_dev_infos_get,
1359 	.read_clock = mlx5_read_clock,
1360 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1361 	.vlan_filter_set = mlx5_vlan_filter_set,
1362 	.rx_queue_setup = mlx5_rx_queue_setup,
1363 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1364 	.tx_queue_setup = mlx5_tx_queue_setup,
1365 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1366 	.rx_queue_release = mlx5_rx_queue_release,
1367 	.tx_queue_release = mlx5_tx_queue_release,
1368 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1369 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1370 	.mac_addr_remove = mlx5_mac_addr_remove,
1371 	.mac_addr_add = mlx5_mac_addr_add,
1372 	.mac_addr_set = mlx5_mac_addr_set,
1373 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1374 	.mtu_set = mlx5_dev_set_mtu,
1375 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1376 	.vlan_offload_set = mlx5_vlan_offload_set,
1377 	.reta_update = mlx5_dev_rss_reta_update,
1378 	.reta_query = mlx5_dev_rss_reta_query,
1379 	.rss_hash_update = mlx5_rss_hash_update,
1380 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1381 	.filter_ctrl = mlx5_dev_filter_ctrl,
1382 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1383 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1384 	.rx_queue_count = mlx5_rx_queue_count,
1385 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1386 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1387 	.is_removed = mlx5_is_removed,
1388 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1389 	.get_module_info = mlx5_get_module_info,
1390 	.get_module_eeprom = mlx5_get_module_eeprom,
1391 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1392 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1393 };
1394 
1395 /* Available operations from secondary process. */
1396 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1397 	.stats_get = mlx5_stats_get,
1398 	.stats_reset = mlx5_stats_reset,
1399 	.xstats_get = mlx5_xstats_get,
1400 	.xstats_reset = mlx5_xstats_reset,
1401 	.xstats_get_names = mlx5_xstats_get_names,
1402 	.fw_version_get = mlx5_fw_version_get,
1403 	.dev_infos_get = mlx5_dev_infos_get,
1404 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1405 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1406 	.get_module_info = mlx5_get_module_info,
1407 	.get_module_eeprom = mlx5_get_module_eeprom,
1408 };
1409 
1410 /* Available operations in flow isolated mode. */
1411 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1412 	.dev_configure = mlx5_dev_configure,
1413 	.dev_start = mlx5_dev_start,
1414 	.dev_stop = mlx5_dev_stop,
1415 	.dev_set_link_down = mlx5_set_link_down,
1416 	.dev_set_link_up = mlx5_set_link_up,
1417 	.dev_close = mlx5_dev_close,
1418 	.promiscuous_enable = mlx5_promiscuous_enable,
1419 	.promiscuous_disable = mlx5_promiscuous_disable,
1420 	.allmulticast_enable = mlx5_allmulticast_enable,
1421 	.allmulticast_disable = mlx5_allmulticast_disable,
1422 	.link_update = mlx5_link_update,
1423 	.stats_get = mlx5_stats_get,
1424 	.stats_reset = mlx5_stats_reset,
1425 	.xstats_get = mlx5_xstats_get,
1426 	.xstats_reset = mlx5_xstats_reset,
1427 	.xstats_get_names = mlx5_xstats_get_names,
1428 	.fw_version_get = mlx5_fw_version_get,
1429 	.dev_infos_get = mlx5_dev_infos_get,
1430 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1431 	.vlan_filter_set = mlx5_vlan_filter_set,
1432 	.rx_queue_setup = mlx5_rx_queue_setup,
1433 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1434 	.tx_queue_setup = mlx5_tx_queue_setup,
1435 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1436 	.rx_queue_release = mlx5_rx_queue_release,
1437 	.tx_queue_release = mlx5_tx_queue_release,
1438 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1439 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1440 	.mac_addr_remove = mlx5_mac_addr_remove,
1441 	.mac_addr_add = mlx5_mac_addr_add,
1442 	.mac_addr_set = mlx5_mac_addr_set,
1443 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1444 	.mtu_set = mlx5_dev_set_mtu,
1445 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1446 	.vlan_offload_set = mlx5_vlan_offload_set,
1447 	.filter_ctrl = mlx5_dev_filter_ctrl,
1448 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1449 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1450 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1451 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1452 	.is_removed = mlx5_is_removed,
1453 	.get_module_info = mlx5_get_module_info,
1454 	.get_module_eeprom = mlx5_get_module_eeprom,
1455 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1456 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1457 };
1458 
1459 /**
1460  * Verify and store value for device argument.
1461  *
1462  * @param[in] key
1463  *   Key argument to verify.
1464  * @param[in] val
1465  *   Value associated with key.
1466  * @param opaque
1467  *   User data.
1468  *
1469  * @return
1470  *   0 on success, a negative errno value otherwise and rte_errno is set.
1471  */
1472 static int
1473 mlx5_args_check(const char *key, const char *val, void *opaque)
1474 {
1475 	struct mlx5_dev_config *config = opaque;
1476 	unsigned long tmp;
1477 
1478 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1479 	if (!strcmp(MLX5_REPRESENTOR, key))
1480 		return 0;
1481 	errno = 0;
1482 	tmp = strtoul(val, NULL, 0);
1483 	if (errno) {
1484 		rte_errno = errno;
1485 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1486 		return -rte_errno;
1487 	}
1488 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1489 		config->cqe_comp = !!tmp;
1490 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1491 		config->cqe_pad = !!tmp;
1492 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1493 		config->hw_padding = !!tmp;
1494 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1495 		config->mprq.enabled = !!tmp;
1496 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1497 		config->mprq.stride_num_n = tmp;
1498 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1499 		config->mprq.max_memcpy_len = tmp;
1500 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1501 		config->mprq.min_rxqs_num = tmp;
1502 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1503 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1504 				 " converted to txq_inline_max", key);
1505 		config->txq_inline_max = tmp;
1506 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1507 		config->txq_inline_max = tmp;
1508 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1509 		config->txq_inline_min = tmp;
1510 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1511 		config->txq_inline_mpw = tmp;
1512 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1513 		config->txqs_inline = tmp;
1514 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1515 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1516 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1517 		config->mps = !!tmp;
1518 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1519 		if (tmp != MLX5_TXDB_CACHED &&
1520 		    tmp != MLX5_TXDB_NCACHED &&
1521 		    tmp != MLX5_TXDB_HEURISTIC) {
1522 			DRV_LOG(ERR, "invalid Tx doorbell "
1523 				     "mapping parameter");
1524 			rte_errno = EINVAL;
1525 			return -rte_errno;
1526 		}
1527 		config->dbnc = tmp;
1528 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1529 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1530 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1531 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1532 				 " converted to txq_inline_mpw", key);
1533 		config->txq_inline_mpw = tmp;
1534 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1535 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1536 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1537 		config->rx_vec_en = !!tmp;
1538 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1539 		config->l3_vxlan_en = !!tmp;
1540 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1541 		config->vf_nl_en = !!tmp;
1542 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1543 		config->dv_esw_en = !!tmp;
1544 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1545 		config->dv_flow_en = !!tmp;
1546 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1547 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1548 		    tmp != MLX5_XMETA_MODE_META16 &&
1549 		    tmp != MLX5_XMETA_MODE_META32) {
1550 			DRV_LOG(ERR, "invalid extensive "
1551 				     "metadata parameter");
1552 			rte_errno = EINVAL;
1553 			return -rte_errno;
1554 		}
1555 		config->dv_xmeta_en = tmp;
1556 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1557 		config->mr_ext_memseg_en = !!tmp;
1558 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1559 		config->max_dump_files_num = tmp;
1560 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1561 		config->lro.timeout = tmp;
1562 	} else {
1563 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1564 		rte_errno = EINVAL;
1565 		return -rte_errno;
1566 	}
1567 	return 0;
1568 }
1569 
1570 /**
1571  * Parse device parameters.
1572  *
1573  * @param config
1574  *   Pointer to device configuration structure.
1575  * @param devargs
1576  *   Device arguments structure.
1577  *
1578  * @return
1579  *   0 on success, a negative errno value otherwise and rte_errno is set.
1580  */
1581 static int
1582 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1583 {
1584 	const char **params = (const char *[]){
1585 		MLX5_RXQ_CQE_COMP_EN,
1586 		MLX5_RXQ_CQE_PAD_EN,
1587 		MLX5_RXQ_PKT_PAD_EN,
1588 		MLX5_RX_MPRQ_EN,
1589 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1590 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1591 		MLX5_RXQS_MIN_MPRQ,
1592 		MLX5_TXQ_INLINE,
1593 		MLX5_TXQ_INLINE_MIN,
1594 		MLX5_TXQ_INLINE_MAX,
1595 		MLX5_TXQ_INLINE_MPW,
1596 		MLX5_TXQS_MIN_INLINE,
1597 		MLX5_TXQS_MAX_VEC,
1598 		MLX5_TXQ_MPW_EN,
1599 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1600 		MLX5_TXQ_MAX_INLINE_LEN,
1601 		MLX5_TX_DB_NC,
1602 		MLX5_TX_VEC_EN,
1603 		MLX5_RX_VEC_EN,
1604 		MLX5_L3_VXLAN_EN,
1605 		MLX5_VF_NL_EN,
1606 		MLX5_DV_ESW_EN,
1607 		MLX5_DV_FLOW_EN,
1608 		MLX5_DV_XMETA_EN,
1609 		MLX5_MR_EXT_MEMSEG_EN,
1610 		MLX5_REPRESENTOR,
1611 		MLX5_MAX_DUMP_FILES_NUM,
1612 		MLX5_LRO_TIMEOUT_USEC,
1613 		NULL,
1614 	};
1615 	struct rte_kvargs *kvlist;
1616 	int ret = 0;
1617 	int i;
1618 
1619 	if (devargs == NULL)
1620 		return 0;
1621 	/* Following UGLY cast is done to pass checkpatch. */
1622 	kvlist = rte_kvargs_parse(devargs->args, params);
1623 	if (kvlist == NULL) {
1624 		rte_errno = EINVAL;
1625 		return -rte_errno;
1626 	}
1627 	/* Process parameters. */
1628 	for (i = 0; (params[i] != NULL); ++i) {
1629 		if (rte_kvargs_count(kvlist, params[i])) {
1630 			ret = rte_kvargs_process(kvlist, params[i],
1631 						 mlx5_args_check, config);
1632 			if (ret) {
1633 				rte_errno = EINVAL;
1634 				rte_kvargs_free(kvlist);
1635 				return -rte_errno;
1636 			}
1637 		}
1638 	}
1639 	rte_kvargs_free(kvlist);
1640 	return 0;
1641 }
1642 
1643 static struct rte_pci_driver mlx5_driver;
1644 
1645 /**
1646  * PMD global initialization.
1647  *
1648  * Independent from individual device, this function initializes global
1649  * per-PMD data structures distinguishing primary and secondary processes.
1650  * Hence, each initialization is called once per a process.
1651  *
1652  * @return
1653  *   0 on success, a negative errno value otherwise and rte_errno is set.
1654  */
1655 static int
1656 mlx5_init_once(void)
1657 {
1658 	struct mlx5_shared_data *sd;
1659 	struct mlx5_local_data *ld = &mlx5_local_data;
1660 	int ret = 0;
1661 
1662 	if (mlx5_init_shared_data())
1663 		return -rte_errno;
1664 	sd = mlx5_shared_data;
1665 	assert(sd);
1666 	rte_spinlock_lock(&sd->lock);
1667 	switch (rte_eal_process_type()) {
1668 	case RTE_PROC_PRIMARY:
1669 		if (sd->init_done)
1670 			break;
1671 		LIST_INIT(&sd->mem_event_cb_list);
1672 		rte_rwlock_init(&sd->mem_event_rwlock);
1673 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1674 						mlx5_mr_mem_event_cb, NULL);
1675 		ret = mlx5_mp_init_primary();
1676 		if (ret)
1677 			goto out;
1678 		sd->init_done = true;
1679 		break;
1680 	case RTE_PROC_SECONDARY:
1681 		if (ld->init_done)
1682 			break;
1683 		ret = mlx5_mp_init_secondary();
1684 		if (ret)
1685 			goto out;
1686 		++sd->secondary_cnt;
1687 		ld->init_done = true;
1688 		break;
1689 	default:
1690 		break;
1691 	}
1692 out:
1693 	rte_spinlock_unlock(&sd->lock);
1694 	return ret;
1695 }
1696 
1697 /**
1698  * Configures the minimal amount of data to inline into WQE
1699  * while sending packets.
1700  *
1701  * - the txq_inline_min has the maximal priority, if this
1702  *   key is specified in devargs
1703  * - if DevX is enabled the inline mode is queried from the
1704  *   device (HCA attributes and NIC vport context if needed).
1705  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1706  *   and none (0 bytes) for other NICs
1707  *
1708  * @param spawn
1709  *   Verbs device parameters (name, port, switch_info) to spawn.
1710  * @param config
1711  *   Device configuration parameters.
1712  */
1713 static void
1714 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1715 		    struct mlx5_dev_config *config)
1716 {
1717 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1718 		/* Application defines size of inlined data explicitly. */
1719 		switch (spawn->pci_dev->id.device_id) {
1720 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1721 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1722 			if (config->txq_inline_min <
1723 				       (int)MLX5_INLINE_HSIZE_L2) {
1724 				DRV_LOG(DEBUG,
1725 					"txq_inline_mix aligned to minimal"
1726 					" ConnectX-4 required value %d",
1727 					(int)MLX5_INLINE_HSIZE_L2);
1728 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1729 			}
1730 			break;
1731 		}
1732 		goto exit;
1733 	}
1734 	if (config->hca_attr.eth_net_offloads) {
1735 		/* We have DevX enabled, inline mode queried successfully. */
1736 		switch (config->hca_attr.wqe_inline_mode) {
1737 		case MLX5_CAP_INLINE_MODE_L2:
1738 			/* outer L2 header must be inlined. */
1739 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1740 			goto exit;
1741 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1742 			/* No inline data are required by NIC. */
1743 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1744 			config->hw_vlan_insert =
1745 				config->hca_attr.wqe_vlan_insert;
1746 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1747 			goto exit;
1748 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1749 			/* inline mode is defined by NIC vport context. */
1750 			if (!config->hca_attr.eth_virt)
1751 				break;
1752 			switch (config->hca_attr.vport_inline_mode) {
1753 			case MLX5_INLINE_MODE_NONE:
1754 				config->txq_inline_min =
1755 					MLX5_INLINE_HSIZE_NONE;
1756 				goto exit;
1757 			case MLX5_INLINE_MODE_L2:
1758 				config->txq_inline_min =
1759 					MLX5_INLINE_HSIZE_L2;
1760 				goto exit;
1761 			case MLX5_INLINE_MODE_IP:
1762 				config->txq_inline_min =
1763 					MLX5_INLINE_HSIZE_L3;
1764 				goto exit;
1765 			case MLX5_INLINE_MODE_TCP_UDP:
1766 				config->txq_inline_min =
1767 					MLX5_INLINE_HSIZE_L4;
1768 				goto exit;
1769 			case MLX5_INLINE_MODE_INNER_L2:
1770 				config->txq_inline_min =
1771 					MLX5_INLINE_HSIZE_INNER_L2;
1772 				goto exit;
1773 			case MLX5_INLINE_MODE_INNER_IP:
1774 				config->txq_inline_min =
1775 					MLX5_INLINE_HSIZE_INNER_L3;
1776 				goto exit;
1777 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1778 				config->txq_inline_min =
1779 					MLX5_INLINE_HSIZE_INNER_L4;
1780 				goto exit;
1781 			}
1782 		}
1783 	}
1784 	/*
1785 	 * We get here if we are unable to deduce
1786 	 * inline data size with DevX. Try PCI ID
1787 	 * to determine old NICs.
1788 	 */
1789 	switch (spawn->pci_dev->id.device_id) {
1790 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1791 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1792 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1793 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1794 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1795 		config->hw_vlan_insert = 0;
1796 		break;
1797 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1798 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1799 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1800 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1801 		/*
1802 		 * These NICs support VLAN insertion from WQE and
1803 		 * report the wqe_vlan_insert flag. But there is the bug
1804 		 * and PFC control may be broken, so disable feature.
1805 		 */
1806 		config->hw_vlan_insert = 0;
1807 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1808 		break;
1809 	default:
1810 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1811 		break;
1812 	}
1813 exit:
1814 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1815 }
1816 
1817 /**
1818  * Configures the metadata mask fields in the shared context.
1819  *
1820  * @param [in] dev
1821  *   Pointer to Ethernet device.
1822  */
1823 static void
1824 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1825 {
1826 	struct mlx5_priv *priv = dev->data->dev_private;
1827 	struct mlx5_ibv_shared *sh = priv->sh;
1828 	uint32_t meta, mark, reg_c0;
1829 
1830 	reg_c0 = ~priv->vport_meta_mask;
1831 	switch (priv->config.dv_xmeta_en) {
1832 	case MLX5_XMETA_MODE_LEGACY:
1833 		meta = UINT32_MAX;
1834 		mark = MLX5_FLOW_MARK_MASK;
1835 		break;
1836 	case MLX5_XMETA_MODE_META16:
1837 		meta = reg_c0 >> rte_bsf32(reg_c0);
1838 		mark = MLX5_FLOW_MARK_MASK;
1839 		break;
1840 	case MLX5_XMETA_MODE_META32:
1841 		meta = UINT32_MAX;
1842 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1843 		break;
1844 	default:
1845 		meta = 0;
1846 		mark = 0;
1847 		assert(false);
1848 		break;
1849 	}
1850 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1851 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1852 				 sh->dv_mark_mask, mark);
1853 	else
1854 		sh->dv_mark_mask = mark;
1855 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1856 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1857 				 sh->dv_meta_mask, meta);
1858 	else
1859 		sh->dv_meta_mask = meta;
1860 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1861 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1862 				 sh->dv_meta_mask, reg_c0);
1863 	else
1864 		sh->dv_regc0_mask = reg_c0;
1865 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1866 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1867 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1868 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1869 }
1870 
1871 /**
1872  * Allocate page of door-bells and register it using DevX API.
1873  *
1874  * @param [in] dev
1875  *   Pointer to Ethernet device.
1876  *
1877  * @return
1878  *   Pointer to new page on success, NULL otherwise.
1879  */
1880 static struct mlx5_devx_dbr_page *
1881 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1882 {
1883 	struct mlx5_priv *priv = dev->data->dev_private;
1884 	struct mlx5_devx_dbr_page *page;
1885 
1886 	/* Allocate space for door-bell page and management data. */
1887 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1888 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1889 	if (!page) {
1890 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1891 			dev->data->port_id);
1892 		return NULL;
1893 	}
1894 	/* Register allocated memory. */
1895 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1896 					      MLX5_DBR_PAGE_SIZE, 0);
1897 	if (!page->umem) {
1898 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1899 			dev->data->port_id);
1900 		rte_free(page);
1901 		return NULL;
1902 	}
1903 	return page;
1904 }
1905 
1906 /**
1907  * Find the next available door-bell, allocate new page if needed.
1908  *
1909  * @param [in] dev
1910  *   Pointer to Ethernet device.
1911  * @param [out] dbr_page
1912  *   Door-bell page containing the page data.
1913  *
1914  * @return
1915  *   Door-bell address offset on success, a negative error value otherwise.
1916  */
1917 int64_t
1918 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1919 {
1920 	struct mlx5_priv *priv = dev->data->dev_private;
1921 	struct mlx5_devx_dbr_page *page = NULL;
1922 	uint32_t i, j;
1923 
1924 	LIST_FOREACH(page, &priv->dbrpgs, next)
1925 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1926 			break;
1927 	if (!page) { /* No page with free door-bell exists. */
1928 		page = mlx5_alloc_dbr_page(dev);
1929 		if (!page) /* Failed to allocate new page. */
1930 			return (-1);
1931 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1932 	}
1933 	/* Loop to find bitmap part with clear bit. */
1934 	for (i = 0;
1935 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1936 	     i++)
1937 		; /* Empty. */
1938 	/* Find the first clear bit. */
1939 	j = rte_bsf64(~page->dbr_bitmap[i]);
1940 	assert(i < (MLX5_DBR_PER_PAGE / 64));
1941 	page->dbr_bitmap[i] |= (1 << j);
1942 	page->dbr_count++;
1943 	*dbr_page = page;
1944 	return (((i * 64) + j) * sizeof(uint64_t));
1945 }
1946 
1947 /**
1948  * Release a door-bell record.
1949  *
1950  * @param [in] dev
1951  *   Pointer to Ethernet device.
1952  * @param [in] umem_id
1953  *   UMEM ID of page containing the door-bell record to release.
1954  * @param [in] offset
1955  *   Offset of door-bell record in page.
1956  *
1957  * @return
1958  *   0 on success, a negative error value otherwise.
1959  */
1960 int32_t
1961 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1962 {
1963 	struct mlx5_priv *priv = dev->data->dev_private;
1964 	struct mlx5_devx_dbr_page *page = NULL;
1965 	int ret = 0;
1966 
1967 	LIST_FOREACH(page, &priv->dbrpgs, next)
1968 		/* Find the page this address belongs to. */
1969 		if (page->umem->umem_id == umem_id)
1970 			break;
1971 	if (!page)
1972 		return -EINVAL;
1973 	page->dbr_count--;
1974 	if (!page->dbr_count) {
1975 		/* Page not used, free it and remove from list. */
1976 		LIST_REMOVE(page, next);
1977 		if (page->umem)
1978 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
1979 		rte_free(page);
1980 	} else {
1981 		/* Mark in bitmap that this door-bell is not in use. */
1982 		offset /= MLX5_DBR_SIZE;
1983 		int i = offset / 64;
1984 		int j = offset % 64;
1985 
1986 		page->dbr_bitmap[i] &= ~(1 << j);
1987 	}
1988 	return ret;
1989 }
1990 
1991 /**
1992  * Check sibling device configurations.
1993  *
1994  * Sibling devices sharing the Infiniband device context
1995  * should have compatible configurations. This regards
1996  * representors and bonding slaves.
1997  *
1998  * @param priv
1999  *   Private device descriptor.
2000  * @param config
2001  *   Configuration of the device is going to be created.
2002  *
2003  * @return
2004  *   0 on success, EINVAL otherwise
2005  */
2006 static int
2007 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2008 			      struct mlx5_dev_config *config)
2009 {
2010 	struct mlx5_ibv_shared *sh = priv->sh;
2011 	struct mlx5_dev_config *sh_conf = NULL;
2012 	uint16_t port_id;
2013 
2014 	assert(sh);
2015 	/* Nothing to compare for the single/first device. */
2016 	if (sh->refcnt == 1)
2017 		return 0;
2018 	/* Find the device with shared context. */
2019 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2020 		struct mlx5_priv *opriv =
2021 			rte_eth_devices[port_id].data->dev_private;
2022 
2023 		if (opriv && opriv != priv && opriv->sh == sh) {
2024 			sh_conf = &opriv->config;
2025 			break;
2026 		}
2027 	}
2028 	if (!sh_conf)
2029 		return 0;
2030 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2031 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2032 			     " for shared %s context", sh->ibdev_name);
2033 		rte_errno = EINVAL;
2034 		return rte_errno;
2035 	}
2036 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2037 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2038 			     " for shared %s context", sh->ibdev_name);
2039 		rte_errno = EINVAL;
2040 		return rte_errno;
2041 	}
2042 	return 0;
2043 }
2044 /**
2045  * Spawn an Ethernet device from Verbs information.
2046  *
2047  * @param dpdk_dev
2048  *   Backing DPDK device.
2049  * @param spawn
2050  *   Verbs device parameters (name, port, switch_info) to spawn.
2051  * @param config
2052  *   Device configuration parameters.
2053  *
2054  * @return
2055  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2056  *   is set. The following errors are defined:
2057  *
2058  *   EBUSY: device is not supposed to be spawned.
2059  *   EEXIST: device is already spawned
2060  */
2061 static struct rte_eth_dev *
2062 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2063 	       struct mlx5_dev_spawn_data *spawn,
2064 	       struct mlx5_dev_config config)
2065 {
2066 	const struct mlx5_switch_info *switch_info = &spawn->info;
2067 	struct mlx5_ibv_shared *sh = NULL;
2068 	struct ibv_port_attr port_attr;
2069 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2070 	struct rte_eth_dev *eth_dev = NULL;
2071 	struct mlx5_priv *priv = NULL;
2072 	int err = 0;
2073 	unsigned int hw_padding = 0;
2074 	unsigned int mps;
2075 	unsigned int cqe_comp;
2076 	unsigned int cqe_pad = 0;
2077 	unsigned int tunnel_en = 0;
2078 	unsigned int mpls_en = 0;
2079 	unsigned int swp = 0;
2080 	unsigned int mprq = 0;
2081 	unsigned int mprq_min_stride_size_n = 0;
2082 	unsigned int mprq_max_stride_size_n = 0;
2083 	unsigned int mprq_min_stride_num_n = 0;
2084 	unsigned int mprq_max_stride_num_n = 0;
2085 	struct rte_ether_addr mac;
2086 	char name[RTE_ETH_NAME_MAX_LEN];
2087 	int own_domain_id = 0;
2088 	uint16_t port_id;
2089 	unsigned int i;
2090 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2091 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2092 #endif
2093 
2094 	/* Determine if this port representor is supposed to be spawned. */
2095 	if (switch_info->representor && dpdk_dev->devargs) {
2096 		struct rte_eth_devargs eth_da;
2097 
2098 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2099 		if (err) {
2100 			rte_errno = -err;
2101 			DRV_LOG(ERR, "failed to process device arguments: %s",
2102 				strerror(rte_errno));
2103 			return NULL;
2104 		}
2105 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
2106 			if (eth_da.representor_ports[i] ==
2107 			    (uint16_t)switch_info->port_name)
2108 				break;
2109 		if (i == eth_da.nb_representor_ports) {
2110 			rte_errno = EBUSY;
2111 			return NULL;
2112 		}
2113 	}
2114 	/* Build device name. */
2115 	if (spawn->pf_bond <  0) {
2116 		/* Single device. */
2117 		if (!switch_info->representor)
2118 			strlcpy(name, dpdk_dev->name, sizeof(name));
2119 		else
2120 			snprintf(name, sizeof(name), "%s_representor_%u",
2121 				 dpdk_dev->name, switch_info->port_name);
2122 	} else {
2123 		/* Bonding device. */
2124 		if (!switch_info->representor)
2125 			snprintf(name, sizeof(name), "%s_%s",
2126 				 dpdk_dev->name, spawn->ibv_dev->name);
2127 		else
2128 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
2129 				 dpdk_dev->name, spawn->ibv_dev->name,
2130 				 switch_info->port_name);
2131 	}
2132 	/* check if the device is already spawned */
2133 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2134 		rte_errno = EEXIST;
2135 		return NULL;
2136 	}
2137 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2138 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2139 		eth_dev = rte_eth_dev_attach_secondary(name);
2140 		if (eth_dev == NULL) {
2141 			DRV_LOG(ERR, "can not attach rte ethdev");
2142 			rte_errno = ENOMEM;
2143 			return NULL;
2144 		}
2145 		eth_dev->device = dpdk_dev;
2146 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2147 		err = mlx5_proc_priv_init(eth_dev);
2148 		if (err)
2149 			return NULL;
2150 		/* Receive command fd from primary process */
2151 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2152 		if (err < 0)
2153 			return NULL;
2154 		/* Remap UAR for Tx queues. */
2155 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2156 		if (err)
2157 			return NULL;
2158 		/*
2159 		 * Ethdev pointer is still required as input since
2160 		 * the primary device is not accessible from the
2161 		 * secondary process.
2162 		 */
2163 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2164 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2165 		return eth_dev;
2166 	}
2167 	/*
2168 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2169 	 * advance to create dv/verbs device context. We proceed the
2170 	 * devargs here to get ones, and later proceed devargs again
2171 	 * to override some hardware settings.
2172 	 */
2173 	err = mlx5_args(&config, dpdk_dev->devargs);
2174 	if (err) {
2175 		err = rte_errno;
2176 		DRV_LOG(ERR, "failed to process device arguments: %s",
2177 			strerror(rte_errno));
2178 		goto error;
2179 	}
2180 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2181 	if (!sh)
2182 		return NULL;
2183 	config.devx = sh->devx;
2184 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2185 	config.dest_tir = 1;
2186 #endif
2187 #ifdef HAVE_IBV_MLX5_MOD_SWP
2188 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2189 #endif
2190 	/*
2191 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2192 	 * as all ConnectX-5 devices.
2193 	 */
2194 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2195 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2196 #endif
2197 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2198 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2199 #endif
2200 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2201 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2202 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2203 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2204 			mps = MLX5_MPW_ENHANCED;
2205 		} else {
2206 			DRV_LOG(DEBUG, "MPW is supported");
2207 			mps = MLX5_MPW;
2208 		}
2209 	} else {
2210 		DRV_LOG(DEBUG, "MPW isn't supported");
2211 		mps = MLX5_MPW_DISABLED;
2212 	}
2213 #ifdef HAVE_IBV_MLX5_MOD_SWP
2214 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2215 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2216 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2217 #endif
2218 	config.swp = !!swp;
2219 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2220 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2221 		struct mlx5dv_striding_rq_caps mprq_caps =
2222 			dv_attr.striding_rq_caps;
2223 
2224 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2225 			mprq_caps.min_single_stride_log_num_of_bytes);
2226 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2227 			mprq_caps.max_single_stride_log_num_of_bytes);
2228 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2229 			mprq_caps.min_single_wqe_log_num_of_strides);
2230 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2231 			mprq_caps.max_single_wqe_log_num_of_strides);
2232 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2233 			mprq_caps.supported_qpts);
2234 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2235 		mprq = 1;
2236 		mprq_min_stride_size_n =
2237 			mprq_caps.min_single_stride_log_num_of_bytes;
2238 		mprq_max_stride_size_n =
2239 			mprq_caps.max_single_stride_log_num_of_bytes;
2240 		mprq_min_stride_num_n =
2241 			mprq_caps.min_single_wqe_log_num_of_strides;
2242 		mprq_max_stride_num_n =
2243 			mprq_caps.max_single_wqe_log_num_of_strides;
2244 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2245 						   mprq_min_stride_num_n);
2246 	}
2247 #endif
2248 	if (RTE_CACHE_LINE_SIZE == 128 &&
2249 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2250 		cqe_comp = 0;
2251 	else
2252 		cqe_comp = 1;
2253 	config.cqe_comp = cqe_comp;
2254 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2255 	/* Whether device supports 128B Rx CQE padding. */
2256 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2257 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2258 #endif
2259 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2260 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2261 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2262 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2263 			     (dv_attr.tunnel_offloads_caps &
2264 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2265 			     (dv_attr.tunnel_offloads_caps &
2266 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2267 	}
2268 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2269 		tunnel_en ? "" : "not ");
2270 #else
2271 	DRV_LOG(WARNING,
2272 		"tunnel offloading disabled due to old OFED/rdma-core version");
2273 #endif
2274 	config.tunnel_en = tunnel_en;
2275 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2276 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2277 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2278 		   (dv_attr.tunnel_offloads_caps &
2279 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2280 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2281 		mpls_en ? "" : "not ");
2282 #else
2283 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2284 		" old OFED/rdma-core version or firmware configuration");
2285 #endif
2286 	config.mpls_en = mpls_en;
2287 	/* Check port status. */
2288 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2289 	if (err) {
2290 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2291 		goto error;
2292 	}
2293 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2294 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2295 		err = EINVAL;
2296 		goto error;
2297 	}
2298 	if (port_attr.state != IBV_PORT_ACTIVE)
2299 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2300 			mlx5_glue->port_state_str(port_attr.state),
2301 			port_attr.state);
2302 	/* Allocate private eth device data. */
2303 	priv = rte_zmalloc("ethdev private structure",
2304 			   sizeof(*priv),
2305 			   RTE_CACHE_LINE_SIZE);
2306 	if (priv == NULL) {
2307 		DRV_LOG(ERR, "priv allocation failure");
2308 		err = ENOMEM;
2309 		goto error;
2310 	}
2311 	priv->sh = sh;
2312 	priv->ibv_port = spawn->ibv_port;
2313 	priv->pci_dev = spawn->pci_dev;
2314 	priv->mtu = RTE_ETHER_MTU;
2315 #ifndef RTE_ARCH_64
2316 	/* Initialize UAR access locks for 32bit implementations. */
2317 	rte_spinlock_init(&priv->uar_lock_cq);
2318 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2319 		rte_spinlock_init(&priv->uar_lock[i]);
2320 #endif
2321 	/* Some internal functions rely on Netlink sockets, open them now. */
2322 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2323 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2324 	priv->nl_sn = 0;
2325 	priv->representor = !!switch_info->representor;
2326 	priv->master = !!switch_info->master;
2327 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2328 	priv->vport_meta_tag = 0;
2329 	priv->vport_meta_mask = 0;
2330 	priv->pf_bond = spawn->pf_bond;
2331 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2332 	/*
2333 	 * The DevX port query API is implemented. E-Switch may use
2334 	 * either vport or reg_c[0] metadata register to match on
2335 	 * vport index. The engaged part of metadata register is
2336 	 * defined by mask.
2337 	 */
2338 	if (switch_info->representor || switch_info->master) {
2339 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2340 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2341 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2342 						 &devx_port);
2343 		if (err) {
2344 			DRV_LOG(WARNING,
2345 				"can't query devx port %d on device %s",
2346 				spawn->ibv_port, spawn->ibv_dev->name);
2347 			devx_port.comp_mask = 0;
2348 		}
2349 	}
2350 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2351 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2352 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2353 		if (!priv->vport_meta_mask) {
2354 			DRV_LOG(ERR, "vport zero mask for port %d"
2355 				     " on bonding device %s",
2356 				     spawn->ibv_port, spawn->ibv_dev->name);
2357 			err = ENOTSUP;
2358 			goto error;
2359 		}
2360 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2361 			DRV_LOG(ERR, "invalid vport tag for port %d"
2362 				     " on bonding device %s",
2363 				     spawn->ibv_port, spawn->ibv_dev->name);
2364 			err = ENOTSUP;
2365 			goto error;
2366 		}
2367 	}
2368 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2369 		priv->vport_id = devx_port.vport_num;
2370 	} else if (spawn->pf_bond >= 0) {
2371 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2372 			     " on bonding device %s",
2373 			     spawn->ibv_port, spawn->ibv_dev->name);
2374 		err = ENOTSUP;
2375 		goto error;
2376 	} else {
2377 		/* Suppose vport index in compatible way. */
2378 		priv->vport_id = switch_info->representor ?
2379 				 switch_info->port_name + 1 : -1;
2380 	}
2381 #else
2382 	/*
2383 	 * Kernel/rdma_core support single E-Switch per PF configurations
2384 	 * only and vport_id field contains the vport index for
2385 	 * associated VF, which is deduced from representor port name.
2386 	 * For example, let's have the IB device port 10, it has
2387 	 * attached network device eth0, which has port name attribute
2388 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2389 	 * as 3 (2+1). This assigning schema should be changed if the
2390 	 * multiple E-Switch instances per PF configurations or/and PCI
2391 	 * subfunctions are added.
2392 	 */
2393 	priv->vport_id = switch_info->representor ?
2394 			 switch_info->port_name + 1 : -1;
2395 #endif
2396 	/* representor_id field keeps the unmodified VF index. */
2397 	priv->representor_id = switch_info->representor ?
2398 			       switch_info->port_name : -1;
2399 	/*
2400 	 * Look for sibling devices in order to reuse their switch domain
2401 	 * if any, otherwise allocate one.
2402 	 */
2403 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2404 		const struct mlx5_priv *opriv =
2405 			rte_eth_devices[port_id].data->dev_private;
2406 
2407 		if (!opriv ||
2408 		    opriv->sh != priv->sh ||
2409 			opriv->domain_id ==
2410 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2411 			continue;
2412 		priv->domain_id = opriv->domain_id;
2413 		break;
2414 	}
2415 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2416 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2417 		if (err) {
2418 			err = rte_errno;
2419 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2420 				strerror(rte_errno));
2421 			goto error;
2422 		}
2423 		own_domain_id = 1;
2424 	}
2425 	/* Override some values set by hardware configuration. */
2426 	mlx5_args(&config, dpdk_dev->devargs);
2427 	err = mlx5_dev_check_sibling_config(priv, &config);
2428 	if (err)
2429 		goto error;
2430 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2431 			    IBV_DEVICE_RAW_IP_CSUM);
2432 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2433 		(config.hw_csum ? "" : "not "));
2434 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2435 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2436 	DRV_LOG(DEBUG, "counters are not supported");
2437 #endif
2438 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2439 	if (config.dv_flow_en) {
2440 		DRV_LOG(WARNING, "DV flow is not supported");
2441 		config.dv_flow_en = 0;
2442 	}
2443 #endif
2444 	config.ind_table_max_size =
2445 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2446 	/*
2447 	 * Remove this check once DPDK supports larger/variable
2448 	 * indirection tables.
2449 	 */
2450 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2451 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2452 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2453 		config.ind_table_max_size);
2454 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2455 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2456 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2457 		(config.hw_vlan_strip ? "" : "not "));
2458 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2459 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2460 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2461 		(config.hw_fcs_strip ? "" : "not "));
2462 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2463 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2464 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2465 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2466 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2467 #endif
2468 	if (config.hw_padding && !hw_padding) {
2469 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2470 		config.hw_padding = 0;
2471 	} else if (config.hw_padding) {
2472 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2473 	}
2474 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2475 		      (sh->device_attr.tso_caps.supported_qpts &
2476 		       (1 << IBV_QPT_RAW_PACKET)));
2477 	if (config.tso)
2478 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2479 	/*
2480 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2481 	 * by default.
2482 	 */
2483 	if (config.mps == MLX5_ARG_UNSET)
2484 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2485 							  MLX5_MPW_DISABLED;
2486 	else
2487 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2488 	DRV_LOG(INFO, "%sMPS is %s",
2489 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2490 		config.mps == MLX5_MPW ? "legacy " : "",
2491 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2492 	if (config.cqe_comp && !cqe_comp) {
2493 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2494 		config.cqe_comp = 0;
2495 	}
2496 	if (config.cqe_pad && !cqe_pad) {
2497 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2498 		config.cqe_pad = 0;
2499 	} else if (config.cqe_pad) {
2500 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2501 	}
2502 	if (config.devx) {
2503 		priv->counter_fallback = 0;
2504 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2505 		if (err) {
2506 			err = -err;
2507 			goto error;
2508 		}
2509 		if (!config.hca_attr.flow_counters_dump)
2510 			priv->counter_fallback = 1;
2511 #ifndef HAVE_IBV_DEVX_ASYNC
2512 		priv->counter_fallback = 1;
2513 #endif
2514 		if (priv->counter_fallback)
2515 			DRV_LOG(INFO, "Use fall-back DV counter management");
2516 		/* Check for LRO support. */
2517 		if (config.dest_tir && config.hca_attr.lro_cap &&
2518 		    config.dv_flow_en) {
2519 			/* TBD check tunnel lro caps. */
2520 			config.lro.supported = config.hca_attr.lro_cap;
2521 			DRV_LOG(DEBUG, "Device supports LRO");
2522 			/*
2523 			 * If LRO timeout is not configured by application,
2524 			 * use the minimal supported value.
2525 			 */
2526 			if (!config.lro.timeout)
2527 				config.lro.timeout =
2528 				config.hca_attr.lro_timer_supported_periods[0];
2529 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2530 				config.lro.timeout);
2531 		}
2532 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2533 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2534 		    config.dv_flow_en) {
2535 			uint8_t reg_c_mask =
2536 				config.hca_attr.qos.flow_meter_reg_c_ids;
2537 			/*
2538 			 * Meter needs two REG_C's for color match and pre-sfx
2539 			 * flow match. Here get the REG_C for color match.
2540 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2541 			 */
2542 			reg_c_mask &= 0xfc;
2543 			if (__builtin_popcount(reg_c_mask) < 1) {
2544 				priv->mtr_en = 0;
2545 				DRV_LOG(WARNING, "No available register for"
2546 					" meter.");
2547 			} else {
2548 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2549 						      REG_C_0;
2550 				priv->mtr_en = 1;
2551 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2552 					priv->mtr_color_reg);
2553 			}
2554 		}
2555 #endif
2556 	}
2557 	if (config.mprq.enabled && mprq) {
2558 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2559 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2560 			config.mprq.stride_num_n =
2561 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2562 					mprq_min_stride_num_n);
2563 			DRV_LOG(WARNING,
2564 				"the number of strides"
2565 				" for Multi-Packet RQ is out of range,"
2566 				" setting default value (%u)",
2567 				1 << config.mprq.stride_num_n);
2568 		}
2569 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2570 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2571 	} else if (config.mprq.enabled && !mprq) {
2572 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2573 		config.mprq.enabled = 0;
2574 	}
2575 	if (config.max_dump_files_num == 0)
2576 		config.max_dump_files_num = 128;
2577 	eth_dev = rte_eth_dev_allocate(name);
2578 	if (eth_dev == NULL) {
2579 		DRV_LOG(ERR, "can not allocate rte ethdev");
2580 		err = ENOMEM;
2581 		goto error;
2582 	}
2583 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2584 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2585 	if (priv->representor) {
2586 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2587 		eth_dev->data->representor_id = priv->representor_id;
2588 	}
2589 	/*
2590 	 * Store associated network device interface index. This index
2591 	 * is permanent throughout the lifetime of device. So, we may store
2592 	 * the ifindex here and use the cached value further.
2593 	 */
2594 	assert(spawn->ifindex);
2595 	priv->if_index = spawn->ifindex;
2596 	eth_dev->data->dev_private = priv;
2597 	priv->dev_data = eth_dev->data;
2598 	eth_dev->data->mac_addrs = priv->mac;
2599 	eth_dev->device = dpdk_dev;
2600 	/* Configure the first MAC address by default. */
2601 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2602 		DRV_LOG(ERR,
2603 			"port %u cannot get MAC address, is mlx5_en"
2604 			" loaded? (errno: %s)",
2605 			eth_dev->data->port_id, strerror(rte_errno));
2606 		err = ENODEV;
2607 		goto error;
2608 	}
2609 	DRV_LOG(INFO,
2610 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2611 		eth_dev->data->port_id,
2612 		mac.addr_bytes[0], mac.addr_bytes[1],
2613 		mac.addr_bytes[2], mac.addr_bytes[3],
2614 		mac.addr_bytes[4], mac.addr_bytes[5]);
2615 #ifndef NDEBUG
2616 	{
2617 		char ifname[IF_NAMESIZE];
2618 
2619 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2620 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2621 				eth_dev->data->port_id, ifname);
2622 		else
2623 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2624 				eth_dev->data->port_id);
2625 	}
2626 #endif
2627 	/* Get actual MTU if possible. */
2628 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2629 	if (err) {
2630 		err = rte_errno;
2631 		goto error;
2632 	}
2633 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2634 		priv->mtu);
2635 	/* Initialize burst functions to prevent crashes before link-up. */
2636 	eth_dev->rx_pkt_burst = removed_rx_burst;
2637 	eth_dev->tx_pkt_burst = removed_tx_burst;
2638 	eth_dev->dev_ops = &mlx5_dev_ops;
2639 	/* Register MAC address. */
2640 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2641 	if (config.vf && config.vf_nl_en)
2642 		mlx5_nl_mac_addr_sync(eth_dev);
2643 	TAILQ_INIT(&priv->flows);
2644 	TAILQ_INIT(&priv->ctrl_flows);
2645 	TAILQ_INIT(&priv->flow_meters);
2646 	TAILQ_INIT(&priv->flow_meter_profiles);
2647 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2648 	struct mlx5dv_ctx_allocators alctr = {
2649 		.alloc = &mlx5_alloc_verbs_buf,
2650 		.free = &mlx5_free_verbs_buf,
2651 		.data = priv,
2652 	};
2653 	mlx5_glue->dv_set_context_attr(sh->ctx,
2654 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2655 				       (void *)((uintptr_t)&alctr));
2656 	/* Bring Ethernet device up. */
2657 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2658 		eth_dev->data->port_id);
2659 	mlx5_set_link_up(eth_dev);
2660 	/*
2661 	 * Even though the interrupt handler is not installed yet,
2662 	 * interrupts will still trigger on the async_fd from
2663 	 * Verbs context returned by ibv_open_device().
2664 	 */
2665 	mlx5_link_update(eth_dev, 0);
2666 #ifdef HAVE_MLX5DV_DR_ESWITCH
2667 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2668 	      (switch_info->representor || switch_info->master)))
2669 		config.dv_esw_en = 0;
2670 #else
2671 	config.dv_esw_en = 0;
2672 #endif
2673 	/* Detect minimal data bytes to inline. */
2674 	mlx5_set_min_inline(spawn, &config);
2675 	/* Store device configuration on private structure. */
2676 	priv->config = config;
2677 	/* Create context for virtual machine VLAN workaround. */
2678 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2679 	if (config.dv_flow_en) {
2680 		err = mlx5_alloc_shared_dr(priv);
2681 		if (err)
2682 			goto error;
2683 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc();
2684 		if (!priv->qrss_id_pool) {
2685 			DRV_LOG(ERR, "can't create flow id pool");
2686 			err = ENOMEM;
2687 			goto error;
2688 		}
2689 	}
2690 	/* Supported Verbs flow priority number detection. */
2691 	err = mlx5_flow_discover_priorities(eth_dev);
2692 	if (err < 0) {
2693 		err = -err;
2694 		goto error;
2695 	}
2696 	priv->config.flow_prio = err;
2697 	if (!priv->config.dv_esw_en &&
2698 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2699 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2700 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2701 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2702 	}
2703 	mlx5_set_metadata_mask(eth_dev);
2704 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2705 	    !priv->sh->dv_regc0_mask) {
2706 		DRV_LOG(ERR, "metadata mode %u is not supported "
2707 			     "(no metadata reg_c[0] is available)",
2708 			     priv->config.dv_xmeta_en);
2709 			err = ENOTSUP;
2710 			goto error;
2711 	}
2712 	/* Query availibility of metadata reg_c's. */
2713 	err = mlx5_flow_discover_mreg_c(eth_dev);
2714 	if (err < 0) {
2715 		err = -err;
2716 		goto error;
2717 	}
2718 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2719 		DRV_LOG(DEBUG,
2720 			"port %u extensive metadata register is not supported",
2721 			eth_dev->data->port_id);
2722 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2723 			DRV_LOG(ERR, "metadata mode %u is not supported "
2724 				     "(no metadata registers available)",
2725 				     priv->config.dv_xmeta_en);
2726 			err = ENOTSUP;
2727 			goto error;
2728 		}
2729 	}
2730 	if (priv->config.dv_flow_en &&
2731 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2732 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
2733 	    priv->sh->dv_regc0_mask) {
2734 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2735 						      MLX5_FLOW_MREG_HTABLE_SZ);
2736 		if (!priv->mreg_cp_tbl) {
2737 			err = ENOMEM;
2738 			goto error;
2739 		}
2740 	}
2741 	return eth_dev;
2742 error:
2743 	if (priv) {
2744 		if (priv->mreg_cp_tbl)
2745 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2746 		if (priv->sh)
2747 			mlx5_free_shared_dr(priv);
2748 		if (priv->nl_socket_route >= 0)
2749 			close(priv->nl_socket_route);
2750 		if (priv->nl_socket_rdma >= 0)
2751 			close(priv->nl_socket_rdma);
2752 		if (priv->vmwa_context)
2753 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2754 		if (priv->qrss_id_pool)
2755 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
2756 		if (own_domain_id)
2757 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2758 		rte_free(priv);
2759 		if (eth_dev != NULL)
2760 			eth_dev->data->dev_private = NULL;
2761 	}
2762 	if (eth_dev != NULL) {
2763 		/* mac_addrs must not be freed alone because part of dev_private */
2764 		eth_dev->data->mac_addrs = NULL;
2765 		rte_eth_dev_release_port(eth_dev);
2766 	}
2767 	if (sh)
2768 		mlx5_free_shared_ibctx(sh);
2769 	assert(err > 0);
2770 	rte_errno = err;
2771 	return NULL;
2772 }
2773 
2774 /**
2775  * Comparison callback to sort device data.
2776  *
2777  * This is meant to be used with qsort().
2778  *
2779  * @param a[in]
2780  *   Pointer to pointer to first data object.
2781  * @param b[in]
2782  *   Pointer to pointer to second data object.
2783  *
2784  * @return
2785  *   0 if both objects are equal, less than 0 if the first argument is less
2786  *   than the second, greater than 0 otherwise.
2787  */
2788 static int
2789 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2790 {
2791 	const struct mlx5_switch_info *si_a =
2792 		&((const struct mlx5_dev_spawn_data *)a)->info;
2793 	const struct mlx5_switch_info *si_b =
2794 		&((const struct mlx5_dev_spawn_data *)b)->info;
2795 	int ret;
2796 
2797 	/* Master device first. */
2798 	ret = si_b->master - si_a->master;
2799 	if (ret)
2800 		return ret;
2801 	/* Then representor devices. */
2802 	ret = si_b->representor - si_a->representor;
2803 	if (ret)
2804 		return ret;
2805 	/* Unidentified devices come last in no specific order. */
2806 	if (!si_a->representor)
2807 		return 0;
2808 	/* Order representors by name. */
2809 	return si_a->port_name - si_b->port_name;
2810 }
2811 
2812 /**
2813  * Match PCI information for possible slaves of bonding device.
2814  *
2815  * @param[in] ibv_dev
2816  *   Pointer to Infiniband device structure.
2817  * @param[in] pci_dev
2818  *   Pointer to PCI device structure to match PCI address.
2819  * @param[in] nl_rdma
2820  *   Netlink RDMA group socket handle.
2821  *
2822  * @return
2823  *   negative value if no bonding device found, otherwise
2824  *   positive index of slave PF in bonding.
2825  */
2826 static int
2827 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2828 			   const struct rte_pci_device *pci_dev,
2829 			   int nl_rdma)
2830 {
2831 	char ifname[IF_NAMESIZE + 1];
2832 	unsigned int ifindex;
2833 	unsigned int np, i;
2834 	FILE *file = NULL;
2835 	int pf = -1;
2836 
2837 	/*
2838 	 * Try to get master device name. If something goes
2839 	 * wrong suppose the lack of kernel support and no
2840 	 * bonding devices.
2841 	 */
2842 	if (nl_rdma < 0)
2843 		return -1;
2844 	if (!strstr(ibv_dev->name, "bond"))
2845 		return -1;
2846 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2847 	if (!np)
2848 		return -1;
2849 	/*
2850 	 * The Master device might not be on the predefined
2851 	 * port (not on port index 1, it is not garanted),
2852 	 * we have to scan all Infiniband device port and
2853 	 * find master.
2854 	 */
2855 	for (i = 1; i <= np; ++i) {
2856 		/* Check whether Infiniband port is populated. */
2857 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2858 		if (!ifindex)
2859 			continue;
2860 		if (!if_indextoname(ifindex, ifname))
2861 			continue;
2862 		/* Try to read bonding slave names from sysfs. */
2863 		MKSTR(slaves,
2864 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2865 		file = fopen(slaves, "r");
2866 		if (file)
2867 			break;
2868 	}
2869 	if (!file)
2870 		return -1;
2871 	/* Use safe format to check maximal buffer length. */
2872 	assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2873 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2874 		char tmp_str[IF_NAMESIZE + 32];
2875 		struct rte_pci_addr pci_addr;
2876 		struct mlx5_switch_info	info;
2877 
2878 		/* Process slave interface names in the loop. */
2879 		snprintf(tmp_str, sizeof(tmp_str),
2880 			 "/sys/class/net/%s", ifname);
2881 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2882 			DRV_LOG(WARNING, "can not get PCI address"
2883 					 " for netdev \"%s\"", ifname);
2884 			continue;
2885 		}
2886 		if (pci_dev->addr.domain != pci_addr.domain ||
2887 		    pci_dev->addr.bus != pci_addr.bus ||
2888 		    pci_dev->addr.devid != pci_addr.devid ||
2889 		    pci_dev->addr.function != pci_addr.function)
2890 			continue;
2891 		/* Slave interface PCI address match found. */
2892 		fclose(file);
2893 		snprintf(tmp_str, sizeof(tmp_str),
2894 			 "/sys/class/net/%s/phys_port_name", ifname);
2895 		file = fopen(tmp_str, "rb");
2896 		if (!file)
2897 			break;
2898 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2899 		if (fscanf(file, "%32s", tmp_str) == 1)
2900 			mlx5_translate_port_name(tmp_str, &info);
2901 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2902 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2903 			pf = info.port_name;
2904 		break;
2905 	}
2906 	if (file)
2907 		fclose(file);
2908 	return pf;
2909 }
2910 
2911 /**
2912  * DPDK callback to register a PCI device.
2913  *
2914  * This function spawns Ethernet devices out of a given PCI device.
2915  *
2916  * @param[in] pci_drv
2917  *   PCI driver structure (mlx5_driver).
2918  * @param[in] pci_dev
2919  *   PCI device information.
2920  *
2921  * @return
2922  *   0 on success, a negative errno value otherwise and rte_errno is set.
2923  */
2924 static int
2925 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2926 	       struct rte_pci_device *pci_dev)
2927 {
2928 	struct ibv_device **ibv_list;
2929 	/*
2930 	 * Number of found IB Devices matching with requested PCI BDF.
2931 	 * nd != 1 means there are multiple IB devices over the same
2932 	 * PCI device and we have representors and master.
2933 	 */
2934 	unsigned int nd = 0;
2935 	/*
2936 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2937 	 * we have the single multiport IB device, and there may be
2938 	 * representors attached to some of found ports.
2939 	 */
2940 	unsigned int np = 0;
2941 	/*
2942 	 * Number of DPDK ethernet devices to Spawn - either over
2943 	 * multiple IB devices or multiple ports of single IB device.
2944 	 * Actually this is the number of iterations to spawn.
2945 	 */
2946 	unsigned int ns = 0;
2947 	/*
2948 	 * Bonding device
2949 	 *   < 0 - no bonding device (single one)
2950 	 *  >= 0 - bonding device (value is slave PF index)
2951 	 */
2952 	int bd = -1;
2953 	struct mlx5_dev_spawn_data *list = NULL;
2954 	struct mlx5_dev_config dev_config;
2955 	int ret;
2956 
2957 	ret = mlx5_init_once();
2958 	if (ret) {
2959 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2960 			strerror(rte_errno));
2961 		return -rte_errno;
2962 	}
2963 	assert(pci_drv == &mlx5_driver);
2964 	errno = 0;
2965 	ibv_list = mlx5_glue->get_device_list(&ret);
2966 	if (!ibv_list) {
2967 		rte_errno = errno ? errno : ENOSYS;
2968 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2969 		return -rte_errno;
2970 	}
2971 	/*
2972 	 * First scan the list of all Infiniband devices to find
2973 	 * matching ones, gathering into the list.
2974 	 */
2975 	struct ibv_device *ibv_match[ret + 1];
2976 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2977 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2978 	unsigned int i;
2979 
2980 	while (ret-- > 0) {
2981 		struct rte_pci_addr pci_addr;
2982 
2983 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2984 		bd = mlx5_device_bond_pci_match
2985 				(ibv_list[ret], pci_dev, nl_rdma);
2986 		if (bd >= 0) {
2987 			/*
2988 			 * Bonding device detected. Only one match is allowed,
2989 			 * the bonding is supported over multi-port IB device,
2990 			 * there should be no matches on representor PCI
2991 			 * functions or non VF LAG bonding devices with
2992 			 * specified address.
2993 			 */
2994 			if (nd) {
2995 				DRV_LOG(ERR,
2996 					"multiple PCI match on bonding device"
2997 					"\"%s\" found", ibv_list[ret]->name);
2998 				rte_errno = ENOENT;
2999 				ret = -rte_errno;
3000 				goto exit;
3001 			}
3002 			DRV_LOG(INFO, "PCI information matches for"
3003 				      " slave %d bonding device \"%s\"",
3004 				      bd, ibv_list[ret]->name);
3005 			ibv_match[nd++] = ibv_list[ret];
3006 			break;
3007 		}
3008 		if (mlx5_dev_to_pci_addr
3009 			(ibv_list[ret]->ibdev_path, &pci_addr))
3010 			continue;
3011 		if (pci_dev->addr.domain != pci_addr.domain ||
3012 		    pci_dev->addr.bus != pci_addr.bus ||
3013 		    pci_dev->addr.devid != pci_addr.devid ||
3014 		    pci_dev->addr.function != pci_addr.function)
3015 			continue;
3016 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3017 			ibv_list[ret]->name);
3018 		ibv_match[nd++] = ibv_list[ret];
3019 	}
3020 	ibv_match[nd] = NULL;
3021 	if (!nd) {
3022 		/* No device matches, just complain and bail out. */
3023 		DRV_LOG(WARNING,
3024 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
3025 			" are kernel drivers loaded?",
3026 			pci_dev->addr.domain, pci_dev->addr.bus,
3027 			pci_dev->addr.devid, pci_dev->addr.function);
3028 		rte_errno = ENOENT;
3029 		ret = -rte_errno;
3030 		goto exit;
3031 	}
3032 	if (nd == 1) {
3033 		/*
3034 		 * Found single matching device may have multiple ports.
3035 		 * Each port may be representor, we have to check the port
3036 		 * number and check the representors existence.
3037 		 */
3038 		if (nl_rdma >= 0)
3039 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3040 		if (!np)
3041 			DRV_LOG(WARNING, "can not get IB device \"%s\""
3042 					 " ports number", ibv_match[0]->name);
3043 		if (bd >= 0 && !np) {
3044 			DRV_LOG(ERR, "can not get ports"
3045 				     " for bonding device");
3046 			rte_errno = ENOENT;
3047 			ret = -rte_errno;
3048 			goto exit;
3049 		}
3050 	}
3051 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3052 	if (bd >= 0) {
3053 		/*
3054 		 * This may happen if there is VF LAG kernel support and
3055 		 * application is compiled with older rdma_core library.
3056 		 */
3057 		DRV_LOG(ERR,
3058 			"No kernel/verbs support for VF LAG bonding found.");
3059 		rte_errno = ENOTSUP;
3060 		ret = -rte_errno;
3061 		goto exit;
3062 	}
3063 #endif
3064 	/*
3065 	 * Now we can determine the maximal
3066 	 * amount of devices to be spawned.
3067 	 */
3068 	list = rte_zmalloc("device spawn data",
3069 			 sizeof(struct mlx5_dev_spawn_data) *
3070 			 (np ? np : nd),
3071 			 RTE_CACHE_LINE_SIZE);
3072 	if (!list) {
3073 		DRV_LOG(ERR, "spawn data array allocation failure");
3074 		rte_errno = ENOMEM;
3075 		ret = -rte_errno;
3076 		goto exit;
3077 	}
3078 	if (bd >= 0 || np > 1) {
3079 		/*
3080 		 * Single IB device with multiple ports found,
3081 		 * it may be E-Switch master device and representors.
3082 		 * We have to perform identification trough the ports.
3083 		 */
3084 		assert(nl_rdma >= 0);
3085 		assert(ns == 0);
3086 		assert(nd == 1);
3087 		assert(np);
3088 		for (i = 1; i <= np; ++i) {
3089 			list[ns].max_port = np;
3090 			list[ns].ibv_port = i;
3091 			list[ns].ibv_dev = ibv_match[0];
3092 			list[ns].eth_dev = NULL;
3093 			list[ns].pci_dev = pci_dev;
3094 			list[ns].pf_bond = bd;
3095 			list[ns].ifindex = mlx5_nl_ifindex
3096 					(nl_rdma, list[ns].ibv_dev->name, i);
3097 			if (!list[ns].ifindex) {
3098 				/*
3099 				 * No network interface index found for the
3100 				 * specified port, it means there is no
3101 				 * representor on this port. It's OK,
3102 				 * there can be disabled ports, for example
3103 				 * if sriov_numvfs < sriov_totalvfs.
3104 				 */
3105 				continue;
3106 			}
3107 			ret = -1;
3108 			if (nl_route >= 0)
3109 				ret = mlx5_nl_switch_info
3110 					       (nl_route,
3111 						list[ns].ifindex,
3112 						&list[ns].info);
3113 			if (ret || (!list[ns].info.representor &&
3114 				    !list[ns].info.master)) {
3115 				/*
3116 				 * We failed to recognize representors with
3117 				 * Netlink, let's try to perform the task
3118 				 * with sysfs.
3119 				 */
3120 				ret =  mlx5_sysfs_switch_info
3121 						(list[ns].ifindex,
3122 						 &list[ns].info);
3123 			}
3124 			if (!ret && bd >= 0) {
3125 				switch (list[ns].info.name_type) {
3126 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3127 					if (list[ns].info.port_name == bd)
3128 						ns++;
3129 					break;
3130 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3131 					if (list[ns].info.pf_num == bd)
3132 						ns++;
3133 					break;
3134 				default:
3135 					break;
3136 				}
3137 				continue;
3138 			}
3139 			if (!ret && (list[ns].info.representor ^
3140 				     list[ns].info.master))
3141 				ns++;
3142 		}
3143 		if (!ns) {
3144 			DRV_LOG(ERR,
3145 				"unable to recognize master/representors"
3146 				" on the IB device with multiple ports");
3147 			rte_errno = ENOENT;
3148 			ret = -rte_errno;
3149 			goto exit;
3150 		}
3151 	} else {
3152 		/*
3153 		 * The existence of several matching entries (nd > 1) means
3154 		 * port representors have been instantiated. No existing Verbs
3155 		 * call nor sysfs entries can tell them apart, this can only
3156 		 * be done through Netlink calls assuming kernel drivers are
3157 		 * recent enough to support them.
3158 		 *
3159 		 * In the event of identification failure through Netlink,
3160 		 * try again through sysfs, then:
3161 		 *
3162 		 * 1. A single IB device matches (nd == 1) with single
3163 		 *    port (np=0/1) and is not a representor, assume
3164 		 *    no switch support.
3165 		 *
3166 		 * 2. Otherwise no safe assumptions can be made;
3167 		 *    complain louder and bail out.
3168 		 */
3169 		np = 1;
3170 		for (i = 0; i != nd; ++i) {
3171 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3172 			list[ns].max_port = 1;
3173 			list[ns].ibv_port = 1;
3174 			list[ns].ibv_dev = ibv_match[i];
3175 			list[ns].eth_dev = NULL;
3176 			list[ns].pci_dev = pci_dev;
3177 			list[ns].pf_bond = -1;
3178 			list[ns].ifindex = 0;
3179 			if (nl_rdma >= 0)
3180 				list[ns].ifindex = mlx5_nl_ifindex
3181 					(nl_rdma, list[ns].ibv_dev->name, 1);
3182 			if (!list[ns].ifindex) {
3183 				char ifname[IF_NAMESIZE];
3184 
3185 				/*
3186 				 * Netlink failed, it may happen with old
3187 				 * ib_core kernel driver (before 4.16).
3188 				 * We can assume there is old driver because
3189 				 * here we are processing single ports IB
3190 				 * devices. Let's try sysfs to retrieve
3191 				 * the ifindex. The method works for
3192 				 * master device only.
3193 				 */
3194 				if (nd > 1) {
3195 					/*
3196 					 * Multiple devices found, assume
3197 					 * representors, can not distinguish
3198 					 * master/representor and retrieve
3199 					 * ifindex via sysfs.
3200 					 */
3201 					continue;
3202 				}
3203 				ret = mlx5_get_master_ifname
3204 					(ibv_match[i]->ibdev_path, &ifname);
3205 				if (!ret)
3206 					list[ns].ifindex =
3207 						if_nametoindex(ifname);
3208 				if (!list[ns].ifindex) {
3209 					/*
3210 					 * No network interface index found
3211 					 * for the specified device, it means
3212 					 * there it is neither representor
3213 					 * nor master.
3214 					 */
3215 					continue;
3216 				}
3217 			}
3218 			ret = -1;
3219 			if (nl_route >= 0)
3220 				ret = mlx5_nl_switch_info
3221 					       (nl_route,
3222 						list[ns].ifindex,
3223 						&list[ns].info);
3224 			if (ret || (!list[ns].info.representor &&
3225 				    !list[ns].info.master)) {
3226 				/*
3227 				 * We failed to recognize representors with
3228 				 * Netlink, let's try to perform the task
3229 				 * with sysfs.
3230 				 */
3231 				ret =  mlx5_sysfs_switch_info
3232 						(list[ns].ifindex,
3233 						 &list[ns].info);
3234 			}
3235 			if (!ret && (list[ns].info.representor ^
3236 				     list[ns].info.master)) {
3237 				ns++;
3238 			} else if ((nd == 1) &&
3239 				   !list[ns].info.representor &&
3240 				   !list[ns].info.master) {
3241 				/*
3242 				 * Single IB device with
3243 				 * one physical port and
3244 				 * attached network device.
3245 				 * May be SRIOV is not enabled
3246 				 * or there is no representors.
3247 				 */
3248 				DRV_LOG(INFO, "no E-Switch support detected");
3249 				ns++;
3250 				break;
3251 			}
3252 		}
3253 		if (!ns) {
3254 			DRV_LOG(ERR,
3255 				"unable to recognize master/representors"
3256 				" on the multiple IB devices");
3257 			rte_errno = ENOENT;
3258 			ret = -rte_errno;
3259 			goto exit;
3260 		}
3261 	}
3262 	assert(ns);
3263 	/*
3264 	 * Sort list to probe devices in natural order for users convenience
3265 	 * (i.e. master first, then representors from lowest to highest ID).
3266 	 */
3267 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3268 	/* Default configuration. */
3269 	dev_config = (struct mlx5_dev_config){
3270 		.hw_padding = 0,
3271 		.mps = MLX5_ARG_UNSET,
3272 		.dbnc = MLX5_ARG_UNSET,
3273 		.rx_vec_en = 1,
3274 		.txq_inline_max = MLX5_ARG_UNSET,
3275 		.txq_inline_min = MLX5_ARG_UNSET,
3276 		.txq_inline_mpw = MLX5_ARG_UNSET,
3277 		.txqs_inline = MLX5_ARG_UNSET,
3278 		.vf_nl_en = 1,
3279 		.mr_ext_memseg_en = 1,
3280 		.mprq = {
3281 			.enabled = 0, /* Disabled by default. */
3282 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3283 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3284 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3285 		},
3286 		.dv_esw_en = 1,
3287 		.dv_flow_en = 1,
3288 	};
3289 	/* Device specific configuration. */
3290 	switch (pci_dev->id.device_id) {
3291 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3292 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3293 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3294 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3295 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3296 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3297 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3298 		dev_config.vf = 1;
3299 		break;
3300 	default:
3301 		break;
3302 	}
3303 	for (i = 0; i != ns; ++i) {
3304 		uint32_t restore;
3305 
3306 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3307 						 &list[i],
3308 						 dev_config);
3309 		if (!list[i].eth_dev) {
3310 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3311 				break;
3312 			/* Device is disabled or already spawned. Ignore it. */
3313 			continue;
3314 		}
3315 		restore = list[i].eth_dev->data->dev_flags;
3316 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3317 		/* Restore non-PCI flags cleared by the above call. */
3318 		list[i].eth_dev->data->dev_flags |= restore;
3319 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3320 		rte_eth_dev_probing_finish(list[i].eth_dev);
3321 	}
3322 	if (i != ns) {
3323 		DRV_LOG(ERR,
3324 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3325 			" encountering an error: %s",
3326 			pci_dev->addr.domain, pci_dev->addr.bus,
3327 			pci_dev->addr.devid, pci_dev->addr.function,
3328 			strerror(rte_errno));
3329 		ret = -rte_errno;
3330 		/* Roll back. */
3331 		while (i--) {
3332 			if (!list[i].eth_dev)
3333 				continue;
3334 			mlx5_dev_close(list[i].eth_dev);
3335 			/* mac_addrs must not be freed because in dev_private */
3336 			list[i].eth_dev->data->mac_addrs = NULL;
3337 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3338 		}
3339 		/* Restore original error. */
3340 		rte_errno = -ret;
3341 	} else {
3342 		ret = 0;
3343 	}
3344 exit:
3345 	/*
3346 	 * Do the routine cleanup:
3347 	 * - close opened Netlink sockets
3348 	 * - free allocated spawn data array
3349 	 * - free the Infiniband device list
3350 	 */
3351 	if (nl_rdma >= 0)
3352 		close(nl_rdma);
3353 	if (nl_route >= 0)
3354 		close(nl_route);
3355 	if (list)
3356 		rte_free(list);
3357 	assert(ibv_list);
3358 	mlx5_glue->free_device_list(ibv_list);
3359 	return ret;
3360 }
3361 
3362 /**
3363  * Look for the ethernet device belonging to mlx5 driver.
3364  *
3365  * @param[in] port_id
3366  *   port_id to start looking for device.
3367  * @param[in] pci_dev
3368  *   Pointer to the hint PCI device. When device is being probed
3369  *   the its siblings (master and preceding representors might
3370  *   not have assigned driver yet (because the mlx5_pci_probe()
3371  *   is not completed yet, for this case match on hint PCI
3372  *   device may be used to detect sibling device.
3373  *
3374  * @return
3375  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3376  */
3377 uint16_t
3378 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3379 {
3380 	while (port_id < RTE_MAX_ETHPORTS) {
3381 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3382 
3383 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3384 		    dev->device &&
3385 		    (dev->device == &pci_dev->device ||
3386 		     (dev->device->driver &&
3387 		     dev->device->driver->name &&
3388 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3389 			break;
3390 		port_id++;
3391 	}
3392 	if (port_id >= RTE_MAX_ETHPORTS)
3393 		return RTE_MAX_ETHPORTS;
3394 	return port_id;
3395 }
3396 
3397 /**
3398  * DPDK callback to remove a PCI device.
3399  *
3400  * This function removes all Ethernet devices belong to a given PCI device.
3401  *
3402  * @param[in] pci_dev
3403  *   Pointer to the PCI device.
3404  *
3405  * @return
3406  *   0 on success, the function cannot fail.
3407  */
3408 static int
3409 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3410 {
3411 	uint16_t port_id;
3412 
3413 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3414 		rte_eth_dev_close(port_id);
3415 	return 0;
3416 }
3417 
3418 static const struct rte_pci_id mlx5_pci_id_map[] = {
3419 	{
3420 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3421 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3422 	},
3423 	{
3424 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3425 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3426 	},
3427 	{
3428 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3429 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3430 	},
3431 	{
3432 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3433 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3434 	},
3435 	{
3436 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3437 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3438 	},
3439 	{
3440 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3441 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3442 	},
3443 	{
3444 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3445 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3446 	},
3447 	{
3448 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3449 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3450 	},
3451 	{
3452 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3453 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3454 	},
3455 	{
3456 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3457 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3458 	},
3459 	{
3460 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3461 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3462 	},
3463 	{
3464 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3465 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3466 	},
3467 	{
3468 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3469 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3470 	},
3471 	{
3472 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3473 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3474 	},
3475 	{
3476 		.vendor_id = 0
3477 	}
3478 };
3479 
3480 static struct rte_pci_driver mlx5_driver = {
3481 	.driver = {
3482 		.name = MLX5_DRIVER_NAME
3483 	},
3484 	.id_table = mlx5_pci_id_map,
3485 	.probe = mlx5_pci_probe,
3486 	.remove = mlx5_pci_remove,
3487 	.dma_map = mlx5_dma_map,
3488 	.dma_unmap = mlx5_dma_unmap,
3489 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3490 		     RTE_PCI_DRV_PROBE_AGAIN,
3491 };
3492 
3493 #ifdef RTE_IBVERBS_LINK_DLOPEN
3494 
3495 /**
3496  * Suffix RTE_EAL_PMD_PATH with "-glue".
3497  *
3498  * This function performs a sanity check on RTE_EAL_PMD_PATH before
3499  * suffixing its last component.
3500  *
3501  * @param buf[out]
3502  *   Output buffer, should be large enough otherwise NULL is returned.
3503  * @param size
3504  *   Size of @p out.
3505  *
3506  * @return
3507  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
3508  */
3509 static char *
3510 mlx5_glue_path(char *buf, size_t size)
3511 {
3512 	static const char *const bad[] = { "/", ".", "..", NULL };
3513 	const char *path = RTE_EAL_PMD_PATH;
3514 	size_t len = strlen(path);
3515 	size_t off;
3516 	int i;
3517 
3518 	while (len && path[len - 1] == '/')
3519 		--len;
3520 	for (off = len; off && path[off - 1] != '/'; --off)
3521 		;
3522 	for (i = 0; bad[i]; ++i)
3523 		if (!strncmp(path + off, bad[i], (int)(len - off)))
3524 			goto error;
3525 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
3526 	if (i == -1 || (size_t)i >= size)
3527 		goto error;
3528 	return buf;
3529 error:
3530 	DRV_LOG(ERR,
3531 		"unable to append \"-glue\" to last component of"
3532 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
3533 		" please re-configure DPDK");
3534 	return NULL;
3535 }
3536 
3537 /**
3538  * Initialization routine for run-time dependency on rdma-core.
3539  */
3540 static int
3541 mlx5_glue_init(void)
3542 {
3543 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
3544 	const char *path[] = {
3545 		/*
3546 		 * A basic security check is necessary before trusting
3547 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
3548 		 */
3549 		(geteuid() == getuid() && getegid() == getgid() ?
3550 		 getenv("MLX5_GLUE_PATH") : NULL),
3551 		/*
3552 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
3553 		 * variant, otherwise let dlopen() look up libraries on its
3554 		 * own.
3555 		 */
3556 		(*RTE_EAL_PMD_PATH ?
3557 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
3558 	};
3559 	unsigned int i = 0;
3560 	void *handle = NULL;
3561 	void **sym;
3562 	const char *dlmsg;
3563 
3564 	while (!handle && i != RTE_DIM(path)) {
3565 		const char *end;
3566 		size_t len;
3567 		int ret;
3568 
3569 		if (!path[i]) {
3570 			++i;
3571 			continue;
3572 		}
3573 		end = strpbrk(path[i], ":;");
3574 		if (!end)
3575 			end = path[i] + strlen(path[i]);
3576 		len = end - path[i];
3577 		ret = 0;
3578 		do {
3579 			char name[ret + 1];
3580 
3581 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
3582 				       (int)len, path[i],
3583 				       (!len || *(end - 1) == '/') ? "" : "/");
3584 			if (ret == -1)
3585 				break;
3586 			if (sizeof(name) != (size_t)ret + 1)
3587 				continue;
3588 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
3589 				name);
3590 			handle = dlopen(name, RTLD_LAZY);
3591 			break;
3592 		} while (1);
3593 		path[i] = end + 1;
3594 		if (!*end)
3595 			++i;
3596 	}
3597 	if (!handle) {
3598 		rte_errno = EINVAL;
3599 		dlmsg = dlerror();
3600 		if (dlmsg)
3601 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
3602 		goto glue_error;
3603 	}
3604 	sym = dlsym(handle, "mlx5_glue");
3605 	if (!sym || !*sym) {
3606 		rte_errno = EINVAL;
3607 		dlmsg = dlerror();
3608 		if (dlmsg)
3609 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
3610 		goto glue_error;
3611 	}
3612 	mlx5_glue = *sym;
3613 	return 0;
3614 glue_error:
3615 	if (handle)
3616 		dlclose(handle);
3617 	DRV_LOG(WARNING,
3618 		"cannot initialize PMD due to missing run-time dependency on"
3619 		" rdma-core libraries (libibverbs, libmlx5)");
3620 	return -rte_errno;
3621 }
3622 
3623 #endif
3624 
3625 /**
3626  * Driver initialization routine.
3627  */
3628 RTE_INIT(rte_mlx5_pmd_init)
3629 {
3630 	/* Initialize driver log type. */
3631 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3632 	if (mlx5_logtype >= 0)
3633 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3634 
3635 	/* Build the static tables for Verbs conversion. */
3636 	mlx5_set_ptype_table();
3637 	mlx5_set_cksum_table();
3638 	mlx5_set_swp_types_table();
3639 	/*
3640 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3641 	 * huge pages. Calling ibv_fork_init() during init allows
3642 	 * applications to use fork() safely for purposes other than
3643 	 * using this PMD, which is not supported in forked processes.
3644 	 */
3645 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3646 	/* Match the size of Rx completion entry to the size of a cacheline. */
3647 	if (RTE_CACHE_LINE_SIZE == 128)
3648 		setenv("MLX5_CQE_SIZE", "128", 0);
3649 	/*
3650 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
3651 	 * cleanup all the Verbs resources even when the device was removed.
3652 	 */
3653 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
3654 #ifdef RTE_IBVERBS_LINK_DLOPEN
3655 	if (mlx5_glue_init())
3656 		return;
3657 	assert(mlx5_glue);
3658 #endif
3659 #ifndef NDEBUG
3660 	/* Glue structure must not contain any NULL pointers. */
3661 	{
3662 		unsigned int i;
3663 
3664 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
3665 			assert(((const void *const *)mlx5_glue)[i]);
3666 	}
3667 #endif
3668 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
3669 		DRV_LOG(ERR,
3670 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
3671 			mlx5_glue->version, MLX5_GLUE_VERSION);
3672 		return;
3673 	}
3674 	mlx5_glue->fork_init();
3675 	rte_pci_register(&mlx5_driver);
3676 }
3677 
3678 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3679 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3680 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3681