xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 1f41d98c207aee8982ced709864c96c463d4503a)
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 	int err = mlx5_alloc_table_hash_list(priv);
870 
871 	if (err)
872 		return err;
873 #ifdef HAVE_MLX5DV_DR
874 	struct mlx5_ibv_shared *sh = priv->sh;
875 	char s[MLX5_HLIST_NAMESIZE];
876 	void *domain;
877 
878 	if (sh->dv_refcnt) {
879 		/* Shared DV/DR structures is already initialized. */
880 		sh->dv_refcnt++;
881 		priv->dr_shared = 1;
882 		return 0;
883 	}
884 	/* Reference counter is zero, we should initialize structures. */
885 	domain = mlx5_glue->dr_create_domain(sh->ctx,
886 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
887 	if (!domain) {
888 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
889 		err = errno;
890 		goto error;
891 	}
892 	sh->rx_domain = domain;
893 	domain = mlx5_glue->dr_create_domain(sh->ctx,
894 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
895 	if (!domain) {
896 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
897 		err = errno;
898 		goto error;
899 	}
900 	pthread_mutex_init(&sh->dv_mutex, NULL);
901 	sh->tx_domain = domain;
902 #ifdef HAVE_MLX5DV_DR_ESWITCH
903 	if (priv->config.dv_esw_en) {
904 		domain  = mlx5_glue->dr_create_domain
905 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
906 		if (!domain) {
907 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
908 			err = errno;
909 			goto error;
910 		}
911 		sh->fdb_domain = domain;
912 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
913 	}
914 #endif
915 	/* create tags hash list table. */
916 	snprintf(s, sizeof(s), "%s_tags", priv->sh->ibdev_name);
917 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
918 	if (!sh->flow_tbls) {
919 		DRV_LOG(ERR, "tags with hash creation failed.\n");
920 		goto error;
921 	}
922 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
923 	sh->dv_refcnt++;
924 	priv->dr_shared = 1;
925 	return 0;
926 
927 error:
928        /* Rollback the created objects. */
929 	if (sh->rx_domain) {
930 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
931 		sh->rx_domain = NULL;
932 	}
933 	if (sh->tx_domain) {
934 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
935 		sh->tx_domain = NULL;
936 	}
937 	if (sh->fdb_domain) {
938 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
939 		sh->fdb_domain = NULL;
940 	}
941 	if (sh->esw_drop_action) {
942 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
943 		sh->esw_drop_action = NULL;
944 	}
945 	if (sh->pop_vlan_action) {
946 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
947 		sh->pop_vlan_action = NULL;
948 	}
949 	mlx5_free_table_hash_list(priv);
950 #endif
951 	return err;
952 }
953 
954 /**
955  * Destroy DR related data within private structure.
956  *
957  * @param[in] priv
958  *   Pointer to the private device data structure.
959  */
960 static void
961 mlx5_free_shared_dr(struct mlx5_priv *priv)
962 {
963 #ifdef HAVE_MLX5DV_DR
964 	struct mlx5_ibv_shared *sh;
965 
966 	if (!priv->dr_shared)
967 		return;
968 	priv->dr_shared = 0;
969 	sh = priv->sh;
970 	assert(sh);
971 	assert(sh->dv_refcnt);
972 	if (sh->dv_refcnt && --sh->dv_refcnt)
973 		return;
974 	if (sh->rx_domain) {
975 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
976 		sh->rx_domain = NULL;
977 	}
978 	if (sh->tx_domain) {
979 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
980 		sh->tx_domain = NULL;
981 	}
982 #ifdef HAVE_MLX5DV_DR_ESWITCH
983 	if (sh->fdb_domain) {
984 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
985 		sh->fdb_domain = NULL;
986 	}
987 	if (sh->esw_drop_action) {
988 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
989 		sh->esw_drop_action = NULL;
990 	}
991 #endif
992 	if (sh->tag_table) {
993 		/* tags should be destroyed with flow before. */
994 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
995 		sh->tag_table = NULL;
996 	}
997 	if (sh->pop_vlan_action) {
998 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
999 		sh->pop_vlan_action = NULL;
1000 	}
1001 	pthread_mutex_destroy(&sh->dv_mutex);
1002 #endif /* HAVE_MLX5DV_DR */
1003 	mlx5_free_table_hash_list(priv);
1004 }
1005 
1006 /**
1007  * Initialize shared data between primary and secondary process.
1008  *
1009  * A memzone is reserved by primary process and secondary processes attach to
1010  * the memzone.
1011  *
1012  * @return
1013  *   0 on success, a negative errno value otherwise and rte_errno is set.
1014  */
1015 static int
1016 mlx5_init_shared_data(void)
1017 {
1018 	const struct rte_memzone *mz;
1019 	int ret = 0;
1020 
1021 	rte_spinlock_lock(&mlx5_shared_data_lock);
1022 	if (mlx5_shared_data == NULL) {
1023 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1024 			/* Allocate shared memory. */
1025 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1026 						 sizeof(*mlx5_shared_data),
1027 						 SOCKET_ID_ANY, 0);
1028 			if (mz == NULL) {
1029 				DRV_LOG(ERR,
1030 					"Cannot allocate mlx5 shared data");
1031 				ret = -rte_errno;
1032 				goto error;
1033 			}
1034 			mlx5_shared_data = mz->addr;
1035 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1036 			rte_spinlock_init(&mlx5_shared_data->lock);
1037 		} else {
1038 			/* Lookup allocated shared memory. */
1039 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1040 			if (mz == NULL) {
1041 				DRV_LOG(ERR,
1042 					"Cannot attach mlx5 shared data");
1043 				ret = -rte_errno;
1044 				goto error;
1045 			}
1046 			mlx5_shared_data = mz->addr;
1047 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1048 		}
1049 	}
1050 error:
1051 	rte_spinlock_unlock(&mlx5_shared_data_lock);
1052 	return ret;
1053 }
1054 
1055 /**
1056  * Retrieve integer value from environment variable.
1057  *
1058  * @param[in] name
1059  *   Environment variable name.
1060  *
1061  * @return
1062  *   Integer value, 0 if the variable is not set.
1063  */
1064 int
1065 mlx5_getenv_int(const char *name)
1066 {
1067 	const char *val = getenv(name);
1068 
1069 	if (val == NULL)
1070 		return 0;
1071 	return atoi(val);
1072 }
1073 
1074 /**
1075  * Verbs callback to allocate a memory. This function should allocate the space
1076  * according to the size provided residing inside a huge page.
1077  * Please note that all allocation must respect the alignment from libmlx5
1078  * (i.e. currently sysconf(_SC_PAGESIZE)).
1079  *
1080  * @param[in] size
1081  *   The size in bytes of the memory to allocate.
1082  * @param[in] data
1083  *   A pointer to the callback data.
1084  *
1085  * @return
1086  *   Allocated buffer, NULL otherwise and rte_errno is set.
1087  */
1088 static void *
1089 mlx5_alloc_verbs_buf(size_t size, void *data)
1090 {
1091 	struct mlx5_priv *priv = data;
1092 	void *ret;
1093 	size_t alignment = sysconf(_SC_PAGESIZE);
1094 	unsigned int socket = SOCKET_ID_ANY;
1095 
1096 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1097 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1098 
1099 		socket = ctrl->socket;
1100 	} else if (priv->verbs_alloc_ctx.type ==
1101 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1102 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1103 
1104 		socket = ctrl->socket;
1105 	}
1106 	assert(data != NULL);
1107 	ret = rte_malloc_socket(__func__, size, alignment, socket);
1108 	if (!ret && size)
1109 		rte_errno = ENOMEM;
1110 	return ret;
1111 }
1112 
1113 /**
1114  * Verbs callback to free a memory.
1115  *
1116  * @param[in] ptr
1117  *   A pointer to the memory to free.
1118  * @param[in] data
1119  *   A pointer to the callback data.
1120  */
1121 static void
1122 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1123 {
1124 	assert(data != NULL);
1125 	rte_free(ptr);
1126 }
1127 
1128 /**
1129  * DPDK callback to add udp tunnel port
1130  *
1131  * @param[in] dev
1132  *   A pointer to eth_dev
1133  * @param[in] udp_tunnel
1134  *   A pointer to udp tunnel
1135  *
1136  * @return
1137  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1138  */
1139 int
1140 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1141 			 struct rte_eth_udp_tunnel *udp_tunnel)
1142 {
1143 	assert(udp_tunnel != NULL);
1144 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1145 	    udp_tunnel->udp_port == 4789)
1146 		return 0;
1147 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1148 	    udp_tunnel->udp_port == 4790)
1149 		return 0;
1150 	return -ENOTSUP;
1151 }
1152 
1153 /**
1154  * Initialize process private data structure.
1155  *
1156  * @param dev
1157  *   Pointer to Ethernet device structure.
1158  *
1159  * @return
1160  *   0 on success, a negative errno value otherwise and rte_errno is set.
1161  */
1162 int
1163 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1164 {
1165 	struct mlx5_priv *priv = dev->data->dev_private;
1166 	struct mlx5_proc_priv *ppriv;
1167 	size_t ppriv_size;
1168 
1169 	/*
1170 	 * UAR register table follows the process private structure. BlueFlame
1171 	 * registers for Tx queues are stored in the table.
1172 	 */
1173 	ppriv_size =
1174 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1175 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1176 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1177 	if (!ppriv) {
1178 		rte_errno = ENOMEM;
1179 		return -rte_errno;
1180 	}
1181 	ppriv->uar_table_sz = ppriv_size;
1182 	dev->process_private = ppriv;
1183 	return 0;
1184 }
1185 
1186 /**
1187  * Un-initialize process private data structure.
1188  *
1189  * @param dev
1190  *   Pointer to Ethernet device structure.
1191  */
1192 static void
1193 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1194 {
1195 	if (!dev->process_private)
1196 		return;
1197 	rte_free(dev->process_private);
1198 	dev->process_private = NULL;
1199 }
1200 
1201 /**
1202  * DPDK callback to close the device.
1203  *
1204  * Destroy all queues and objects, free memory.
1205  *
1206  * @param dev
1207  *   Pointer to Ethernet device structure.
1208  */
1209 static void
1210 mlx5_dev_close(struct rte_eth_dev *dev)
1211 {
1212 	struct mlx5_priv *priv = dev->data->dev_private;
1213 	unsigned int i;
1214 	int ret;
1215 
1216 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1217 		dev->data->port_id,
1218 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1219 	/* In case mlx5_dev_stop() has not been called. */
1220 	mlx5_dev_interrupt_handler_uninstall(dev);
1221 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1222 	mlx5_traffic_disable(dev);
1223 	mlx5_flow_flush(dev, NULL);
1224 	mlx5_flow_meter_flush(dev, NULL);
1225 	/* Prevent crashes when queues are still in use. */
1226 	dev->rx_pkt_burst = removed_rx_burst;
1227 	dev->tx_pkt_burst = removed_tx_burst;
1228 	rte_wmb();
1229 	/* Disable datapath on secondary process. */
1230 	mlx5_mp_req_stop_rxtx(dev);
1231 	if (priv->rxqs != NULL) {
1232 		/* XXX race condition if mlx5_rx_burst() is still running. */
1233 		usleep(1000);
1234 		for (i = 0; (i != priv->rxqs_n); ++i)
1235 			mlx5_rxq_release(dev, i);
1236 		priv->rxqs_n = 0;
1237 		priv->rxqs = NULL;
1238 	}
1239 	if (priv->txqs != NULL) {
1240 		/* XXX race condition if mlx5_tx_burst() is still running. */
1241 		usleep(1000);
1242 		for (i = 0; (i != priv->txqs_n); ++i)
1243 			mlx5_txq_release(dev, i);
1244 		priv->txqs_n = 0;
1245 		priv->txqs = NULL;
1246 	}
1247 	mlx5_proc_priv_uninit(dev);
1248 	if (priv->mreg_cp_tbl)
1249 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1250 	mlx5_mprq_free_mp(dev);
1251 	mlx5_free_shared_dr(priv);
1252 	if (priv->rss_conf.rss_key != NULL)
1253 		rte_free(priv->rss_conf.rss_key);
1254 	if (priv->reta_idx != NULL)
1255 		rte_free(priv->reta_idx);
1256 	if (priv->config.vf)
1257 		mlx5_nl_mac_addr_flush(dev);
1258 	if (priv->nl_socket_route >= 0)
1259 		close(priv->nl_socket_route);
1260 	if (priv->nl_socket_rdma >= 0)
1261 		close(priv->nl_socket_rdma);
1262 	if (priv->vmwa_context)
1263 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1264 	if (priv->sh) {
1265 		/*
1266 		 * Free the shared context in last turn, because the cleanup
1267 		 * routines above may use some shared fields, like
1268 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1269 		 * ifindex if Netlink fails.
1270 		 */
1271 		mlx5_free_shared_ibctx(priv->sh);
1272 		priv->sh = NULL;
1273 	}
1274 	ret = mlx5_hrxq_verify(dev);
1275 	if (ret)
1276 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1277 			dev->data->port_id);
1278 	ret = mlx5_ind_table_obj_verify(dev);
1279 	if (ret)
1280 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1281 			dev->data->port_id);
1282 	ret = mlx5_rxq_obj_verify(dev);
1283 	if (ret)
1284 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1285 			dev->data->port_id);
1286 	ret = mlx5_rxq_verify(dev);
1287 	if (ret)
1288 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1289 			dev->data->port_id);
1290 	ret = mlx5_txq_obj_verify(dev);
1291 	if (ret)
1292 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1293 			dev->data->port_id);
1294 	ret = mlx5_txq_verify(dev);
1295 	if (ret)
1296 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1297 			dev->data->port_id);
1298 	ret = mlx5_flow_verify(dev);
1299 	if (ret)
1300 		DRV_LOG(WARNING, "port %u some flows still remain",
1301 			dev->data->port_id);
1302 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1303 		unsigned int c = 0;
1304 		uint16_t port_id;
1305 
1306 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1307 			struct mlx5_priv *opriv =
1308 				rte_eth_devices[port_id].data->dev_private;
1309 
1310 			if (!opriv ||
1311 			    opriv->domain_id != priv->domain_id ||
1312 			    &rte_eth_devices[port_id] == dev)
1313 				continue;
1314 			++c;
1315 			break;
1316 		}
1317 		if (!c)
1318 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1319 	}
1320 	memset(priv, 0, sizeof(*priv));
1321 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1322 	/*
1323 	 * Reset mac_addrs to NULL such that it is not freed as part of
1324 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1325 	 * it is freed when dev_private is freed.
1326 	 */
1327 	dev->data->mac_addrs = NULL;
1328 }
1329 
1330 const struct eth_dev_ops mlx5_dev_ops = {
1331 	.dev_configure = mlx5_dev_configure,
1332 	.dev_start = mlx5_dev_start,
1333 	.dev_stop = mlx5_dev_stop,
1334 	.dev_set_link_down = mlx5_set_link_down,
1335 	.dev_set_link_up = mlx5_set_link_up,
1336 	.dev_close = mlx5_dev_close,
1337 	.promiscuous_enable = mlx5_promiscuous_enable,
1338 	.promiscuous_disable = mlx5_promiscuous_disable,
1339 	.allmulticast_enable = mlx5_allmulticast_enable,
1340 	.allmulticast_disable = mlx5_allmulticast_disable,
1341 	.link_update = mlx5_link_update,
1342 	.stats_get = mlx5_stats_get,
1343 	.stats_reset = mlx5_stats_reset,
1344 	.xstats_get = mlx5_xstats_get,
1345 	.xstats_reset = mlx5_xstats_reset,
1346 	.xstats_get_names = mlx5_xstats_get_names,
1347 	.fw_version_get = mlx5_fw_version_get,
1348 	.dev_infos_get = mlx5_dev_infos_get,
1349 	.read_clock = mlx5_read_clock,
1350 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1351 	.vlan_filter_set = mlx5_vlan_filter_set,
1352 	.rx_queue_setup = mlx5_rx_queue_setup,
1353 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1354 	.tx_queue_setup = mlx5_tx_queue_setup,
1355 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1356 	.rx_queue_release = mlx5_rx_queue_release,
1357 	.tx_queue_release = mlx5_tx_queue_release,
1358 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1359 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1360 	.mac_addr_remove = mlx5_mac_addr_remove,
1361 	.mac_addr_add = mlx5_mac_addr_add,
1362 	.mac_addr_set = mlx5_mac_addr_set,
1363 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1364 	.mtu_set = mlx5_dev_set_mtu,
1365 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1366 	.vlan_offload_set = mlx5_vlan_offload_set,
1367 	.reta_update = mlx5_dev_rss_reta_update,
1368 	.reta_query = mlx5_dev_rss_reta_query,
1369 	.rss_hash_update = mlx5_rss_hash_update,
1370 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1371 	.filter_ctrl = mlx5_dev_filter_ctrl,
1372 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1373 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1374 	.rx_queue_count = mlx5_rx_queue_count,
1375 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1376 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1377 	.is_removed = mlx5_is_removed,
1378 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1379 	.get_module_info = mlx5_get_module_info,
1380 	.get_module_eeprom = mlx5_get_module_eeprom,
1381 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1382 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1383 };
1384 
1385 /* Available operations from secondary process. */
1386 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1387 	.stats_get = mlx5_stats_get,
1388 	.stats_reset = mlx5_stats_reset,
1389 	.xstats_get = mlx5_xstats_get,
1390 	.xstats_reset = mlx5_xstats_reset,
1391 	.xstats_get_names = mlx5_xstats_get_names,
1392 	.fw_version_get = mlx5_fw_version_get,
1393 	.dev_infos_get = mlx5_dev_infos_get,
1394 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1395 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1396 	.get_module_info = mlx5_get_module_info,
1397 	.get_module_eeprom = mlx5_get_module_eeprom,
1398 };
1399 
1400 /* Available operations in flow isolated mode. */
1401 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1402 	.dev_configure = mlx5_dev_configure,
1403 	.dev_start = mlx5_dev_start,
1404 	.dev_stop = mlx5_dev_stop,
1405 	.dev_set_link_down = mlx5_set_link_down,
1406 	.dev_set_link_up = mlx5_set_link_up,
1407 	.dev_close = mlx5_dev_close,
1408 	.promiscuous_enable = mlx5_promiscuous_enable,
1409 	.promiscuous_disable = mlx5_promiscuous_disable,
1410 	.allmulticast_enable = mlx5_allmulticast_enable,
1411 	.allmulticast_disable = mlx5_allmulticast_disable,
1412 	.link_update = mlx5_link_update,
1413 	.stats_get = mlx5_stats_get,
1414 	.stats_reset = mlx5_stats_reset,
1415 	.xstats_get = mlx5_xstats_get,
1416 	.xstats_reset = mlx5_xstats_reset,
1417 	.xstats_get_names = mlx5_xstats_get_names,
1418 	.fw_version_get = mlx5_fw_version_get,
1419 	.dev_infos_get = mlx5_dev_infos_get,
1420 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1421 	.vlan_filter_set = mlx5_vlan_filter_set,
1422 	.rx_queue_setup = mlx5_rx_queue_setup,
1423 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1424 	.tx_queue_setup = mlx5_tx_queue_setup,
1425 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1426 	.rx_queue_release = mlx5_rx_queue_release,
1427 	.tx_queue_release = mlx5_tx_queue_release,
1428 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1429 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1430 	.mac_addr_remove = mlx5_mac_addr_remove,
1431 	.mac_addr_add = mlx5_mac_addr_add,
1432 	.mac_addr_set = mlx5_mac_addr_set,
1433 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1434 	.mtu_set = mlx5_dev_set_mtu,
1435 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1436 	.vlan_offload_set = mlx5_vlan_offload_set,
1437 	.filter_ctrl = mlx5_dev_filter_ctrl,
1438 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1439 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1440 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1441 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1442 	.is_removed = mlx5_is_removed,
1443 	.get_module_info = mlx5_get_module_info,
1444 	.get_module_eeprom = mlx5_get_module_eeprom,
1445 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1446 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1447 };
1448 
1449 /**
1450  * Verify and store value for device argument.
1451  *
1452  * @param[in] key
1453  *   Key argument to verify.
1454  * @param[in] val
1455  *   Value associated with key.
1456  * @param opaque
1457  *   User data.
1458  *
1459  * @return
1460  *   0 on success, a negative errno value otherwise and rte_errno is set.
1461  */
1462 static int
1463 mlx5_args_check(const char *key, const char *val, void *opaque)
1464 {
1465 	struct mlx5_dev_config *config = opaque;
1466 	unsigned long tmp;
1467 
1468 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1469 	if (!strcmp(MLX5_REPRESENTOR, key))
1470 		return 0;
1471 	errno = 0;
1472 	tmp = strtoul(val, NULL, 0);
1473 	if (errno) {
1474 		rte_errno = errno;
1475 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1476 		return -rte_errno;
1477 	}
1478 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1479 		config->cqe_comp = !!tmp;
1480 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1481 		config->cqe_pad = !!tmp;
1482 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1483 		config->hw_padding = !!tmp;
1484 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1485 		config->mprq.enabled = !!tmp;
1486 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1487 		config->mprq.stride_num_n = tmp;
1488 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1489 		config->mprq.max_memcpy_len = tmp;
1490 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1491 		config->mprq.min_rxqs_num = tmp;
1492 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1493 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1494 				 " converted to txq_inline_max", key);
1495 		config->txq_inline_max = tmp;
1496 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1497 		config->txq_inline_max = tmp;
1498 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1499 		config->txq_inline_min = tmp;
1500 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1501 		config->txq_inline_mpw = tmp;
1502 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1503 		config->txqs_inline = tmp;
1504 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1505 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1506 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1507 		config->mps = !!tmp;
1508 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1509 		if (tmp != MLX5_TXDB_CACHED &&
1510 		    tmp != MLX5_TXDB_NCACHED &&
1511 		    tmp != MLX5_TXDB_HEURISTIC) {
1512 			DRV_LOG(ERR, "invalid Tx doorbell "
1513 				     "mapping parameter");
1514 			rte_errno = EINVAL;
1515 			return -rte_errno;
1516 		}
1517 		config->dbnc = tmp;
1518 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1519 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1520 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1521 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1522 				 " converted to txq_inline_mpw", key);
1523 		config->txq_inline_mpw = tmp;
1524 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1525 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1526 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1527 		config->rx_vec_en = !!tmp;
1528 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1529 		config->l3_vxlan_en = !!tmp;
1530 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1531 		config->vf_nl_en = !!tmp;
1532 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1533 		config->dv_esw_en = !!tmp;
1534 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1535 		config->dv_flow_en = !!tmp;
1536 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1537 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1538 		    tmp != MLX5_XMETA_MODE_META16 &&
1539 		    tmp != MLX5_XMETA_MODE_META32) {
1540 			DRV_LOG(ERR, "invalid extensive "
1541 				     "metadata parameter");
1542 			rte_errno = EINVAL;
1543 			return -rte_errno;
1544 		}
1545 		config->dv_xmeta_en = tmp;
1546 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1547 		config->mr_ext_memseg_en = !!tmp;
1548 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1549 		config->max_dump_files_num = tmp;
1550 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1551 		config->lro.timeout = tmp;
1552 	} else {
1553 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1554 		rte_errno = EINVAL;
1555 		return -rte_errno;
1556 	}
1557 	return 0;
1558 }
1559 
1560 /**
1561  * Parse device parameters.
1562  *
1563  * @param config
1564  *   Pointer to device configuration structure.
1565  * @param devargs
1566  *   Device arguments structure.
1567  *
1568  * @return
1569  *   0 on success, a negative errno value otherwise and rte_errno is set.
1570  */
1571 static int
1572 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1573 {
1574 	const char **params = (const char *[]){
1575 		MLX5_RXQ_CQE_COMP_EN,
1576 		MLX5_RXQ_CQE_PAD_EN,
1577 		MLX5_RXQ_PKT_PAD_EN,
1578 		MLX5_RX_MPRQ_EN,
1579 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1580 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1581 		MLX5_RXQS_MIN_MPRQ,
1582 		MLX5_TXQ_INLINE,
1583 		MLX5_TXQ_INLINE_MIN,
1584 		MLX5_TXQ_INLINE_MAX,
1585 		MLX5_TXQ_INLINE_MPW,
1586 		MLX5_TXQS_MIN_INLINE,
1587 		MLX5_TXQS_MAX_VEC,
1588 		MLX5_TXQ_MPW_EN,
1589 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1590 		MLX5_TXQ_MAX_INLINE_LEN,
1591 		MLX5_TX_DB_NC,
1592 		MLX5_TX_VEC_EN,
1593 		MLX5_RX_VEC_EN,
1594 		MLX5_L3_VXLAN_EN,
1595 		MLX5_VF_NL_EN,
1596 		MLX5_DV_ESW_EN,
1597 		MLX5_DV_FLOW_EN,
1598 		MLX5_DV_XMETA_EN,
1599 		MLX5_MR_EXT_MEMSEG_EN,
1600 		MLX5_REPRESENTOR,
1601 		MLX5_MAX_DUMP_FILES_NUM,
1602 		MLX5_LRO_TIMEOUT_USEC,
1603 		NULL,
1604 	};
1605 	struct rte_kvargs *kvlist;
1606 	int ret = 0;
1607 	int i;
1608 
1609 	if (devargs == NULL)
1610 		return 0;
1611 	/* Following UGLY cast is done to pass checkpatch. */
1612 	kvlist = rte_kvargs_parse(devargs->args, params);
1613 	if (kvlist == NULL) {
1614 		rte_errno = EINVAL;
1615 		return -rte_errno;
1616 	}
1617 	/* Process parameters. */
1618 	for (i = 0; (params[i] != NULL); ++i) {
1619 		if (rte_kvargs_count(kvlist, params[i])) {
1620 			ret = rte_kvargs_process(kvlist, params[i],
1621 						 mlx5_args_check, config);
1622 			if (ret) {
1623 				rte_errno = EINVAL;
1624 				rte_kvargs_free(kvlist);
1625 				return -rte_errno;
1626 			}
1627 		}
1628 	}
1629 	rte_kvargs_free(kvlist);
1630 	return 0;
1631 }
1632 
1633 static struct rte_pci_driver mlx5_driver;
1634 
1635 /**
1636  * PMD global initialization.
1637  *
1638  * Independent from individual device, this function initializes global
1639  * per-PMD data structures distinguishing primary and secondary processes.
1640  * Hence, each initialization is called once per a process.
1641  *
1642  * @return
1643  *   0 on success, a negative errno value otherwise and rte_errno is set.
1644  */
1645 static int
1646 mlx5_init_once(void)
1647 {
1648 	struct mlx5_shared_data *sd;
1649 	struct mlx5_local_data *ld = &mlx5_local_data;
1650 	int ret = 0;
1651 
1652 	if (mlx5_init_shared_data())
1653 		return -rte_errno;
1654 	sd = mlx5_shared_data;
1655 	assert(sd);
1656 	rte_spinlock_lock(&sd->lock);
1657 	switch (rte_eal_process_type()) {
1658 	case RTE_PROC_PRIMARY:
1659 		if (sd->init_done)
1660 			break;
1661 		LIST_INIT(&sd->mem_event_cb_list);
1662 		rte_rwlock_init(&sd->mem_event_rwlock);
1663 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1664 						mlx5_mr_mem_event_cb, NULL);
1665 		ret = mlx5_mp_init_primary();
1666 		if (ret)
1667 			goto out;
1668 		sd->init_done = true;
1669 		break;
1670 	case RTE_PROC_SECONDARY:
1671 		if (ld->init_done)
1672 			break;
1673 		ret = mlx5_mp_init_secondary();
1674 		if (ret)
1675 			goto out;
1676 		++sd->secondary_cnt;
1677 		ld->init_done = true;
1678 		break;
1679 	default:
1680 		break;
1681 	}
1682 out:
1683 	rte_spinlock_unlock(&sd->lock);
1684 	return ret;
1685 }
1686 
1687 /**
1688  * Configures the minimal amount of data to inline into WQE
1689  * while sending packets.
1690  *
1691  * - the txq_inline_min has the maximal priority, if this
1692  *   key is specified in devargs
1693  * - if DevX is enabled the inline mode is queried from the
1694  *   device (HCA attributes and NIC vport context if needed).
1695  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1696  *   and none (0 bytes) for other NICs
1697  *
1698  * @param spawn
1699  *   Verbs device parameters (name, port, switch_info) to spawn.
1700  * @param config
1701  *   Device configuration parameters.
1702  */
1703 static void
1704 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1705 		    struct mlx5_dev_config *config)
1706 {
1707 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1708 		/* Application defines size of inlined data explicitly. */
1709 		switch (spawn->pci_dev->id.device_id) {
1710 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1711 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1712 			if (config->txq_inline_min <
1713 				       (int)MLX5_INLINE_HSIZE_L2) {
1714 				DRV_LOG(DEBUG,
1715 					"txq_inline_mix aligned to minimal"
1716 					" ConnectX-4 required value %d",
1717 					(int)MLX5_INLINE_HSIZE_L2);
1718 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1719 			}
1720 			break;
1721 		}
1722 		goto exit;
1723 	}
1724 	if (config->hca_attr.eth_net_offloads) {
1725 		/* We have DevX enabled, inline mode queried successfully. */
1726 		switch (config->hca_attr.wqe_inline_mode) {
1727 		case MLX5_CAP_INLINE_MODE_L2:
1728 			/* outer L2 header must be inlined. */
1729 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1730 			goto exit;
1731 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1732 			/* No inline data are required by NIC. */
1733 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1734 			config->hw_vlan_insert =
1735 				config->hca_attr.wqe_vlan_insert;
1736 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1737 			goto exit;
1738 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1739 			/* inline mode is defined by NIC vport context. */
1740 			if (!config->hca_attr.eth_virt)
1741 				break;
1742 			switch (config->hca_attr.vport_inline_mode) {
1743 			case MLX5_INLINE_MODE_NONE:
1744 				config->txq_inline_min =
1745 					MLX5_INLINE_HSIZE_NONE;
1746 				goto exit;
1747 			case MLX5_INLINE_MODE_L2:
1748 				config->txq_inline_min =
1749 					MLX5_INLINE_HSIZE_L2;
1750 				goto exit;
1751 			case MLX5_INLINE_MODE_IP:
1752 				config->txq_inline_min =
1753 					MLX5_INLINE_HSIZE_L3;
1754 				goto exit;
1755 			case MLX5_INLINE_MODE_TCP_UDP:
1756 				config->txq_inline_min =
1757 					MLX5_INLINE_HSIZE_L4;
1758 				goto exit;
1759 			case MLX5_INLINE_MODE_INNER_L2:
1760 				config->txq_inline_min =
1761 					MLX5_INLINE_HSIZE_INNER_L2;
1762 				goto exit;
1763 			case MLX5_INLINE_MODE_INNER_IP:
1764 				config->txq_inline_min =
1765 					MLX5_INLINE_HSIZE_INNER_L3;
1766 				goto exit;
1767 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1768 				config->txq_inline_min =
1769 					MLX5_INLINE_HSIZE_INNER_L4;
1770 				goto exit;
1771 			}
1772 		}
1773 	}
1774 	/*
1775 	 * We get here if we are unable to deduce
1776 	 * inline data size with DevX. Try PCI ID
1777 	 * to determine old NICs.
1778 	 */
1779 	switch (spawn->pci_dev->id.device_id) {
1780 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1781 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1782 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1783 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1784 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1785 		config->hw_vlan_insert = 0;
1786 		break;
1787 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1788 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1789 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1790 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1791 		/*
1792 		 * These NICs support VLAN insertion from WQE and
1793 		 * report the wqe_vlan_insert flag. But there is the bug
1794 		 * and PFC control may be broken, so disable feature.
1795 		 */
1796 		config->hw_vlan_insert = 0;
1797 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1798 		break;
1799 	default:
1800 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1801 		break;
1802 	}
1803 exit:
1804 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1805 }
1806 
1807 /**
1808  * Configures the metadata mask fields in the shared context.
1809  *
1810  * @param [in] dev
1811  *   Pointer to Ethernet device.
1812  */
1813 static void
1814 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1815 {
1816 	struct mlx5_priv *priv = dev->data->dev_private;
1817 	struct mlx5_ibv_shared *sh = priv->sh;
1818 	uint32_t meta, mark, reg_c0;
1819 
1820 	reg_c0 = ~priv->vport_meta_mask;
1821 	switch (priv->config.dv_xmeta_en) {
1822 	case MLX5_XMETA_MODE_LEGACY:
1823 		meta = UINT32_MAX;
1824 		mark = MLX5_FLOW_MARK_MASK;
1825 		break;
1826 	case MLX5_XMETA_MODE_META16:
1827 		meta = reg_c0 >> rte_bsf32(reg_c0);
1828 		mark = MLX5_FLOW_MARK_MASK;
1829 		break;
1830 	case MLX5_XMETA_MODE_META32:
1831 		meta = UINT32_MAX;
1832 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1833 		break;
1834 	default:
1835 		meta = 0;
1836 		mark = 0;
1837 		assert(false);
1838 		break;
1839 	}
1840 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1841 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1842 				 sh->dv_mark_mask, mark);
1843 	else
1844 		sh->dv_mark_mask = mark;
1845 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1846 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1847 				 sh->dv_meta_mask, meta);
1848 	else
1849 		sh->dv_meta_mask = meta;
1850 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1851 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1852 				 sh->dv_meta_mask, reg_c0);
1853 	else
1854 		sh->dv_regc0_mask = reg_c0;
1855 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1856 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1857 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1858 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1859 }
1860 
1861 /**
1862  * Allocate page of door-bells and register it using DevX API.
1863  *
1864  * @param [in] dev
1865  *   Pointer to Ethernet device.
1866  *
1867  * @return
1868  *   Pointer to new page on success, NULL otherwise.
1869  */
1870 static struct mlx5_devx_dbr_page *
1871 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1872 {
1873 	struct mlx5_priv *priv = dev->data->dev_private;
1874 	struct mlx5_devx_dbr_page *page;
1875 
1876 	/* Allocate space for door-bell page and management data. */
1877 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1878 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1879 	if (!page) {
1880 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1881 			dev->data->port_id);
1882 		return NULL;
1883 	}
1884 	/* Register allocated memory. */
1885 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1886 					      MLX5_DBR_PAGE_SIZE, 0);
1887 	if (!page->umem) {
1888 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1889 			dev->data->port_id);
1890 		rte_free(page);
1891 		return NULL;
1892 	}
1893 	return page;
1894 }
1895 
1896 /**
1897  * Find the next available door-bell, allocate new page if needed.
1898  *
1899  * @param [in] dev
1900  *   Pointer to Ethernet device.
1901  * @param [out] dbr_page
1902  *   Door-bell page containing the page data.
1903  *
1904  * @return
1905  *   Door-bell address offset on success, a negative error value otherwise.
1906  */
1907 int64_t
1908 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1909 {
1910 	struct mlx5_priv *priv = dev->data->dev_private;
1911 	struct mlx5_devx_dbr_page *page = NULL;
1912 	uint32_t i, j;
1913 
1914 	LIST_FOREACH(page, &priv->dbrpgs, next)
1915 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1916 			break;
1917 	if (!page) { /* No page with free door-bell exists. */
1918 		page = mlx5_alloc_dbr_page(dev);
1919 		if (!page) /* Failed to allocate new page. */
1920 			return (-1);
1921 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1922 	}
1923 	/* Loop to find bitmap part with clear bit. */
1924 	for (i = 0;
1925 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1926 	     i++)
1927 		; /* Empty. */
1928 	/* Find the first clear bit. */
1929 	j = rte_bsf64(~page->dbr_bitmap[i]);
1930 	assert(i < (MLX5_DBR_PER_PAGE / 64));
1931 	page->dbr_bitmap[i] |= (1 << j);
1932 	page->dbr_count++;
1933 	*dbr_page = page;
1934 	return (((i * 64) + j) * sizeof(uint64_t));
1935 }
1936 
1937 /**
1938  * Release a door-bell record.
1939  *
1940  * @param [in] dev
1941  *   Pointer to Ethernet device.
1942  * @param [in] umem_id
1943  *   UMEM ID of page containing the door-bell record to release.
1944  * @param [in] offset
1945  *   Offset of door-bell record in page.
1946  *
1947  * @return
1948  *   0 on success, a negative error value otherwise.
1949  */
1950 int32_t
1951 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1952 {
1953 	struct mlx5_priv *priv = dev->data->dev_private;
1954 	struct mlx5_devx_dbr_page *page = NULL;
1955 	int ret = 0;
1956 
1957 	LIST_FOREACH(page, &priv->dbrpgs, next)
1958 		/* Find the page this address belongs to. */
1959 		if (page->umem->umem_id == umem_id)
1960 			break;
1961 	if (!page)
1962 		return -EINVAL;
1963 	page->dbr_count--;
1964 	if (!page->dbr_count) {
1965 		/* Page not used, free it and remove from list. */
1966 		LIST_REMOVE(page, next);
1967 		if (page->umem)
1968 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
1969 		rte_free(page);
1970 	} else {
1971 		/* Mark in bitmap that this door-bell is not in use. */
1972 		offset /= MLX5_DBR_SIZE;
1973 		int i = offset / 64;
1974 		int j = offset % 64;
1975 
1976 		page->dbr_bitmap[i] &= ~(1 << j);
1977 	}
1978 	return ret;
1979 }
1980 
1981 /**
1982  * Check sibling device configurations.
1983  *
1984  * Sibling devices sharing the Infiniband device context
1985  * should have compatible configurations. This regards
1986  * representors and bonding slaves.
1987  *
1988  * @param priv
1989  *   Private device descriptor.
1990  * @param config
1991  *   Configuration of the device is going to be created.
1992  *
1993  * @return
1994  *   0 on success, EINVAL otherwise
1995  */
1996 static int
1997 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1998 			      struct mlx5_dev_config *config)
1999 {
2000 	struct mlx5_ibv_shared *sh = priv->sh;
2001 	struct mlx5_dev_config *sh_conf = NULL;
2002 	uint16_t port_id;
2003 
2004 	assert(sh);
2005 	/* Nothing to compare for the single/first device. */
2006 	if (sh->refcnt == 1)
2007 		return 0;
2008 	/* Find the device with shared context. */
2009 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2010 		struct mlx5_priv *opriv =
2011 			rte_eth_devices[port_id].data->dev_private;
2012 
2013 		if (opriv && opriv != priv && opriv->sh == sh) {
2014 			sh_conf = &opriv->config;
2015 			break;
2016 		}
2017 	}
2018 	if (!sh_conf)
2019 		return 0;
2020 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2021 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2022 			     " for shared %s context", sh->ibdev_name);
2023 		rte_errno = EINVAL;
2024 		return rte_errno;
2025 	}
2026 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2027 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2028 			     " for shared %s context", sh->ibdev_name);
2029 		rte_errno = EINVAL;
2030 		return rte_errno;
2031 	}
2032 	return 0;
2033 }
2034 /**
2035  * Spawn an Ethernet device from Verbs information.
2036  *
2037  * @param dpdk_dev
2038  *   Backing DPDK device.
2039  * @param spawn
2040  *   Verbs device parameters (name, port, switch_info) to spawn.
2041  * @param config
2042  *   Device configuration parameters.
2043  *
2044  * @return
2045  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2046  *   is set. The following errors are defined:
2047  *
2048  *   EBUSY: device is not supposed to be spawned.
2049  *   EEXIST: device is already spawned
2050  */
2051 static struct rte_eth_dev *
2052 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2053 	       struct mlx5_dev_spawn_data *spawn,
2054 	       struct mlx5_dev_config config)
2055 {
2056 	const struct mlx5_switch_info *switch_info = &spawn->info;
2057 	struct mlx5_ibv_shared *sh = NULL;
2058 	struct ibv_port_attr port_attr;
2059 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2060 	struct rte_eth_dev *eth_dev = NULL;
2061 	struct mlx5_priv *priv = NULL;
2062 	int err = 0;
2063 	unsigned int hw_padding = 0;
2064 	unsigned int mps;
2065 	unsigned int cqe_comp;
2066 	unsigned int cqe_pad = 0;
2067 	unsigned int tunnel_en = 0;
2068 	unsigned int mpls_en = 0;
2069 	unsigned int swp = 0;
2070 	unsigned int mprq = 0;
2071 	unsigned int mprq_min_stride_size_n = 0;
2072 	unsigned int mprq_max_stride_size_n = 0;
2073 	unsigned int mprq_min_stride_num_n = 0;
2074 	unsigned int mprq_max_stride_num_n = 0;
2075 	struct rte_ether_addr mac;
2076 	char name[RTE_ETH_NAME_MAX_LEN];
2077 	int own_domain_id = 0;
2078 	uint16_t port_id;
2079 	unsigned int i;
2080 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2081 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2082 #endif
2083 
2084 	/* Determine if this port representor is supposed to be spawned. */
2085 	if (switch_info->representor && dpdk_dev->devargs) {
2086 		struct rte_eth_devargs eth_da;
2087 
2088 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2089 		if (err) {
2090 			rte_errno = -err;
2091 			DRV_LOG(ERR, "failed to process device arguments: %s",
2092 				strerror(rte_errno));
2093 			return NULL;
2094 		}
2095 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
2096 			if (eth_da.representor_ports[i] ==
2097 			    (uint16_t)switch_info->port_name)
2098 				break;
2099 		if (i == eth_da.nb_representor_ports) {
2100 			rte_errno = EBUSY;
2101 			return NULL;
2102 		}
2103 	}
2104 	/* Build device name. */
2105 	if (spawn->pf_bond <  0) {
2106 		/* Single device. */
2107 		if (!switch_info->representor)
2108 			strlcpy(name, dpdk_dev->name, sizeof(name));
2109 		else
2110 			snprintf(name, sizeof(name), "%s_representor_%u",
2111 				 dpdk_dev->name, switch_info->port_name);
2112 	} else {
2113 		/* Bonding device. */
2114 		if (!switch_info->representor)
2115 			snprintf(name, sizeof(name), "%s_%s",
2116 				 dpdk_dev->name, spawn->ibv_dev->name);
2117 		else
2118 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
2119 				 dpdk_dev->name, spawn->ibv_dev->name,
2120 				 switch_info->port_name);
2121 	}
2122 	/* check if the device is already spawned */
2123 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2124 		rte_errno = EEXIST;
2125 		return NULL;
2126 	}
2127 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2128 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2129 		eth_dev = rte_eth_dev_attach_secondary(name);
2130 		if (eth_dev == NULL) {
2131 			DRV_LOG(ERR, "can not attach rte ethdev");
2132 			rte_errno = ENOMEM;
2133 			return NULL;
2134 		}
2135 		eth_dev->device = dpdk_dev;
2136 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2137 		err = mlx5_proc_priv_init(eth_dev);
2138 		if (err)
2139 			return NULL;
2140 		/* Receive command fd from primary process */
2141 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2142 		if (err < 0)
2143 			return NULL;
2144 		/* Remap UAR for Tx queues. */
2145 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2146 		if (err)
2147 			return NULL;
2148 		/*
2149 		 * Ethdev pointer is still required as input since
2150 		 * the primary device is not accessible from the
2151 		 * secondary process.
2152 		 */
2153 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2154 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2155 		return eth_dev;
2156 	}
2157 	/*
2158 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2159 	 * advance to create dv/verbs device context. We proceed the
2160 	 * devargs here to get ones, and later proceed devargs again
2161 	 * to override some hardware settings.
2162 	 */
2163 	err = mlx5_args(&config, dpdk_dev->devargs);
2164 	if (err) {
2165 		err = rte_errno;
2166 		DRV_LOG(ERR, "failed to process device arguments: %s",
2167 			strerror(rte_errno));
2168 		goto error;
2169 	}
2170 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2171 	if (!sh)
2172 		return NULL;
2173 	config.devx = sh->devx;
2174 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2175 	config.dest_tir = 1;
2176 #endif
2177 #ifdef HAVE_IBV_MLX5_MOD_SWP
2178 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2179 #endif
2180 	/*
2181 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2182 	 * as all ConnectX-5 devices.
2183 	 */
2184 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2185 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2186 #endif
2187 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2188 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2189 #endif
2190 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2191 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2192 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2193 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2194 			mps = MLX5_MPW_ENHANCED;
2195 		} else {
2196 			DRV_LOG(DEBUG, "MPW is supported");
2197 			mps = MLX5_MPW;
2198 		}
2199 	} else {
2200 		DRV_LOG(DEBUG, "MPW isn't supported");
2201 		mps = MLX5_MPW_DISABLED;
2202 	}
2203 #ifdef HAVE_IBV_MLX5_MOD_SWP
2204 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2205 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2206 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2207 #endif
2208 	config.swp = !!swp;
2209 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2210 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2211 		struct mlx5dv_striding_rq_caps mprq_caps =
2212 			dv_attr.striding_rq_caps;
2213 
2214 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2215 			mprq_caps.min_single_stride_log_num_of_bytes);
2216 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2217 			mprq_caps.max_single_stride_log_num_of_bytes);
2218 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2219 			mprq_caps.min_single_wqe_log_num_of_strides);
2220 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2221 			mprq_caps.max_single_wqe_log_num_of_strides);
2222 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2223 			mprq_caps.supported_qpts);
2224 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2225 		mprq = 1;
2226 		mprq_min_stride_size_n =
2227 			mprq_caps.min_single_stride_log_num_of_bytes;
2228 		mprq_max_stride_size_n =
2229 			mprq_caps.max_single_stride_log_num_of_bytes;
2230 		mprq_min_stride_num_n =
2231 			mprq_caps.min_single_wqe_log_num_of_strides;
2232 		mprq_max_stride_num_n =
2233 			mprq_caps.max_single_wqe_log_num_of_strides;
2234 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2235 						   mprq_min_stride_num_n);
2236 	}
2237 #endif
2238 	if (RTE_CACHE_LINE_SIZE == 128 &&
2239 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2240 		cqe_comp = 0;
2241 	else
2242 		cqe_comp = 1;
2243 	config.cqe_comp = cqe_comp;
2244 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2245 	/* Whether device supports 128B Rx CQE padding. */
2246 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2247 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2248 #endif
2249 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2250 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2251 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2252 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2253 			     (dv_attr.tunnel_offloads_caps &
2254 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
2255 	}
2256 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2257 		tunnel_en ? "" : "not ");
2258 #else
2259 	DRV_LOG(WARNING,
2260 		"tunnel offloading disabled due to old OFED/rdma-core version");
2261 #endif
2262 	config.tunnel_en = tunnel_en;
2263 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2264 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2265 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2266 		   (dv_attr.tunnel_offloads_caps &
2267 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2268 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2269 		mpls_en ? "" : "not ");
2270 #else
2271 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2272 		" old OFED/rdma-core version or firmware configuration");
2273 #endif
2274 	config.mpls_en = mpls_en;
2275 	/* Check port status. */
2276 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2277 	if (err) {
2278 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2279 		goto error;
2280 	}
2281 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2282 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2283 		err = EINVAL;
2284 		goto error;
2285 	}
2286 	if (port_attr.state != IBV_PORT_ACTIVE)
2287 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2288 			mlx5_glue->port_state_str(port_attr.state),
2289 			port_attr.state);
2290 	/* Allocate private eth device data. */
2291 	priv = rte_zmalloc("ethdev private structure",
2292 			   sizeof(*priv),
2293 			   RTE_CACHE_LINE_SIZE);
2294 	if (priv == NULL) {
2295 		DRV_LOG(ERR, "priv allocation failure");
2296 		err = ENOMEM;
2297 		goto error;
2298 	}
2299 	priv->sh = sh;
2300 	priv->ibv_port = spawn->ibv_port;
2301 	priv->pci_dev = spawn->pci_dev;
2302 	priv->mtu = RTE_ETHER_MTU;
2303 #ifndef RTE_ARCH_64
2304 	/* Initialize UAR access locks for 32bit implementations. */
2305 	rte_spinlock_init(&priv->uar_lock_cq);
2306 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2307 		rte_spinlock_init(&priv->uar_lock[i]);
2308 #endif
2309 	/* Some internal functions rely on Netlink sockets, open them now. */
2310 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2311 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2312 	priv->nl_sn = 0;
2313 	priv->representor = !!switch_info->representor;
2314 	priv->master = !!switch_info->master;
2315 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2316 	priv->vport_meta_tag = 0;
2317 	priv->vport_meta_mask = 0;
2318 	priv->pf_bond = spawn->pf_bond;
2319 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2320 	/*
2321 	 * The DevX port query API is implemented. E-Switch may use
2322 	 * either vport or reg_c[0] metadata register to match on
2323 	 * vport index. The engaged part of metadata register is
2324 	 * defined by mask.
2325 	 */
2326 	if (switch_info->representor || switch_info->master) {
2327 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2328 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2329 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2330 						 &devx_port);
2331 		if (err) {
2332 			DRV_LOG(WARNING,
2333 				"can't query devx port %d on device %s",
2334 				spawn->ibv_port, spawn->ibv_dev->name);
2335 			devx_port.comp_mask = 0;
2336 		}
2337 	}
2338 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2339 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2340 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2341 		if (!priv->vport_meta_mask) {
2342 			DRV_LOG(ERR, "vport zero mask for port %d"
2343 				     " on bonding device %s",
2344 				     spawn->ibv_port, spawn->ibv_dev->name);
2345 			err = ENOTSUP;
2346 			goto error;
2347 		}
2348 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2349 			DRV_LOG(ERR, "invalid vport tag for port %d"
2350 				     " on bonding device %s",
2351 				     spawn->ibv_port, spawn->ibv_dev->name);
2352 			err = ENOTSUP;
2353 			goto error;
2354 		}
2355 	}
2356 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2357 		priv->vport_id = devx_port.vport_num;
2358 	} else if (spawn->pf_bond >= 0) {
2359 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2360 			     " on bonding device %s",
2361 			     spawn->ibv_port, spawn->ibv_dev->name);
2362 		err = ENOTSUP;
2363 		goto error;
2364 	} else {
2365 		/* Suppose vport index in compatible way. */
2366 		priv->vport_id = switch_info->representor ?
2367 				 switch_info->port_name + 1 : -1;
2368 	}
2369 #else
2370 	/*
2371 	 * Kernel/rdma_core support single E-Switch per PF configurations
2372 	 * only and vport_id field contains the vport index for
2373 	 * associated VF, which is deduced from representor port name.
2374 	 * For example, let's have the IB device port 10, it has
2375 	 * attached network device eth0, which has port name attribute
2376 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2377 	 * as 3 (2+1). This assigning schema should be changed if the
2378 	 * multiple E-Switch instances per PF configurations or/and PCI
2379 	 * subfunctions are added.
2380 	 */
2381 	priv->vport_id = switch_info->representor ?
2382 			 switch_info->port_name + 1 : -1;
2383 #endif
2384 	/* representor_id field keeps the unmodified VF index. */
2385 	priv->representor_id = switch_info->representor ?
2386 			       switch_info->port_name : -1;
2387 	/*
2388 	 * Look for sibling devices in order to reuse their switch domain
2389 	 * if any, otherwise allocate one.
2390 	 */
2391 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2392 		const struct mlx5_priv *opriv =
2393 			rte_eth_devices[port_id].data->dev_private;
2394 
2395 		if (!opriv ||
2396 		    opriv->sh != priv->sh ||
2397 			opriv->domain_id ==
2398 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2399 			continue;
2400 		priv->domain_id = opriv->domain_id;
2401 		break;
2402 	}
2403 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2404 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2405 		if (err) {
2406 			err = rte_errno;
2407 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2408 				strerror(rte_errno));
2409 			goto error;
2410 		}
2411 		own_domain_id = 1;
2412 	}
2413 	/* Override some values set by hardware configuration. */
2414 	mlx5_args(&config, dpdk_dev->devargs);
2415 	err = mlx5_dev_check_sibling_config(priv, &config);
2416 	if (err)
2417 		goto error;
2418 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2419 			    IBV_DEVICE_RAW_IP_CSUM);
2420 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2421 		(config.hw_csum ? "" : "not "));
2422 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2423 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2424 	DRV_LOG(DEBUG, "counters are not supported");
2425 #endif
2426 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
2427 	if (config.dv_flow_en) {
2428 		DRV_LOG(WARNING, "DV flow is not supported");
2429 		config.dv_flow_en = 0;
2430 	}
2431 #endif
2432 	config.ind_table_max_size =
2433 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2434 	/*
2435 	 * Remove this check once DPDK supports larger/variable
2436 	 * indirection tables.
2437 	 */
2438 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2439 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2440 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2441 		config.ind_table_max_size);
2442 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2443 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2444 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2445 		(config.hw_vlan_strip ? "" : "not "));
2446 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2447 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2448 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2449 		(config.hw_fcs_strip ? "" : "not "));
2450 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2451 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2452 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2453 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2454 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2455 #endif
2456 	if (config.hw_padding && !hw_padding) {
2457 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2458 		config.hw_padding = 0;
2459 	} else if (config.hw_padding) {
2460 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2461 	}
2462 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2463 		      (sh->device_attr.tso_caps.supported_qpts &
2464 		       (1 << IBV_QPT_RAW_PACKET)));
2465 	if (config.tso)
2466 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2467 	/*
2468 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2469 	 * by default.
2470 	 */
2471 	if (config.mps == MLX5_ARG_UNSET)
2472 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2473 							  MLX5_MPW_DISABLED;
2474 	else
2475 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2476 	DRV_LOG(INFO, "%sMPS is %s",
2477 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
2478 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2479 	if (config.cqe_comp && !cqe_comp) {
2480 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2481 		config.cqe_comp = 0;
2482 	}
2483 	if (config.cqe_pad && !cqe_pad) {
2484 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2485 		config.cqe_pad = 0;
2486 	} else if (config.cqe_pad) {
2487 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2488 	}
2489 	if (config.devx) {
2490 		priv->counter_fallback = 0;
2491 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2492 		if (err) {
2493 			err = -err;
2494 			goto error;
2495 		}
2496 		if (!config.hca_attr.flow_counters_dump)
2497 			priv->counter_fallback = 1;
2498 #ifndef HAVE_IBV_DEVX_ASYNC
2499 		priv->counter_fallback = 1;
2500 #endif
2501 		if (priv->counter_fallback)
2502 			DRV_LOG(INFO, "Use fall-back DV counter management");
2503 		/* Check for LRO support. */
2504 		if (config.dest_tir && config.hca_attr.lro_cap &&
2505 		    config.dv_flow_en) {
2506 			/* TBD check tunnel lro caps. */
2507 			config.lro.supported = config.hca_attr.lro_cap;
2508 			DRV_LOG(DEBUG, "Device supports LRO");
2509 			/*
2510 			 * If LRO timeout is not configured by application,
2511 			 * use the minimal supported value.
2512 			 */
2513 			if (!config.lro.timeout)
2514 				config.lro.timeout =
2515 				config.hca_attr.lro_timer_supported_periods[0];
2516 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2517 				config.lro.timeout);
2518 		}
2519 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2520 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2521 		    config.dv_flow_en) {
2522 			uint8_t reg_c_mask =
2523 				config.hca_attr.qos.flow_meter_reg_c_ids;
2524 			/*
2525 			 * Meter needs two REG_C's for color match and pre-sfx
2526 			 * flow match. Here get the REG_C for color match.
2527 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2528 			 */
2529 			reg_c_mask &= 0xfc;
2530 			if (__builtin_popcount(reg_c_mask) < 1) {
2531 				priv->mtr_en = 0;
2532 				DRV_LOG(WARNING, "No available register for"
2533 					" meter.");
2534 			} else {
2535 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2536 						      REG_C_0;
2537 				priv->mtr_en = 1;
2538 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2539 					priv->mtr_color_reg);
2540 			}
2541 		}
2542 #endif
2543 	}
2544 	if (config.mprq.enabled && mprq) {
2545 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2546 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2547 			config.mprq.stride_num_n =
2548 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2549 					mprq_min_stride_num_n);
2550 			DRV_LOG(WARNING,
2551 				"the number of strides"
2552 				" for Multi-Packet RQ is out of range,"
2553 				" setting default value (%u)",
2554 				1 << config.mprq.stride_num_n);
2555 		}
2556 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2557 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2558 	} else if (config.mprq.enabled && !mprq) {
2559 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2560 		config.mprq.enabled = 0;
2561 	}
2562 	if (config.max_dump_files_num == 0)
2563 		config.max_dump_files_num = 128;
2564 	eth_dev = rte_eth_dev_allocate(name);
2565 	if (eth_dev == NULL) {
2566 		DRV_LOG(ERR, "can not allocate rte ethdev");
2567 		err = ENOMEM;
2568 		goto error;
2569 	}
2570 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2571 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2572 	if (priv->representor) {
2573 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2574 		eth_dev->data->representor_id = priv->representor_id;
2575 	}
2576 	/*
2577 	 * Store associated network device interface index. This index
2578 	 * is permanent throughout the lifetime of device. So, we may store
2579 	 * the ifindex here and use the cached value further.
2580 	 */
2581 	assert(spawn->ifindex);
2582 	priv->if_index = spawn->ifindex;
2583 	eth_dev->data->dev_private = priv;
2584 	priv->dev_data = eth_dev->data;
2585 	eth_dev->data->mac_addrs = priv->mac;
2586 	eth_dev->device = dpdk_dev;
2587 	/* Configure the first MAC address by default. */
2588 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2589 		DRV_LOG(ERR,
2590 			"port %u cannot get MAC address, is mlx5_en"
2591 			" loaded? (errno: %s)",
2592 			eth_dev->data->port_id, strerror(rte_errno));
2593 		err = ENODEV;
2594 		goto error;
2595 	}
2596 	DRV_LOG(INFO,
2597 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2598 		eth_dev->data->port_id,
2599 		mac.addr_bytes[0], mac.addr_bytes[1],
2600 		mac.addr_bytes[2], mac.addr_bytes[3],
2601 		mac.addr_bytes[4], mac.addr_bytes[5]);
2602 #ifndef NDEBUG
2603 	{
2604 		char ifname[IF_NAMESIZE];
2605 
2606 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2607 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2608 				eth_dev->data->port_id, ifname);
2609 		else
2610 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2611 				eth_dev->data->port_id);
2612 	}
2613 #endif
2614 	/* Get actual MTU if possible. */
2615 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2616 	if (err) {
2617 		err = rte_errno;
2618 		goto error;
2619 	}
2620 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2621 		priv->mtu);
2622 	/* Initialize burst functions to prevent crashes before link-up. */
2623 	eth_dev->rx_pkt_burst = removed_rx_burst;
2624 	eth_dev->tx_pkt_burst = removed_tx_burst;
2625 	eth_dev->dev_ops = &mlx5_dev_ops;
2626 	/* Register MAC address. */
2627 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2628 	if (config.vf && config.vf_nl_en)
2629 		mlx5_nl_mac_addr_sync(eth_dev);
2630 	TAILQ_INIT(&priv->flows);
2631 	TAILQ_INIT(&priv->ctrl_flows);
2632 	TAILQ_INIT(&priv->flow_meters);
2633 	TAILQ_INIT(&priv->flow_meter_profiles);
2634 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2635 	struct mlx5dv_ctx_allocators alctr = {
2636 		.alloc = &mlx5_alloc_verbs_buf,
2637 		.free = &mlx5_free_verbs_buf,
2638 		.data = priv,
2639 	};
2640 	mlx5_glue->dv_set_context_attr(sh->ctx,
2641 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2642 				       (void *)((uintptr_t)&alctr));
2643 	/* Bring Ethernet device up. */
2644 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2645 		eth_dev->data->port_id);
2646 	mlx5_set_link_up(eth_dev);
2647 	/*
2648 	 * Even though the interrupt handler is not installed yet,
2649 	 * interrupts will still trigger on the async_fd from
2650 	 * Verbs context returned by ibv_open_device().
2651 	 */
2652 	mlx5_link_update(eth_dev, 0);
2653 #ifdef HAVE_MLX5DV_DR_ESWITCH
2654 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2655 	      (switch_info->representor || switch_info->master)))
2656 		config.dv_esw_en = 0;
2657 #else
2658 	config.dv_esw_en = 0;
2659 #endif
2660 	/* Detect minimal data bytes to inline. */
2661 	mlx5_set_min_inline(spawn, &config);
2662 	/* Store device configuration on private structure. */
2663 	priv->config = config;
2664 	/* Create context for virtual machine VLAN workaround. */
2665 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2666 	if (config.dv_flow_en) {
2667 		err = mlx5_alloc_shared_dr(priv);
2668 		if (err)
2669 			goto error;
2670 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc();
2671 		if (!priv->qrss_id_pool) {
2672 			DRV_LOG(ERR, "can't create flow id pool");
2673 			err = ENOMEM;
2674 			goto error;
2675 		}
2676 	}
2677 	/* Supported Verbs flow priority number detection. */
2678 	err = mlx5_flow_discover_priorities(eth_dev);
2679 	if (err < 0) {
2680 		err = -err;
2681 		goto error;
2682 	}
2683 	priv->config.flow_prio = err;
2684 	if (!priv->config.dv_esw_en &&
2685 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2686 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2687 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2688 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2689 	}
2690 	mlx5_set_metadata_mask(eth_dev);
2691 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2692 	    !priv->sh->dv_regc0_mask) {
2693 		DRV_LOG(ERR, "metadata mode %u is not supported "
2694 			     "(no metadata reg_c[0] is available)",
2695 			     priv->config.dv_xmeta_en);
2696 			err = ENOTSUP;
2697 			goto error;
2698 	}
2699 	/* Query availibility of metadata reg_c's. */
2700 	err = mlx5_flow_discover_mreg_c(eth_dev);
2701 	if (err < 0) {
2702 		err = -err;
2703 		goto error;
2704 	}
2705 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2706 		DRV_LOG(DEBUG,
2707 			"port %u extensive metadata register is not supported",
2708 			eth_dev->data->port_id);
2709 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2710 			DRV_LOG(ERR, "metadata mode %u is not supported "
2711 				     "(no metadata registers available)",
2712 				     priv->config.dv_xmeta_en);
2713 			err = ENOTSUP;
2714 			goto error;
2715 		}
2716 	}
2717 	if (priv->config.dv_flow_en &&
2718 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2719 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
2720 	    priv->sh->dv_regc0_mask) {
2721 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2722 						      MLX5_FLOW_MREG_HTABLE_SZ);
2723 		if (!priv->mreg_cp_tbl) {
2724 			err = ENOMEM;
2725 			goto error;
2726 		}
2727 	}
2728 	return eth_dev;
2729 error:
2730 	if (priv) {
2731 		if (priv->mreg_cp_tbl)
2732 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2733 		if (priv->sh)
2734 			mlx5_free_shared_dr(priv);
2735 		if (priv->nl_socket_route >= 0)
2736 			close(priv->nl_socket_route);
2737 		if (priv->nl_socket_rdma >= 0)
2738 			close(priv->nl_socket_rdma);
2739 		if (priv->vmwa_context)
2740 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2741 		if (priv->qrss_id_pool)
2742 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
2743 		if (own_domain_id)
2744 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2745 		rte_free(priv);
2746 		if (eth_dev != NULL)
2747 			eth_dev->data->dev_private = NULL;
2748 	}
2749 	if (eth_dev != NULL) {
2750 		/* mac_addrs must not be freed alone because part of dev_private */
2751 		eth_dev->data->mac_addrs = NULL;
2752 		rte_eth_dev_release_port(eth_dev);
2753 	}
2754 	if (sh)
2755 		mlx5_free_shared_ibctx(sh);
2756 	assert(err > 0);
2757 	rte_errno = err;
2758 	return NULL;
2759 }
2760 
2761 /**
2762  * Comparison callback to sort device data.
2763  *
2764  * This is meant to be used with qsort().
2765  *
2766  * @param a[in]
2767  *   Pointer to pointer to first data object.
2768  * @param b[in]
2769  *   Pointer to pointer to second data object.
2770  *
2771  * @return
2772  *   0 if both objects are equal, less than 0 if the first argument is less
2773  *   than the second, greater than 0 otherwise.
2774  */
2775 static int
2776 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2777 {
2778 	const struct mlx5_switch_info *si_a =
2779 		&((const struct mlx5_dev_spawn_data *)a)->info;
2780 	const struct mlx5_switch_info *si_b =
2781 		&((const struct mlx5_dev_spawn_data *)b)->info;
2782 	int ret;
2783 
2784 	/* Master device first. */
2785 	ret = si_b->master - si_a->master;
2786 	if (ret)
2787 		return ret;
2788 	/* Then representor devices. */
2789 	ret = si_b->representor - si_a->representor;
2790 	if (ret)
2791 		return ret;
2792 	/* Unidentified devices come last in no specific order. */
2793 	if (!si_a->representor)
2794 		return 0;
2795 	/* Order representors by name. */
2796 	return si_a->port_name - si_b->port_name;
2797 }
2798 
2799 /**
2800  * Match PCI information for possible slaves of bonding device.
2801  *
2802  * @param[in] ibv_dev
2803  *   Pointer to Infiniband device structure.
2804  * @param[in] pci_dev
2805  *   Pointer to PCI device structure to match PCI address.
2806  * @param[in] nl_rdma
2807  *   Netlink RDMA group socket handle.
2808  *
2809  * @return
2810  *   negative value if no bonding device found, otherwise
2811  *   positive index of slave PF in bonding.
2812  */
2813 static int
2814 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2815 			   const struct rte_pci_device *pci_dev,
2816 			   int nl_rdma)
2817 {
2818 	char ifname[IF_NAMESIZE + 1];
2819 	unsigned int ifindex;
2820 	unsigned int np, i;
2821 	FILE *file = NULL;
2822 	int pf = -1;
2823 
2824 	/*
2825 	 * Try to get master device name. If something goes
2826 	 * wrong suppose the lack of kernel support and no
2827 	 * bonding devices.
2828 	 */
2829 	if (nl_rdma < 0)
2830 		return -1;
2831 	if (!strstr(ibv_dev->name, "bond"))
2832 		return -1;
2833 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2834 	if (!np)
2835 		return -1;
2836 	/*
2837 	 * The Master device might not be on the predefined
2838 	 * port (not on port index 1, it is not garanted),
2839 	 * we have to scan all Infiniband device port and
2840 	 * find master.
2841 	 */
2842 	for (i = 1; i <= np; ++i) {
2843 		/* Check whether Infiniband port is populated. */
2844 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2845 		if (!ifindex)
2846 			continue;
2847 		if (!if_indextoname(ifindex, ifname))
2848 			continue;
2849 		/* Try to read bonding slave names from sysfs. */
2850 		MKSTR(slaves,
2851 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2852 		file = fopen(slaves, "r");
2853 		if (file)
2854 			break;
2855 	}
2856 	if (!file)
2857 		return -1;
2858 	/* Use safe format to check maximal buffer length. */
2859 	assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2860 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2861 		char tmp_str[IF_NAMESIZE + 32];
2862 		struct rte_pci_addr pci_addr;
2863 		struct mlx5_switch_info	info;
2864 
2865 		/* Process slave interface names in the loop. */
2866 		snprintf(tmp_str, sizeof(tmp_str),
2867 			 "/sys/class/net/%s", ifname);
2868 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2869 			DRV_LOG(WARNING, "can not get PCI address"
2870 					 " for netdev \"%s\"", ifname);
2871 			continue;
2872 		}
2873 		if (pci_dev->addr.domain != pci_addr.domain ||
2874 		    pci_dev->addr.bus != pci_addr.bus ||
2875 		    pci_dev->addr.devid != pci_addr.devid ||
2876 		    pci_dev->addr.function != pci_addr.function)
2877 			continue;
2878 		/* Slave interface PCI address match found. */
2879 		fclose(file);
2880 		snprintf(tmp_str, sizeof(tmp_str),
2881 			 "/sys/class/net/%s/phys_port_name", ifname);
2882 		file = fopen(tmp_str, "rb");
2883 		if (!file)
2884 			break;
2885 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2886 		if (fscanf(file, "%32s", tmp_str) == 1)
2887 			mlx5_translate_port_name(tmp_str, &info);
2888 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2889 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2890 			pf = info.port_name;
2891 		break;
2892 	}
2893 	if (file)
2894 		fclose(file);
2895 	return pf;
2896 }
2897 
2898 /**
2899  * DPDK callback to register a PCI device.
2900  *
2901  * This function spawns Ethernet devices out of a given PCI device.
2902  *
2903  * @param[in] pci_drv
2904  *   PCI driver structure (mlx5_driver).
2905  * @param[in] pci_dev
2906  *   PCI device information.
2907  *
2908  * @return
2909  *   0 on success, a negative errno value otherwise and rte_errno is set.
2910  */
2911 static int
2912 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2913 	       struct rte_pci_device *pci_dev)
2914 {
2915 	struct ibv_device **ibv_list;
2916 	/*
2917 	 * Number of found IB Devices matching with requested PCI BDF.
2918 	 * nd != 1 means there are multiple IB devices over the same
2919 	 * PCI device and we have representors and master.
2920 	 */
2921 	unsigned int nd = 0;
2922 	/*
2923 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2924 	 * we have the single multiport IB device, and there may be
2925 	 * representors attached to some of found ports.
2926 	 */
2927 	unsigned int np = 0;
2928 	/*
2929 	 * Number of DPDK ethernet devices to Spawn - either over
2930 	 * multiple IB devices or multiple ports of single IB device.
2931 	 * Actually this is the number of iterations to spawn.
2932 	 */
2933 	unsigned int ns = 0;
2934 	/*
2935 	 * Bonding device
2936 	 *   < 0 - no bonding device (single one)
2937 	 *  >= 0 - bonding device (value is slave PF index)
2938 	 */
2939 	int bd = -1;
2940 	struct mlx5_dev_spawn_data *list = NULL;
2941 	struct mlx5_dev_config dev_config;
2942 	int ret;
2943 
2944 	ret = mlx5_init_once();
2945 	if (ret) {
2946 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2947 			strerror(rte_errno));
2948 		return -rte_errno;
2949 	}
2950 	assert(pci_drv == &mlx5_driver);
2951 	errno = 0;
2952 	ibv_list = mlx5_glue->get_device_list(&ret);
2953 	if (!ibv_list) {
2954 		rte_errno = errno ? errno : ENOSYS;
2955 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2956 		return -rte_errno;
2957 	}
2958 	/*
2959 	 * First scan the list of all Infiniband devices to find
2960 	 * matching ones, gathering into the list.
2961 	 */
2962 	struct ibv_device *ibv_match[ret + 1];
2963 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2964 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2965 	unsigned int i;
2966 
2967 	while (ret-- > 0) {
2968 		struct rte_pci_addr pci_addr;
2969 
2970 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2971 		bd = mlx5_device_bond_pci_match
2972 				(ibv_list[ret], pci_dev, nl_rdma);
2973 		if (bd >= 0) {
2974 			/*
2975 			 * Bonding device detected. Only one match is allowed,
2976 			 * the bonding is supported over multi-port IB device,
2977 			 * there should be no matches on representor PCI
2978 			 * functions or non VF LAG bonding devices with
2979 			 * specified address.
2980 			 */
2981 			if (nd) {
2982 				DRV_LOG(ERR,
2983 					"multiple PCI match on bonding device"
2984 					"\"%s\" found", ibv_list[ret]->name);
2985 				rte_errno = ENOENT;
2986 				ret = -rte_errno;
2987 				goto exit;
2988 			}
2989 			DRV_LOG(INFO, "PCI information matches for"
2990 				      " slave %d bonding device \"%s\"",
2991 				      bd, ibv_list[ret]->name);
2992 			ibv_match[nd++] = ibv_list[ret];
2993 			break;
2994 		}
2995 		if (mlx5_dev_to_pci_addr
2996 			(ibv_list[ret]->ibdev_path, &pci_addr))
2997 			continue;
2998 		if (pci_dev->addr.domain != pci_addr.domain ||
2999 		    pci_dev->addr.bus != pci_addr.bus ||
3000 		    pci_dev->addr.devid != pci_addr.devid ||
3001 		    pci_dev->addr.function != pci_addr.function)
3002 			continue;
3003 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3004 			ibv_list[ret]->name);
3005 		ibv_match[nd++] = ibv_list[ret];
3006 	}
3007 	ibv_match[nd] = NULL;
3008 	if (!nd) {
3009 		/* No device matches, just complain and bail out. */
3010 		DRV_LOG(WARNING,
3011 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
3012 			" are kernel drivers loaded?",
3013 			pci_dev->addr.domain, pci_dev->addr.bus,
3014 			pci_dev->addr.devid, pci_dev->addr.function);
3015 		rte_errno = ENOENT;
3016 		ret = -rte_errno;
3017 		goto exit;
3018 	}
3019 	if (nd == 1) {
3020 		/*
3021 		 * Found single matching device may have multiple ports.
3022 		 * Each port may be representor, we have to check the port
3023 		 * number and check the representors existence.
3024 		 */
3025 		if (nl_rdma >= 0)
3026 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3027 		if (!np)
3028 			DRV_LOG(WARNING, "can not get IB device \"%s\""
3029 					 " ports number", ibv_match[0]->name);
3030 		if (bd >= 0 && !np) {
3031 			DRV_LOG(ERR, "can not get ports"
3032 				     " for bonding device");
3033 			rte_errno = ENOENT;
3034 			ret = -rte_errno;
3035 			goto exit;
3036 		}
3037 	}
3038 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3039 	if (bd >= 0) {
3040 		/*
3041 		 * This may happen if there is VF LAG kernel support and
3042 		 * application is compiled with older rdma_core library.
3043 		 */
3044 		DRV_LOG(ERR,
3045 			"No kernel/verbs support for VF LAG bonding found.");
3046 		rte_errno = ENOTSUP;
3047 		ret = -rte_errno;
3048 		goto exit;
3049 	}
3050 #endif
3051 	/*
3052 	 * Now we can determine the maximal
3053 	 * amount of devices to be spawned.
3054 	 */
3055 	list = rte_zmalloc("device spawn data",
3056 			 sizeof(struct mlx5_dev_spawn_data) *
3057 			 (np ? np : nd),
3058 			 RTE_CACHE_LINE_SIZE);
3059 	if (!list) {
3060 		DRV_LOG(ERR, "spawn data array allocation failure");
3061 		rte_errno = ENOMEM;
3062 		ret = -rte_errno;
3063 		goto exit;
3064 	}
3065 	if (bd >= 0 || np > 1) {
3066 		/*
3067 		 * Single IB device with multiple ports found,
3068 		 * it may be E-Switch master device and representors.
3069 		 * We have to perform identification trough the ports.
3070 		 */
3071 		assert(nl_rdma >= 0);
3072 		assert(ns == 0);
3073 		assert(nd == 1);
3074 		assert(np);
3075 		for (i = 1; i <= np; ++i) {
3076 			list[ns].max_port = np;
3077 			list[ns].ibv_port = i;
3078 			list[ns].ibv_dev = ibv_match[0];
3079 			list[ns].eth_dev = NULL;
3080 			list[ns].pci_dev = pci_dev;
3081 			list[ns].pf_bond = bd;
3082 			list[ns].ifindex = mlx5_nl_ifindex
3083 					(nl_rdma, list[ns].ibv_dev->name, i);
3084 			if (!list[ns].ifindex) {
3085 				/*
3086 				 * No network interface index found for the
3087 				 * specified port, it means there is no
3088 				 * representor on this port. It's OK,
3089 				 * there can be disabled ports, for example
3090 				 * if sriov_numvfs < sriov_totalvfs.
3091 				 */
3092 				continue;
3093 			}
3094 			ret = -1;
3095 			if (nl_route >= 0)
3096 				ret = mlx5_nl_switch_info
3097 					       (nl_route,
3098 						list[ns].ifindex,
3099 						&list[ns].info);
3100 			if (ret || (!list[ns].info.representor &&
3101 				    !list[ns].info.master)) {
3102 				/*
3103 				 * We failed to recognize representors with
3104 				 * Netlink, let's try to perform the task
3105 				 * with sysfs.
3106 				 */
3107 				ret =  mlx5_sysfs_switch_info
3108 						(list[ns].ifindex,
3109 						 &list[ns].info);
3110 			}
3111 			if (!ret && bd >= 0) {
3112 				switch (list[ns].info.name_type) {
3113 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3114 					if (list[ns].info.port_name == bd)
3115 						ns++;
3116 					break;
3117 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3118 					if (list[ns].info.pf_num == bd)
3119 						ns++;
3120 					break;
3121 				default:
3122 					break;
3123 				}
3124 				continue;
3125 			}
3126 			if (!ret && (list[ns].info.representor ^
3127 				     list[ns].info.master))
3128 				ns++;
3129 		}
3130 		if (!ns) {
3131 			DRV_LOG(ERR,
3132 				"unable to recognize master/representors"
3133 				" on the IB device with multiple ports");
3134 			rte_errno = ENOENT;
3135 			ret = -rte_errno;
3136 			goto exit;
3137 		}
3138 	} else {
3139 		/*
3140 		 * The existence of several matching entries (nd > 1) means
3141 		 * port representors have been instantiated. No existing Verbs
3142 		 * call nor sysfs entries can tell them apart, this can only
3143 		 * be done through Netlink calls assuming kernel drivers are
3144 		 * recent enough to support them.
3145 		 *
3146 		 * In the event of identification failure through Netlink,
3147 		 * try again through sysfs, then:
3148 		 *
3149 		 * 1. A single IB device matches (nd == 1) with single
3150 		 *    port (np=0/1) and is not a representor, assume
3151 		 *    no switch support.
3152 		 *
3153 		 * 2. Otherwise no safe assumptions can be made;
3154 		 *    complain louder and bail out.
3155 		 */
3156 		np = 1;
3157 		for (i = 0; i != nd; ++i) {
3158 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3159 			list[ns].max_port = 1;
3160 			list[ns].ibv_port = 1;
3161 			list[ns].ibv_dev = ibv_match[i];
3162 			list[ns].eth_dev = NULL;
3163 			list[ns].pci_dev = pci_dev;
3164 			list[ns].pf_bond = -1;
3165 			list[ns].ifindex = 0;
3166 			if (nl_rdma >= 0)
3167 				list[ns].ifindex = mlx5_nl_ifindex
3168 					(nl_rdma, list[ns].ibv_dev->name, 1);
3169 			if (!list[ns].ifindex) {
3170 				char ifname[IF_NAMESIZE];
3171 
3172 				/*
3173 				 * Netlink failed, it may happen with old
3174 				 * ib_core kernel driver (before 4.16).
3175 				 * We can assume there is old driver because
3176 				 * here we are processing single ports IB
3177 				 * devices. Let's try sysfs to retrieve
3178 				 * the ifindex. The method works for
3179 				 * master device only.
3180 				 */
3181 				if (nd > 1) {
3182 					/*
3183 					 * Multiple devices found, assume
3184 					 * representors, can not distinguish
3185 					 * master/representor and retrieve
3186 					 * ifindex via sysfs.
3187 					 */
3188 					continue;
3189 				}
3190 				ret = mlx5_get_master_ifname
3191 					(ibv_match[i]->ibdev_path, &ifname);
3192 				if (!ret)
3193 					list[ns].ifindex =
3194 						if_nametoindex(ifname);
3195 				if (!list[ns].ifindex) {
3196 					/*
3197 					 * No network interface index found
3198 					 * for the specified device, it means
3199 					 * there it is neither representor
3200 					 * nor master.
3201 					 */
3202 					continue;
3203 				}
3204 			}
3205 			ret = -1;
3206 			if (nl_route >= 0)
3207 				ret = mlx5_nl_switch_info
3208 					       (nl_route,
3209 						list[ns].ifindex,
3210 						&list[ns].info);
3211 			if (ret || (!list[ns].info.representor &&
3212 				    !list[ns].info.master)) {
3213 				/*
3214 				 * We failed to recognize representors with
3215 				 * Netlink, let's try to perform the task
3216 				 * with sysfs.
3217 				 */
3218 				ret =  mlx5_sysfs_switch_info
3219 						(list[ns].ifindex,
3220 						 &list[ns].info);
3221 			}
3222 			if (!ret && (list[ns].info.representor ^
3223 				     list[ns].info.master)) {
3224 				ns++;
3225 			} else if ((nd == 1) &&
3226 				   !list[ns].info.representor &&
3227 				   !list[ns].info.master) {
3228 				/*
3229 				 * Single IB device with
3230 				 * one physical port and
3231 				 * attached network device.
3232 				 * May be SRIOV is not enabled
3233 				 * or there is no representors.
3234 				 */
3235 				DRV_LOG(INFO, "no E-Switch support detected");
3236 				ns++;
3237 				break;
3238 			}
3239 		}
3240 		if (!ns) {
3241 			DRV_LOG(ERR,
3242 				"unable to recognize master/representors"
3243 				" on the multiple IB devices");
3244 			rte_errno = ENOENT;
3245 			ret = -rte_errno;
3246 			goto exit;
3247 		}
3248 	}
3249 	assert(ns);
3250 	/*
3251 	 * Sort list to probe devices in natural order for users convenience
3252 	 * (i.e. master first, then representors from lowest to highest ID).
3253 	 */
3254 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3255 	/* Default configuration. */
3256 	dev_config = (struct mlx5_dev_config){
3257 		.hw_padding = 0,
3258 		.mps = MLX5_ARG_UNSET,
3259 		.dbnc = MLX5_ARG_UNSET,
3260 		.rx_vec_en = 1,
3261 		.txq_inline_max = MLX5_ARG_UNSET,
3262 		.txq_inline_min = MLX5_ARG_UNSET,
3263 		.txq_inline_mpw = MLX5_ARG_UNSET,
3264 		.txqs_inline = MLX5_ARG_UNSET,
3265 		.vf_nl_en = 1,
3266 		.mr_ext_memseg_en = 1,
3267 		.mprq = {
3268 			.enabled = 0, /* Disabled by default. */
3269 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3270 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3271 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3272 		},
3273 		.dv_esw_en = 1,
3274 		.dv_flow_en = 1,
3275 	};
3276 	/* Device specific configuration. */
3277 	switch (pci_dev->id.device_id) {
3278 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3279 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3280 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3281 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3282 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3283 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3284 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3285 		dev_config.vf = 1;
3286 		break;
3287 	default:
3288 		break;
3289 	}
3290 	for (i = 0; i != ns; ++i) {
3291 		uint32_t restore;
3292 
3293 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3294 						 &list[i],
3295 						 dev_config);
3296 		if (!list[i].eth_dev) {
3297 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3298 				break;
3299 			/* Device is disabled or already spawned. Ignore it. */
3300 			continue;
3301 		}
3302 		restore = list[i].eth_dev->data->dev_flags;
3303 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3304 		/* Restore non-PCI flags cleared by the above call. */
3305 		list[i].eth_dev->data->dev_flags |= restore;
3306 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3307 		rte_eth_dev_probing_finish(list[i].eth_dev);
3308 	}
3309 	if (i != ns) {
3310 		DRV_LOG(ERR,
3311 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3312 			" encountering an error: %s",
3313 			pci_dev->addr.domain, pci_dev->addr.bus,
3314 			pci_dev->addr.devid, pci_dev->addr.function,
3315 			strerror(rte_errno));
3316 		ret = -rte_errno;
3317 		/* Roll back. */
3318 		while (i--) {
3319 			if (!list[i].eth_dev)
3320 				continue;
3321 			mlx5_dev_close(list[i].eth_dev);
3322 			/* mac_addrs must not be freed because in dev_private */
3323 			list[i].eth_dev->data->mac_addrs = NULL;
3324 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3325 		}
3326 		/* Restore original error. */
3327 		rte_errno = -ret;
3328 	} else {
3329 		ret = 0;
3330 	}
3331 exit:
3332 	/*
3333 	 * Do the routine cleanup:
3334 	 * - close opened Netlink sockets
3335 	 * - free allocated spawn data array
3336 	 * - free the Infiniband device list
3337 	 */
3338 	if (nl_rdma >= 0)
3339 		close(nl_rdma);
3340 	if (nl_route >= 0)
3341 		close(nl_route);
3342 	if (list)
3343 		rte_free(list);
3344 	assert(ibv_list);
3345 	mlx5_glue->free_device_list(ibv_list);
3346 	return ret;
3347 }
3348 
3349 /**
3350  * Look for the ethernet device belonging to mlx5 driver.
3351  *
3352  * @param[in] port_id
3353  *   port_id to start looking for device.
3354  * @param[in] pci_dev
3355  *   Pointer to the hint PCI device. When device is being probed
3356  *   the its siblings (master and preceding representors might
3357  *   not have assigned driver yet (because the mlx5_pci_probe()
3358  *   is not completed yet, for this case match on hint PCI
3359  *   device may be used to detect sibling device.
3360  *
3361  * @return
3362  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3363  */
3364 uint16_t
3365 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3366 {
3367 	while (port_id < RTE_MAX_ETHPORTS) {
3368 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3369 
3370 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3371 		    dev->device &&
3372 		    (dev->device == &pci_dev->device ||
3373 		     (dev->device->driver &&
3374 		     dev->device->driver->name &&
3375 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3376 			break;
3377 		port_id++;
3378 	}
3379 	if (port_id >= RTE_MAX_ETHPORTS)
3380 		return RTE_MAX_ETHPORTS;
3381 	return port_id;
3382 }
3383 
3384 /**
3385  * DPDK callback to remove a PCI device.
3386  *
3387  * This function removes all Ethernet devices belong to a given PCI device.
3388  *
3389  * @param[in] pci_dev
3390  *   Pointer to the PCI device.
3391  *
3392  * @return
3393  *   0 on success, the function cannot fail.
3394  */
3395 static int
3396 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3397 {
3398 	uint16_t port_id;
3399 
3400 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3401 		rte_eth_dev_close(port_id);
3402 	return 0;
3403 }
3404 
3405 static const struct rte_pci_id mlx5_pci_id_map[] = {
3406 	{
3407 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3408 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3409 	},
3410 	{
3411 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3412 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3413 	},
3414 	{
3415 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3416 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3417 	},
3418 	{
3419 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3420 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3421 	},
3422 	{
3423 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3424 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3425 	},
3426 	{
3427 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3428 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3429 	},
3430 	{
3431 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3432 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3433 	},
3434 	{
3435 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3436 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3437 	},
3438 	{
3439 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3440 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3441 	},
3442 	{
3443 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3444 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3445 	},
3446 	{
3447 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3448 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3449 	},
3450 	{
3451 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3452 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3453 	},
3454 	{
3455 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3456 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3457 	},
3458 	{
3459 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3460 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3461 	},
3462 	{
3463 		.vendor_id = 0
3464 	}
3465 };
3466 
3467 static struct rte_pci_driver mlx5_driver = {
3468 	.driver = {
3469 		.name = MLX5_DRIVER_NAME
3470 	},
3471 	.id_table = mlx5_pci_id_map,
3472 	.probe = mlx5_pci_probe,
3473 	.remove = mlx5_pci_remove,
3474 	.dma_map = mlx5_dma_map,
3475 	.dma_unmap = mlx5_dma_unmap,
3476 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3477 		     RTE_PCI_DRV_PROBE_AGAIN,
3478 };
3479 
3480 #ifdef RTE_IBVERBS_LINK_DLOPEN
3481 
3482 /**
3483  * Suffix RTE_EAL_PMD_PATH with "-glue".
3484  *
3485  * This function performs a sanity check on RTE_EAL_PMD_PATH before
3486  * suffixing its last component.
3487  *
3488  * @param buf[out]
3489  *   Output buffer, should be large enough otherwise NULL is returned.
3490  * @param size
3491  *   Size of @p out.
3492  *
3493  * @return
3494  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
3495  */
3496 static char *
3497 mlx5_glue_path(char *buf, size_t size)
3498 {
3499 	static const char *const bad[] = { "/", ".", "..", NULL };
3500 	const char *path = RTE_EAL_PMD_PATH;
3501 	size_t len = strlen(path);
3502 	size_t off;
3503 	int i;
3504 
3505 	while (len && path[len - 1] == '/')
3506 		--len;
3507 	for (off = len; off && path[off - 1] != '/'; --off)
3508 		;
3509 	for (i = 0; bad[i]; ++i)
3510 		if (!strncmp(path + off, bad[i], (int)(len - off)))
3511 			goto error;
3512 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
3513 	if (i == -1 || (size_t)i >= size)
3514 		goto error;
3515 	return buf;
3516 error:
3517 	DRV_LOG(ERR,
3518 		"unable to append \"-glue\" to last component of"
3519 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
3520 		" please re-configure DPDK");
3521 	return NULL;
3522 }
3523 
3524 /**
3525  * Initialization routine for run-time dependency on rdma-core.
3526  */
3527 static int
3528 mlx5_glue_init(void)
3529 {
3530 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
3531 	const char *path[] = {
3532 		/*
3533 		 * A basic security check is necessary before trusting
3534 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
3535 		 */
3536 		(geteuid() == getuid() && getegid() == getgid() ?
3537 		 getenv("MLX5_GLUE_PATH") : NULL),
3538 		/*
3539 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
3540 		 * variant, otherwise let dlopen() look up libraries on its
3541 		 * own.
3542 		 */
3543 		(*RTE_EAL_PMD_PATH ?
3544 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
3545 	};
3546 	unsigned int i = 0;
3547 	void *handle = NULL;
3548 	void **sym;
3549 	const char *dlmsg;
3550 
3551 	while (!handle && i != RTE_DIM(path)) {
3552 		const char *end;
3553 		size_t len;
3554 		int ret;
3555 
3556 		if (!path[i]) {
3557 			++i;
3558 			continue;
3559 		}
3560 		end = strpbrk(path[i], ":;");
3561 		if (!end)
3562 			end = path[i] + strlen(path[i]);
3563 		len = end - path[i];
3564 		ret = 0;
3565 		do {
3566 			char name[ret + 1];
3567 
3568 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
3569 				       (int)len, path[i],
3570 				       (!len || *(end - 1) == '/') ? "" : "/");
3571 			if (ret == -1)
3572 				break;
3573 			if (sizeof(name) != (size_t)ret + 1)
3574 				continue;
3575 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
3576 				name);
3577 			handle = dlopen(name, RTLD_LAZY);
3578 			break;
3579 		} while (1);
3580 		path[i] = end + 1;
3581 		if (!*end)
3582 			++i;
3583 	}
3584 	if (!handle) {
3585 		rte_errno = EINVAL;
3586 		dlmsg = dlerror();
3587 		if (dlmsg)
3588 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
3589 		goto glue_error;
3590 	}
3591 	sym = dlsym(handle, "mlx5_glue");
3592 	if (!sym || !*sym) {
3593 		rte_errno = EINVAL;
3594 		dlmsg = dlerror();
3595 		if (dlmsg)
3596 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
3597 		goto glue_error;
3598 	}
3599 	mlx5_glue = *sym;
3600 	return 0;
3601 glue_error:
3602 	if (handle)
3603 		dlclose(handle);
3604 	DRV_LOG(WARNING,
3605 		"cannot initialize PMD due to missing run-time dependency on"
3606 		" rdma-core libraries (libibverbs, libmlx5)");
3607 	return -rte_errno;
3608 }
3609 
3610 #endif
3611 
3612 /**
3613  * Driver initialization routine.
3614  */
3615 RTE_INIT(rte_mlx5_pmd_init)
3616 {
3617 	/* Initialize driver log type. */
3618 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3619 	if (mlx5_logtype >= 0)
3620 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3621 
3622 	/* Build the static tables for Verbs conversion. */
3623 	mlx5_set_ptype_table();
3624 	mlx5_set_cksum_table();
3625 	mlx5_set_swp_types_table();
3626 	/*
3627 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3628 	 * huge pages. Calling ibv_fork_init() during init allows
3629 	 * applications to use fork() safely for purposes other than
3630 	 * using this PMD, which is not supported in forked processes.
3631 	 */
3632 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3633 	/* Match the size of Rx completion entry to the size of a cacheline. */
3634 	if (RTE_CACHE_LINE_SIZE == 128)
3635 		setenv("MLX5_CQE_SIZE", "128", 0);
3636 	/*
3637 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
3638 	 * cleanup all the Verbs resources even when the device was removed.
3639 	 */
3640 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
3641 #ifdef RTE_IBVERBS_LINK_DLOPEN
3642 	if (mlx5_glue_init())
3643 		return;
3644 	assert(mlx5_glue);
3645 #endif
3646 #ifndef NDEBUG
3647 	/* Glue structure must not contain any NULL pointers. */
3648 	{
3649 		unsigned int i;
3650 
3651 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
3652 			assert(((const void *const *)mlx5_glue)[i]);
3653 	}
3654 #endif
3655 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
3656 		DRV_LOG(ERR,
3657 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
3658 			mlx5_glue->version, MLX5_GLUE_VERSION);
3659 		return;
3660 	}
3661 	mlx5_glue->fork_init();
3662 	rte_pci_register(&mlx5_driver);
3663 }
3664 
3665 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3666 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3667 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3668