xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 9cd9d3e702fba4700539c1a2eddac13dd14ecf70)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <sys/mman.h>
14 #include <linux/rtnetlink.h>
15 
16 /* Verbs header. */
17 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
18 #ifdef PEDANTIC
19 #pragma GCC diagnostic ignored "-Wpedantic"
20 #endif
21 #include <infiniband/verbs.h>
22 #ifdef PEDANTIC
23 #pragma GCC diagnostic error "-Wpedantic"
24 #endif
25 
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_ethdev_pci.h>
29 #include <rte_pci.h>
30 #include <rte_bus_pci.h>
31 #include <rte_common.h>
32 #include <rte_kvargs.h>
33 #include <rte_rwlock.h>
34 #include <rte_spinlock.h>
35 #include <rte_string_fns.h>
36 #include <rte_alarm.h>
37 
38 #include <mlx5_glue.h>
39 #include <mlx5_devx_cmds.h>
40 #include <mlx5_common.h>
41 
42 #include "mlx5_defs.h"
43 #include "mlx5.h"
44 #include "mlx5_utils.h"
45 #include "mlx5_rxtx.h"
46 #include "mlx5_autoconf.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49 #include "rte_pmd_mlx5.h"
50 
51 /* Device parameter to enable RX completion queue compression. */
52 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
53 
54 /* Device parameter to enable RX completion entry padding to 128B. */
55 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
56 
57 /* Device parameter to enable padding Rx packet to cacheline size. */
58 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
59 
60 /* Device parameter to enable Multi-Packet Rx queue. */
61 #define MLX5_RX_MPRQ_EN "mprq_en"
62 
63 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
64 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
65 
66 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
67 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
68 
69 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
70 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
71 
72 /* Device parameter to configure inline send. Deprecated, ignored.*/
73 #define MLX5_TXQ_INLINE "txq_inline"
74 
75 /* Device parameter to limit packet size to inline with ordinary SEND. */
76 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
77 
78 /* Device parameter to configure minimal data size to inline. */
79 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
80 
81 /* Device parameter to limit packet size to inline with Enhanced MPW. */
82 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
83 
84 /*
85  * Device parameter to configure the number of TX queues threshold for
86  * enabling inline send.
87  */
88 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
89 
90 /*
91  * Device parameter to configure the number of TX queues threshold for
92  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
93  */
94 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
95 
96 /* Device parameter to enable multi-packet send WQEs. */
97 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
98 
99 /*
100  * Device parameter to force doorbell register mapping
101  * to non-cahed region eliminating the extra write memory barrier.
102  */
103 #define MLX5_TX_DB_NC "tx_db_nc"
104 
105 /*
106  * Device parameter to include 2 dsegs in the title WQEBB.
107  * Deprecated, ignored.
108  */
109 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
110 
111 /*
112  * Device parameter to limit the size of inlining packet.
113  * Deprecated, ignored.
114  */
115 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
116 
117 /*
118  * Device parameter to enable hardware Tx vector.
119  * Deprecated, ignored (no vectorized Tx routines anymore).
120  */
121 #define MLX5_TX_VEC_EN "tx_vec_en"
122 
123 /* Device parameter to enable hardware Rx vector. */
124 #define MLX5_RX_VEC_EN "rx_vec_en"
125 
126 /* Allow L3 VXLAN flow creation. */
127 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
128 
129 /* Activate DV E-Switch flow steering. */
130 #define MLX5_DV_ESW_EN "dv_esw_en"
131 
132 /* Activate DV flow steering. */
133 #define MLX5_DV_FLOW_EN "dv_flow_en"
134 
135 /* Enable extensive flow metadata support. */
136 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
137 
138 /* Activate Netlink support in VF mode. */
139 #define MLX5_VF_NL_EN "vf_nl_en"
140 
141 /* Enable extending memsegs when creating a MR. */
142 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
143 
144 /* Select port representors to instantiate. */
145 #define MLX5_REPRESENTOR "representor"
146 
147 /* Device parameter to configure the maximum number of dump files per queue. */
148 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
149 
150 /* Configure timeout of LRO session (in microseconds). */
151 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
152 
153 #ifndef HAVE_IBV_MLX5_MOD_MPW
154 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
155 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
156 #endif
157 
158 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
159 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
160 #endif
161 
162 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
163 
164 /* Shared memory between primary and secondary processes. */
165 struct mlx5_shared_data *mlx5_shared_data;
166 
167 /* Spinlock for mlx5_shared_data allocation. */
168 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
169 
170 /* Process local data for secondary processes. */
171 static struct mlx5_local_data mlx5_local_data;
172 
173 /** Driver-specific log messages type. */
174 int mlx5_logtype;
175 
176 /** Data associated with devices to spawn. */
177 struct mlx5_dev_spawn_data {
178 	uint32_t ifindex; /**< Network interface index. */
179 	uint32_t max_port; /**< IB device maximal port index. */
180 	uint32_t ibv_port; /**< IB device physical port index. */
181 	int pf_bond; /**< bonding device PF index. < 0 - no bonding */
182 	struct mlx5_switch_info info; /**< Switch information. */
183 	struct ibv_device *ibv_dev; /**< Associated IB device. */
184 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
185 	struct rte_pci_device *pci_dev; /**< Backend PCI device. */
186 };
187 
188 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
189 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
190 
191 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
192 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
193 
194 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
195 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
196 
197 /**
198  * Allocate ID pool structure.
199  *
200  * @param[in] max_id
201  *   The maximum id can be allocated from the pool.
202  *
203  * @return
204  *   Pointer to pool object, NULL value otherwise.
205  */
206 struct mlx5_flow_id_pool *
207 mlx5_flow_id_pool_alloc(uint32_t max_id)
208 {
209 	struct mlx5_flow_id_pool *pool;
210 	void *mem;
211 
212 	pool = rte_zmalloc("id pool allocation", sizeof(*pool),
213 			   RTE_CACHE_LINE_SIZE);
214 	if (!pool) {
215 		DRV_LOG(ERR, "can't allocate id pool");
216 		rte_errno  = ENOMEM;
217 		return NULL;
218 	}
219 	mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
220 			  RTE_CACHE_LINE_SIZE);
221 	if (!mem) {
222 		DRV_LOG(ERR, "can't allocate mem for id pool");
223 		rte_errno  = ENOMEM;
224 		goto error;
225 	}
226 	pool->free_arr = mem;
227 	pool->curr = pool->free_arr;
228 	pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
229 	pool->base_index = 0;
230 	pool->max_id = max_id;
231 	return pool;
232 error:
233 	rte_free(pool);
234 	return NULL;
235 }
236 
237 /**
238  * Release ID pool structure.
239  *
240  * @param[in] pool
241  *   Pointer to flow id pool object to free.
242  */
243 void
244 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
245 {
246 	rte_free(pool->free_arr);
247 	rte_free(pool);
248 }
249 
250 /**
251  * Generate ID.
252  *
253  * @param[in] pool
254  *   Pointer to flow id pool.
255  * @param[out] id
256  *   The generated ID.
257  *
258  * @return
259  *   0 on success, error value otherwise.
260  */
261 uint32_t
262 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
263 {
264 	if (pool->curr == pool->free_arr) {
265 		if (pool->base_index == pool->max_id) {
266 			rte_errno  = ENOMEM;
267 			DRV_LOG(ERR, "no free id");
268 			return -rte_errno;
269 		}
270 		*id = ++pool->base_index;
271 		return 0;
272 	}
273 	*id = *(--pool->curr);
274 	return 0;
275 }
276 
277 /**
278  * Release ID.
279  *
280  * @param[in] pool
281  *   Pointer to flow id pool.
282  * @param[out] id
283  *   The generated ID.
284  *
285  * @return
286  *   0 on success, error value otherwise.
287  */
288 uint32_t
289 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
290 {
291 	uint32_t size;
292 	uint32_t size2;
293 	void *mem;
294 
295 	if (pool->curr == pool->last) {
296 		size = pool->curr - pool->free_arr;
297 		size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
298 		MLX5_ASSERT(size2 > size);
299 		mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
300 		if (!mem) {
301 			DRV_LOG(ERR, "can't allocate mem for id pool");
302 			rte_errno  = ENOMEM;
303 			return -rte_errno;
304 		}
305 		memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
306 		rte_free(pool->free_arr);
307 		pool->free_arr = mem;
308 		pool->curr = pool->free_arr + size;
309 		pool->last = pool->free_arr + size2;
310 	}
311 	*pool->curr = id;
312 	pool->curr++;
313 	return 0;
314 }
315 
316 /**
317  * Initialize the counters management structure.
318  *
319  * @param[in] sh
320  *   Pointer to mlx5_ibv_shared object to free
321  */
322 static void
323 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
324 {
325 	uint8_t i;
326 
327 	TAILQ_INIT(&sh->cmng.flow_counters);
328 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
329 		TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
330 }
331 
332 /**
333  * Destroy all the resources allocated for a counter memory management.
334  *
335  * @param[in] mng
336  *   Pointer to the memory management structure.
337  */
338 static void
339 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
340 {
341 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
342 
343 	LIST_REMOVE(mng, next);
344 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
345 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
346 	rte_free(mem);
347 }
348 
349 /**
350  * Close and release all the resources of the counters management.
351  *
352  * @param[in] sh
353  *   Pointer to mlx5_ibv_shared object to free.
354  */
355 static void
356 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
357 {
358 	struct mlx5_counter_stats_mem_mng *mng;
359 	uint8_t i;
360 	int j;
361 	int retries = 1024;
362 
363 	rte_errno = 0;
364 	while (--retries) {
365 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
366 		if (rte_errno != EINPROGRESS)
367 			break;
368 		rte_pause();
369 	}
370 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
371 		struct mlx5_flow_counter_pool *pool;
372 		uint32_t batch = !!(i % 2);
373 
374 		if (!sh->cmng.ccont[i].pools)
375 			continue;
376 		pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
377 		while (pool) {
378 			if (batch) {
379 				if (pool->min_dcs)
380 					claim_zero
381 					(mlx5_devx_cmd_destroy(pool->min_dcs));
382 			}
383 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
384 				if (pool->counters_raw[j].action)
385 					claim_zero
386 					(mlx5_glue->destroy_flow_action
387 					       (pool->counters_raw[j].action));
388 				if (!batch && pool->counters_raw[j].dcs)
389 					claim_zero(mlx5_devx_cmd_destroy
390 						  (pool->counters_raw[j].dcs));
391 			}
392 			TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
393 				     next);
394 			rte_free(pool);
395 			pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
396 		}
397 		rte_free(sh->cmng.ccont[i].pools);
398 	}
399 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
400 	while (mng) {
401 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
402 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
403 	}
404 	memset(&sh->cmng, 0, sizeof(sh->cmng));
405 }
406 
407 /**
408  * Extract pdn of PD object using DV API.
409  *
410  * @param[in] pd
411  *   Pointer to the verbs PD object.
412  * @param[out] pdn
413  *   Pointer to the PD object number variable.
414  *
415  * @return
416  *   0 on success, error value otherwise.
417  */
418 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
419 static int
420 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
421 {
422 	struct mlx5dv_obj obj;
423 	struct mlx5dv_pd pd_info;
424 	int ret = 0;
425 
426 	obj.pd.in = pd;
427 	obj.pd.out = &pd_info;
428 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
429 	if (ret) {
430 		DRV_LOG(DEBUG, "Fail to get PD object info");
431 		return ret;
432 	}
433 	*pdn = pd_info.pdn;
434 	return 0;
435 }
436 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
437 
438 static int
439 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
440 {
441 	char *env;
442 	int value;
443 
444 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
445 	/* Get environment variable to store. */
446 	env = getenv(MLX5_SHUT_UP_BF);
447 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
448 	if (config->dbnc == MLX5_ARG_UNSET)
449 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
450 	else
451 		setenv(MLX5_SHUT_UP_BF,
452 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
453 	return value;
454 }
455 
456 static void
457 mlx5_restore_doorbell_mapping_env(int value)
458 {
459 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
460 	/* Restore the original environment variable state. */
461 	if (value == MLX5_ARG_UNSET)
462 		unsetenv(MLX5_SHUT_UP_BF);
463 	else
464 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
465 }
466 
467 /**
468  * Allocate shared IB device context. If there is multiport device the
469  * master and representors will share this context, if there is single
470  * port dedicated IB device, the context will be used by only given
471  * port due to unification.
472  *
473  * Routine first searches the context for the specified IB device name,
474  * if found the shared context assumed and reference counter is incremented.
475  * If no context found the new one is created and initialized with specified
476  * IB device context and parameters.
477  *
478  * @param[in] spawn
479  *   Pointer to the IB device attributes (name, port, etc).
480  * @param[in] config
481  *   Pointer to device configuration structure.
482  *
483  * @return
484  *   Pointer to mlx5_ibv_shared object on success,
485  *   otherwise NULL and rte_errno is set.
486  */
487 static struct mlx5_ibv_shared *
488 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
489 			const struct mlx5_dev_config *config)
490 {
491 	struct mlx5_ibv_shared *sh;
492 	int dbmap_env;
493 	int err = 0;
494 	uint32_t i;
495 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
496 	struct mlx5_devx_tis_attr tis_attr = { 0 };
497 #endif
498 
499 	MLX5_ASSERT(spawn);
500 	/* Secondary process should not create the shared context. */
501 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
502 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
503 	/* Search for IB context by device name. */
504 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
505 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
506 			sh->refcnt++;
507 			goto exit;
508 		}
509 	}
510 	/* No device found, we have to create new shared context. */
511 	MLX5_ASSERT(spawn->max_port);
512 	sh = rte_zmalloc("ethdev shared ib context",
513 			 sizeof(struct mlx5_ibv_shared) +
514 			 spawn->max_port *
515 			 sizeof(struct mlx5_ibv_shared_port),
516 			 RTE_CACHE_LINE_SIZE);
517 	if (!sh) {
518 		DRV_LOG(ERR, "shared context allocation failure");
519 		rte_errno  = ENOMEM;
520 		goto exit;
521 	}
522 	/*
523 	 * Configure environment variable "MLX5_BF_SHUT_UP"
524 	 * before the device creation. The rdma_core library
525 	 * checks the variable at device creation and
526 	 * stores the result internally.
527 	 */
528 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
529 	/* Try to open IB device with DV first, then usual Verbs. */
530 	errno = 0;
531 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
532 	if (sh->ctx) {
533 		sh->devx = 1;
534 		DRV_LOG(DEBUG, "DevX is supported");
535 		/* The device is created, no need for environment. */
536 		mlx5_restore_doorbell_mapping_env(dbmap_env);
537 	} else {
538 		/* The environment variable is still configured. */
539 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
540 		err = errno ? errno : ENODEV;
541 		/*
542 		 * The environment variable is not needed anymore,
543 		 * all device creation attempts are completed.
544 		 */
545 		mlx5_restore_doorbell_mapping_env(dbmap_env);
546 		if (!sh->ctx)
547 			goto error;
548 		DRV_LOG(DEBUG, "DevX is NOT supported");
549 	}
550 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
551 	if (err) {
552 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
553 		goto error;
554 	}
555 	sh->refcnt = 1;
556 	sh->max_port = spawn->max_port;
557 	strncpy(sh->ibdev_name, sh->ctx->device->name,
558 		sizeof(sh->ibdev_name));
559 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
560 		sizeof(sh->ibdev_path));
561 	pthread_mutex_init(&sh->intr_mutex, NULL);
562 	/*
563 	 * Setting port_id to max unallowed value means
564 	 * there is no interrupt subhandler installed for
565 	 * the given port index i.
566 	 */
567 	for (i = 0; i < sh->max_port; i++) {
568 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
569 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
570 	}
571 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
572 	if (sh->pd == NULL) {
573 		DRV_LOG(ERR, "PD allocation failure");
574 		err = ENOMEM;
575 		goto error;
576 	}
577 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
578 	if (sh->devx) {
579 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
580 		if (err) {
581 			DRV_LOG(ERR, "Fail to extract pdn from PD");
582 			goto error;
583 		}
584 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
585 		if (!sh->td) {
586 			DRV_LOG(ERR, "TD allocation failure");
587 			err = ENOMEM;
588 			goto error;
589 		}
590 		tis_attr.transport_domain = sh->td->id;
591 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
592 		if (!sh->tis) {
593 			DRV_LOG(ERR, "TIS allocation failure");
594 			err = ENOMEM;
595 			goto error;
596 		}
597 	}
598 	sh->flow_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX);
599 	if (!sh->flow_id_pool) {
600 		DRV_LOG(ERR, "can't create flow id pool");
601 		err = ENOMEM;
602 		goto error;
603 	}
604 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
605 	/*
606 	 * Once the device is added to the list of memory event
607 	 * callback, its global MR cache table cannot be expanded
608 	 * on the fly because of deadlock. If it overflows, lookup
609 	 * should be done by searching MR list linearly, which is slow.
610 	 *
611 	 * At this point the device is not added to the memory
612 	 * event list yet, context is just being created.
613 	 */
614 	err = mlx5_mr_btree_init(&sh->mr.cache,
615 				 MLX5_MR_BTREE_CACHE_N * 2,
616 				 spawn->pci_dev->device.numa_node);
617 	if (err) {
618 		err = rte_errno;
619 		goto error;
620 	}
621 	mlx5_flow_counters_mng_init(sh);
622 	/* Add device to memory callback list. */
623 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
624 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
625 			 sh, mem_event_cb);
626 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
627 	/* Add context to the global device list. */
628 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
629 exit:
630 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
631 	return sh;
632 error:
633 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
634 	MLX5_ASSERT(sh);
635 	if (sh->tis)
636 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
637 	if (sh->td)
638 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
639 	if (sh->pd)
640 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
641 	if (sh->ctx)
642 		claim_zero(mlx5_glue->close_device(sh->ctx));
643 	if (sh->flow_id_pool)
644 		mlx5_flow_id_pool_release(sh->flow_id_pool);
645 	rte_free(sh);
646 	MLX5_ASSERT(err > 0);
647 	rte_errno = err;
648 	return NULL;
649 }
650 
651 /**
652  * Free shared IB device context. Decrement counter and if zero free
653  * all allocated resources and close handles.
654  *
655  * @param[in] sh
656  *   Pointer to mlx5_ibv_shared object to free
657  */
658 static void
659 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
660 {
661 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
662 #ifdef RTE_LIBRTE_MLX5_DEBUG
663 	/* Check the object presence in the list. */
664 	struct mlx5_ibv_shared *lctx;
665 
666 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
667 		if (lctx == sh)
668 			break;
669 	MLX5_ASSERT(lctx);
670 	if (lctx != sh) {
671 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
672 		goto exit;
673 	}
674 #endif
675 	MLX5_ASSERT(sh);
676 	MLX5_ASSERT(sh->refcnt);
677 	/* Secondary process should not free the shared context. */
678 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
679 	if (--sh->refcnt)
680 		goto exit;
681 	/* Remove from memory callback device list. */
682 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
683 	LIST_REMOVE(sh, mem_event_cb);
684 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
685 	/* Release created Memory Regions. */
686 	mlx5_mr_release(sh);
687 	/* Remove context from the global device list. */
688 	LIST_REMOVE(sh, next);
689 	/*
690 	 *  Ensure there is no async event handler installed.
691 	 *  Only primary process handles async device events.
692 	 **/
693 	mlx5_flow_counters_mng_close(sh);
694 	MLX5_ASSERT(!sh->intr_cnt);
695 	if (sh->intr_cnt)
696 		mlx5_intr_callback_unregister
697 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
698 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
699 	if (sh->devx_intr_cnt) {
700 		if (sh->intr_handle_devx.fd)
701 			rte_intr_callback_unregister(&sh->intr_handle_devx,
702 					  mlx5_dev_interrupt_handler_devx, sh);
703 		if (sh->devx_comp)
704 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
705 	}
706 #endif
707 	pthread_mutex_destroy(&sh->intr_mutex);
708 	if (sh->pd)
709 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
710 	if (sh->tis)
711 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
712 	if (sh->td)
713 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
714 	if (sh->ctx)
715 		claim_zero(mlx5_glue->close_device(sh->ctx));
716 	if (sh->flow_id_pool)
717 		mlx5_flow_id_pool_release(sh->flow_id_pool);
718 	rte_free(sh);
719 exit:
720 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
721 }
722 
723 /**
724  * Destroy table hash list and all the root entries per domain.
725  *
726  * @param[in] priv
727  *   Pointer to the private device data structure.
728  */
729 static void
730 mlx5_free_table_hash_list(struct mlx5_priv *priv)
731 {
732 	struct mlx5_ibv_shared *sh = priv->sh;
733 	struct mlx5_flow_tbl_data_entry *tbl_data;
734 	union mlx5_flow_tbl_key table_key = {
735 		{
736 			.table_id = 0,
737 			.reserved = 0,
738 			.domain = 0,
739 			.direction = 0,
740 		}
741 	};
742 	struct mlx5_hlist_entry *pos;
743 
744 	if (!sh->flow_tbls)
745 		return;
746 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
747 	if (pos) {
748 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
749 					entry);
750 		MLX5_ASSERT(tbl_data);
751 		mlx5_hlist_remove(sh->flow_tbls, pos);
752 		rte_free(tbl_data);
753 	}
754 	table_key.direction = 1;
755 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
756 	if (pos) {
757 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
758 					entry);
759 		MLX5_ASSERT(tbl_data);
760 		mlx5_hlist_remove(sh->flow_tbls, pos);
761 		rte_free(tbl_data);
762 	}
763 	table_key.direction = 0;
764 	table_key.domain = 1;
765 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
766 	if (pos) {
767 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
768 					entry);
769 		MLX5_ASSERT(tbl_data);
770 		mlx5_hlist_remove(sh->flow_tbls, pos);
771 		rte_free(tbl_data);
772 	}
773 	mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
774 }
775 
776 /**
777  * Initialize flow table hash list and create the root tables entry
778  * for each domain.
779  *
780  * @param[in] priv
781  *   Pointer to the private device data structure.
782  *
783  * @return
784  *   Zero on success, positive error code otherwise.
785  */
786 static int
787 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
788 {
789 	struct mlx5_ibv_shared *sh = priv->sh;
790 	char s[MLX5_HLIST_NAMESIZE];
791 	int err = 0;
792 
793 	MLX5_ASSERT(sh);
794 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
795 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
796 	if (!sh->flow_tbls) {
797 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
798 		err = ENOMEM;
799 		return err;
800 	}
801 #ifndef HAVE_MLX5DV_DR
802 	/*
803 	 * In case we have not DR support, the zero tables should be created
804 	 * because DV expect to see them even if they cannot be created by
805 	 * RDMA-CORE.
806 	 */
807 	union mlx5_flow_tbl_key table_key = {
808 		{
809 			.table_id = 0,
810 			.reserved = 0,
811 			.domain = 0,
812 			.direction = 0,
813 		}
814 	};
815 	struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
816 							  sizeof(*tbl_data), 0);
817 
818 	if (!tbl_data) {
819 		err = ENOMEM;
820 		goto error;
821 	}
822 	tbl_data->entry.key = table_key.v64;
823 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
824 	if (err)
825 		goto error;
826 	rte_atomic32_init(&tbl_data->tbl.refcnt);
827 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
828 	table_key.direction = 1;
829 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
830 	if (!tbl_data) {
831 		err = ENOMEM;
832 		goto error;
833 	}
834 	tbl_data->entry.key = table_key.v64;
835 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
836 	if (err)
837 		goto error;
838 	rte_atomic32_init(&tbl_data->tbl.refcnt);
839 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
840 	table_key.direction = 0;
841 	table_key.domain = 1;
842 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
843 	if (!tbl_data) {
844 		err = ENOMEM;
845 		goto error;
846 	}
847 	tbl_data->entry.key = table_key.v64;
848 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
849 	if (err)
850 		goto error;
851 	rte_atomic32_init(&tbl_data->tbl.refcnt);
852 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
853 	return err;
854 error:
855 	mlx5_free_table_hash_list(priv);
856 #endif /* HAVE_MLX5DV_DR */
857 	return err;
858 }
859 
860 /**
861  * Initialize DR related data within private structure.
862  * Routine checks the reference counter and does actual
863  * resources creation/initialization only if counter is zero.
864  *
865  * @param[in] priv
866  *   Pointer to the private device data structure.
867  *
868  * @return
869  *   Zero on success, positive error code otherwise.
870  */
871 static int
872 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
873 {
874 	struct mlx5_ibv_shared *sh = priv->sh;
875 	char s[MLX5_HLIST_NAMESIZE];
876 	int err = 0;
877 
878 	if (!sh->flow_tbls)
879 		err = mlx5_alloc_table_hash_list(priv);
880 	else
881 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
882 			(void *)sh->flow_tbls);
883 	if (err)
884 		return err;
885 	/* Create tags hash list table. */
886 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
887 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
888 	if (!sh->tag_table) {
889 		DRV_LOG(ERR, "tags with hash creation failed.\n");
890 		err = ENOMEM;
891 		goto error;
892 	}
893 #ifdef HAVE_MLX5DV_DR
894 	void *domain;
895 
896 	if (sh->dv_refcnt) {
897 		/* Shared DV/DR structures is already initialized. */
898 		sh->dv_refcnt++;
899 		priv->dr_shared = 1;
900 		return 0;
901 	}
902 	/* Reference counter is zero, we should initialize structures. */
903 	domain = mlx5_glue->dr_create_domain(sh->ctx,
904 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
905 	if (!domain) {
906 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
907 		err = errno;
908 		goto error;
909 	}
910 	sh->rx_domain = domain;
911 	domain = mlx5_glue->dr_create_domain(sh->ctx,
912 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
913 	if (!domain) {
914 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
915 		err = errno;
916 		goto error;
917 	}
918 	pthread_mutex_init(&sh->dv_mutex, NULL);
919 	sh->tx_domain = domain;
920 #ifdef HAVE_MLX5DV_DR_ESWITCH
921 	if (priv->config.dv_esw_en) {
922 		domain  = mlx5_glue->dr_create_domain
923 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
924 		if (!domain) {
925 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
926 			err = errno;
927 			goto error;
928 		}
929 		sh->fdb_domain = domain;
930 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
931 	}
932 #endif
933 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
934 #endif /* HAVE_MLX5DV_DR */
935 	sh->dv_refcnt++;
936 	priv->dr_shared = 1;
937 	return 0;
938 error:
939 	/* Rollback the created objects. */
940 	if (sh->rx_domain) {
941 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
942 		sh->rx_domain = NULL;
943 	}
944 	if (sh->tx_domain) {
945 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
946 		sh->tx_domain = NULL;
947 	}
948 	if (sh->fdb_domain) {
949 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
950 		sh->fdb_domain = NULL;
951 	}
952 	if (sh->esw_drop_action) {
953 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
954 		sh->esw_drop_action = NULL;
955 	}
956 	if (sh->pop_vlan_action) {
957 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
958 		sh->pop_vlan_action = NULL;
959 	}
960 	if (sh->tag_table) {
961 		/* tags should be destroyed with flow before. */
962 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
963 		sh->tag_table = NULL;
964 	}
965 	mlx5_free_table_hash_list(priv);
966 	return err;
967 }
968 
969 /**
970  * Destroy DR related data within private structure.
971  *
972  * @param[in] priv
973  *   Pointer to the private device data structure.
974  */
975 static void
976 mlx5_free_shared_dr(struct mlx5_priv *priv)
977 {
978 	struct mlx5_ibv_shared *sh;
979 
980 	if (!priv->dr_shared)
981 		return;
982 	priv->dr_shared = 0;
983 	sh = priv->sh;
984 	MLX5_ASSERT(sh);
985 #ifdef HAVE_MLX5DV_DR
986 	MLX5_ASSERT(sh->dv_refcnt);
987 	if (sh->dv_refcnt && --sh->dv_refcnt)
988 		return;
989 	if (sh->rx_domain) {
990 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
991 		sh->rx_domain = NULL;
992 	}
993 	if (sh->tx_domain) {
994 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
995 		sh->tx_domain = NULL;
996 	}
997 #ifdef HAVE_MLX5DV_DR_ESWITCH
998 	if (sh->fdb_domain) {
999 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1000 		sh->fdb_domain = NULL;
1001 	}
1002 	if (sh->esw_drop_action) {
1003 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1004 		sh->esw_drop_action = NULL;
1005 	}
1006 #endif
1007 	if (sh->pop_vlan_action) {
1008 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1009 		sh->pop_vlan_action = NULL;
1010 	}
1011 	pthread_mutex_destroy(&sh->dv_mutex);
1012 #endif /* HAVE_MLX5DV_DR */
1013 	if (sh->tag_table) {
1014 		/* tags should be destroyed with flow before. */
1015 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1016 		sh->tag_table = NULL;
1017 	}
1018 	mlx5_free_table_hash_list(priv);
1019 }
1020 
1021 /**
1022  * Initialize shared data between primary and secondary process.
1023  *
1024  * A memzone is reserved by primary process and secondary processes attach to
1025  * the memzone.
1026  *
1027  * @return
1028  *   0 on success, a negative errno value otherwise and rte_errno is set.
1029  */
1030 static int
1031 mlx5_init_shared_data(void)
1032 {
1033 	const struct rte_memzone *mz;
1034 	int ret = 0;
1035 
1036 	rte_spinlock_lock(&mlx5_shared_data_lock);
1037 	if (mlx5_shared_data == NULL) {
1038 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1039 			/* Allocate shared memory. */
1040 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1041 						 sizeof(*mlx5_shared_data),
1042 						 SOCKET_ID_ANY, 0);
1043 			if (mz == NULL) {
1044 				DRV_LOG(ERR,
1045 					"Cannot allocate mlx5 shared data");
1046 				ret = -rte_errno;
1047 				goto error;
1048 			}
1049 			mlx5_shared_data = mz->addr;
1050 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1051 			rte_spinlock_init(&mlx5_shared_data->lock);
1052 		} else {
1053 			/* Lookup allocated shared memory. */
1054 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1055 			if (mz == NULL) {
1056 				DRV_LOG(ERR,
1057 					"Cannot attach mlx5 shared data");
1058 				ret = -rte_errno;
1059 				goto error;
1060 			}
1061 			mlx5_shared_data = mz->addr;
1062 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1063 		}
1064 	}
1065 error:
1066 	rte_spinlock_unlock(&mlx5_shared_data_lock);
1067 	return ret;
1068 }
1069 
1070 /**
1071  * Retrieve integer value from environment variable.
1072  *
1073  * @param[in] name
1074  *   Environment variable name.
1075  *
1076  * @return
1077  *   Integer value, 0 if the variable is not set.
1078  */
1079 int
1080 mlx5_getenv_int(const char *name)
1081 {
1082 	const char *val = getenv(name);
1083 
1084 	if (val == NULL)
1085 		return 0;
1086 	return atoi(val);
1087 }
1088 
1089 /**
1090  * Verbs callback to allocate a memory. This function should allocate the space
1091  * according to the size provided residing inside a huge page.
1092  * Please note that all allocation must respect the alignment from libmlx5
1093  * (i.e. currently sysconf(_SC_PAGESIZE)).
1094  *
1095  * @param[in] size
1096  *   The size in bytes of the memory to allocate.
1097  * @param[in] data
1098  *   A pointer to the callback data.
1099  *
1100  * @return
1101  *   Allocated buffer, NULL otherwise and rte_errno is set.
1102  */
1103 static void *
1104 mlx5_alloc_verbs_buf(size_t size, void *data)
1105 {
1106 	struct mlx5_priv *priv = data;
1107 	void *ret;
1108 	size_t alignment = sysconf(_SC_PAGESIZE);
1109 	unsigned int socket = SOCKET_ID_ANY;
1110 
1111 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1112 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1113 
1114 		socket = ctrl->socket;
1115 	} else if (priv->verbs_alloc_ctx.type ==
1116 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1117 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1118 
1119 		socket = ctrl->socket;
1120 	}
1121 	MLX5_ASSERT(data != NULL);
1122 	ret = rte_malloc_socket(__func__, size, alignment, socket);
1123 	if (!ret && size)
1124 		rte_errno = ENOMEM;
1125 	return ret;
1126 }
1127 
1128 /**
1129  * Verbs callback to free a memory.
1130  *
1131  * @param[in] ptr
1132  *   A pointer to the memory to free.
1133  * @param[in] data
1134  *   A pointer to the callback data.
1135  */
1136 static void
1137 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1138 {
1139 	MLX5_ASSERT(data != NULL);
1140 	rte_free(ptr);
1141 }
1142 
1143 /**
1144  * DPDK callback to add udp tunnel port
1145  *
1146  * @param[in] dev
1147  *   A pointer to eth_dev
1148  * @param[in] udp_tunnel
1149  *   A pointer to udp tunnel
1150  *
1151  * @return
1152  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1153  */
1154 int
1155 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1156 			 struct rte_eth_udp_tunnel *udp_tunnel)
1157 {
1158 	MLX5_ASSERT(udp_tunnel != NULL);
1159 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1160 	    udp_tunnel->udp_port == 4789)
1161 		return 0;
1162 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1163 	    udp_tunnel->udp_port == 4790)
1164 		return 0;
1165 	return -ENOTSUP;
1166 }
1167 
1168 /**
1169  * Initialize process private data structure.
1170  *
1171  * @param dev
1172  *   Pointer to Ethernet device structure.
1173  *
1174  * @return
1175  *   0 on success, a negative errno value otherwise and rte_errno is set.
1176  */
1177 int
1178 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1179 {
1180 	struct mlx5_priv *priv = dev->data->dev_private;
1181 	struct mlx5_proc_priv *ppriv;
1182 	size_t ppriv_size;
1183 
1184 	/*
1185 	 * UAR register table follows the process private structure. BlueFlame
1186 	 * registers for Tx queues are stored in the table.
1187 	 */
1188 	ppriv_size =
1189 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1190 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1191 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1192 	if (!ppriv) {
1193 		rte_errno = ENOMEM;
1194 		return -rte_errno;
1195 	}
1196 	ppriv->uar_table_sz = ppriv_size;
1197 	dev->process_private = ppriv;
1198 	return 0;
1199 }
1200 
1201 /**
1202  * Un-initialize process private data structure.
1203  *
1204  * @param dev
1205  *   Pointer to Ethernet device structure.
1206  */
1207 static void
1208 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1209 {
1210 	if (!dev->process_private)
1211 		return;
1212 	rte_free(dev->process_private);
1213 	dev->process_private = NULL;
1214 }
1215 
1216 /**
1217  * DPDK callback to close the device.
1218  *
1219  * Destroy all queues and objects, free memory.
1220  *
1221  * @param dev
1222  *   Pointer to Ethernet device structure.
1223  */
1224 static void
1225 mlx5_dev_close(struct rte_eth_dev *dev)
1226 {
1227 	struct mlx5_priv *priv = dev->data->dev_private;
1228 	unsigned int i;
1229 	int ret;
1230 
1231 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1232 		dev->data->port_id,
1233 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1234 	/* In case mlx5_dev_stop() has not been called. */
1235 	mlx5_dev_interrupt_handler_uninstall(dev);
1236 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1237 	mlx5_traffic_disable(dev);
1238 	mlx5_flow_flush(dev, NULL);
1239 	mlx5_flow_meter_flush(dev, NULL);
1240 	/* Prevent crashes when queues are still in use. */
1241 	dev->rx_pkt_burst = removed_rx_burst;
1242 	dev->tx_pkt_burst = removed_tx_burst;
1243 	rte_wmb();
1244 	/* Disable datapath on secondary process. */
1245 	mlx5_mp_req_stop_rxtx(dev);
1246 	if (priv->rxqs != NULL) {
1247 		/* XXX race condition if mlx5_rx_burst() is still running. */
1248 		usleep(1000);
1249 		for (i = 0; (i != priv->rxqs_n); ++i)
1250 			mlx5_rxq_release(dev, i);
1251 		priv->rxqs_n = 0;
1252 		priv->rxqs = NULL;
1253 	}
1254 	if (priv->txqs != NULL) {
1255 		/* XXX race condition if mlx5_tx_burst() is still running. */
1256 		usleep(1000);
1257 		for (i = 0; (i != priv->txqs_n); ++i)
1258 			mlx5_txq_release(dev, i);
1259 		priv->txqs_n = 0;
1260 		priv->txqs = NULL;
1261 	}
1262 	mlx5_proc_priv_uninit(dev);
1263 	if (priv->mreg_cp_tbl)
1264 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1265 	mlx5_mprq_free_mp(dev);
1266 	mlx5_free_shared_dr(priv);
1267 	if (priv->rss_conf.rss_key != NULL)
1268 		rte_free(priv->rss_conf.rss_key);
1269 	if (priv->reta_idx != NULL)
1270 		rte_free(priv->reta_idx);
1271 	if (priv->config.vf)
1272 		mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1273 				       dev->data->mac_addrs,
1274 				       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1275 	if (priv->nl_socket_route >= 0)
1276 		close(priv->nl_socket_route);
1277 	if (priv->nl_socket_rdma >= 0)
1278 		close(priv->nl_socket_rdma);
1279 	if (priv->vmwa_context)
1280 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1281 	if (priv->sh) {
1282 		/*
1283 		 * Free the shared context in last turn, because the cleanup
1284 		 * routines above may use some shared fields, like
1285 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1286 		 * ifindex if Netlink fails.
1287 		 */
1288 		mlx5_free_shared_ibctx(priv->sh);
1289 		priv->sh = NULL;
1290 	}
1291 	ret = mlx5_hrxq_verify(dev);
1292 	if (ret)
1293 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1294 			dev->data->port_id);
1295 	ret = mlx5_ind_table_obj_verify(dev);
1296 	if (ret)
1297 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1298 			dev->data->port_id);
1299 	ret = mlx5_rxq_obj_verify(dev);
1300 	if (ret)
1301 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1302 			dev->data->port_id);
1303 	ret = mlx5_rxq_verify(dev);
1304 	if (ret)
1305 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1306 			dev->data->port_id);
1307 	ret = mlx5_txq_obj_verify(dev);
1308 	if (ret)
1309 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1310 			dev->data->port_id);
1311 	ret = mlx5_txq_verify(dev);
1312 	if (ret)
1313 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1314 			dev->data->port_id);
1315 	ret = mlx5_flow_verify(dev);
1316 	if (ret)
1317 		DRV_LOG(WARNING, "port %u some flows still remain",
1318 			dev->data->port_id);
1319 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1320 		unsigned int c = 0;
1321 		uint16_t port_id;
1322 
1323 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1324 			struct mlx5_priv *opriv =
1325 				rte_eth_devices[port_id].data->dev_private;
1326 
1327 			if (!opriv ||
1328 			    opriv->domain_id != priv->domain_id ||
1329 			    &rte_eth_devices[port_id] == dev)
1330 				continue;
1331 			++c;
1332 			break;
1333 		}
1334 		if (!c)
1335 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1336 	}
1337 	memset(priv, 0, sizeof(*priv));
1338 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1339 	/*
1340 	 * Reset mac_addrs to NULL such that it is not freed as part of
1341 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1342 	 * it is freed when dev_private is freed.
1343 	 */
1344 	dev->data->mac_addrs = NULL;
1345 }
1346 
1347 const struct eth_dev_ops mlx5_dev_ops = {
1348 	.dev_configure = mlx5_dev_configure,
1349 	.dev_start = mlx5_dev_start,
1350 	.dev_stop = mlx5_dev_stop,
1351 	.dev_set_link_down = mlx5_set_link_down,
1352 	.dev_set_link_up = mlx5_set_link_up,
1353 	.dev_close = mlx5_dev_close,
1354 	.promiscuous_enable = mlx5_promiscuous_enable,
1355 	.promiscuous_disable = mlx5_promiscuous_disable,
1356 	.allmulticast_enable = mlx5_allmulticast_enable,
1357 	.allmulticast_disable = mlx5_allmulticast_disable,
1358 	.link_update = mlx5_link_update,
1359 	.stats_get = mlx5_stats_get,
1360 	.stats_reset = mlx5_stats_reset,
1361 	.xstats_get = mlx5_xstats_get,
1362 	.xstats_reset = mlx5_xstats_reset,
1363 	.xstats_get_names = mlx5_xstats_get_names,
1364 	.fw_version_get = mlx5_fw_version_get,
1365 	.dev_infos_get = mlx5_dev_infos_get,
1366 	.read_clock = mlx5_read_clock,
1367 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1368 	.vlan_filter_set = mlx5_vlan_filter_set,
1369 	.rx_queue_setup = mlx5_rx_queue_setup,
1370 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1371 	.tx_queue_setup = mlx5_tx_queue_setup,
1372 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1373 	.rx_queue_release = mlx5_rx_queue_release,
1374 	.tx_queue_release = mlx5_tx_queue_release,
1375 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1376 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1377 	.mac_addr_remove = mlx5_mac_addr_remove,
1378 	.mac_addr_add = mlx5_mac_addr_add,
1379 	.mac_addr_set = mlx5_mac_addr_set,
1380 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1381 	.mtu_set = mlx5_dev_set_mtu,
1382 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1383 	.vlan_offload_set = mlx5_vlan_offload_set,
1384 	.reta_update = mlx5_dev_rss_reta_update,
1385 	.reta_query = mlx5_dev_rss_reta_query,
1386 	.rss_hash_update = mlx5_rss_hash_update,
1387 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1388 	.filter_ctrl = mlx5_dev_filter_ctrl,
1389 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1390 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1391 	.rxq_info_get = mlx5_rxq_info_get,
1392 	.txq_info_get = mlx5_txq_info_get,
1393 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1394 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1395 	.rx_queue_count = mlx5_rx_queue_count,
1396 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1397 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1398 	.is_removed = mlx5_is_removed,
1399 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1400 	.get_module_info = mlx5_get_module_info,
1401 	.get_module_eeprom = mlx5_get_module_eeprom,
1402 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1403 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1404 };
1405 
1406 /* Available operations from secondary process. */
1407 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1408 	.stats_get = mlx5_stats_get,
1409 	.stats_reset = mlx5_stats_reset,
1410 	.xstats_get = mlx5_xstats_get,
1411 	.xstats_reset = mlx5_xstats_reset,
1412 	.xstats_get_names = mlx5_xstats_get_names,
1413 	.fw_version_get = mlx5_fw_version_get,
1414 	.dev_infos_get = mlx5_dev_infos_get,
1415 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1416 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1417 	.rxq_info_get = mlx5_rxq_info_get,
1418 	.txq_info_get = mlx5_txq_info_get,
1419 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1420 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1421 	.get_module_info = mlx5_get_module_info,
1422 	.get_module_eeprom = mlx5_get_module_eeprom,
1423 };
1424 
1425 /* Available operations in flow isolated mode. */
1426 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1427 	.dev_configure = mlx5_dev_configure,
1428 	.dev_start = mlx5_dev_start,
1429 	.dev_stop = mlx5_dev_stop,
1430 	.dev_set_link_down = mlx5_set_link_down,
1431 	.dev_set_link_up = mlx5_set_link_up,
1432 	.dev_close = mlx5_dev_close,
1433 	.promiscuous_enable = mlx5_promiscuous_enable,
1434 	.promiscuous_disable = mlx5_promiscuous_disable,
1435 	.allmulticast_enable = mlx5_allmulticast_enable,
1436 	.allmulticast_disable = mlx5_allmulticast_disable,
1437 	.link_update = mlx5_link_update,
1438 	.stats_get = mlx5_stats_get,
1439 	.stats_reset = mlx5_stats_reset,
1440 	.xstats_get = mlx5_xstats_get,
1441 	.xstats_reset = mlx5_xstats_reset,
1442 	.xstats_get_names = mlx5_xstats_get_names,
1443 	.fw_version_get = mlx5_fw_version_get,
1444 	.dev_infos_get = mlx5_dev_infos_get,
1445 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1446 	.vlan_filter_set = mlx5_vlan_filter_set,
1447 	.rx_queue_setup = mlx5_rx_queue_setup,
1448 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1449 	.tx_queue_setup = mlx5_tx_queue_setup,
1450 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1451 	.rx_queue_release = mlx5_rx_queue_release,
1452 	.tx_queue_release = mlx5_tx_queue_release,
1453 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1454 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1455 	.mac_addr_remove = mlx5_mac_addr_remove,
1456 	.mac_addr_add = mlx5_mac_addr_add,
1457 	.mac_addr_set = mlx5_mac_addr_set,
1458 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1459 	.mtu_set = mlx5_dev_set_mtu,
1460 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1461 	.vlan_offload_set = mlx5_vlan_offload_set,
1462 	.filter_ctrl = mlx5_dev_filter_ctrl,
1463 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1464 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1465 	.rxq_info_get = mlx5_rxq_info_get,
1466 	.txq_info_get = mlx5_txq_info_get,
1467 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1468 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1469 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1470 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1471 	.is_removed = mlx5_is_removed,
1472 	.get_module_info = mlx5_get_module_info,
1473 	.get_module_eeprom = mlx5_get_module_eeprom,
1474 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1475 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1476 };
1477 
1478 /**
1479  * Verify and store value for device argument.
1480  *
1481  * @param[in] key
1482  *   Key argument to verify.
1483  * @param[in] val
1484  *   Value associated with key.
1485  * @param opaque
1486  *   User data.
1487  *
1488  * @return
1489  *   0 on success, a negative errno value otherwise and rte_errno is set.
1490  */
1491 static int
1492 mlx5_args_check(const char *key, const char *val, void *opaque)
1493 {
1494 	struct mlx5_dev_config *config = opaque;
1495 	unsigned long tmp;
1496 
1497 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1498 	if (!strcmp(MLX5_REPRESENTOR, key))
1499 		return 0;
1500 	errno = 0;
1501 	tmp = strtoul(val, NULL, 0);
1502 	if (errno) {
1503 		rte_errno = errno;
1504 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1505 		return -rte_errno;
1506 	}
1507 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1508 		config->cqe_comp = !!tmp;
1509 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1510 		config->cqe_pad = !!tmp;
1511 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1512 		config->hw_padding = !!tmp;
1513 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1514 		config->mprq.enabled = !!tmp;
1515 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1516 		config->mprq.stride_num_n = tmp;
1517 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1518 		config->mprq.max_memcpy_len = tmp;
1519 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1520 		config->mprq.min_rxqs_num = tmp;
1521 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1522 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1523 				 " converted to txq_inline_max", key);
1524 		config->txq_inline_max = tmp;
1525 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1526 		config->txq_inline_max = tmp;
1527 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1528 		config->txq_inline_min = tmp;
1529 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1530 		config->txq_inline_mpw = tmp;
1531 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1532 		config->txqs_inline = tmp;
1533 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1534 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1535 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1536 		config->mps = !!tmp;
1537 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1538 		if (tmp != MLX5_TXDB_CACHED &&
1539 		    tmp != MLX5_TXDB_NCACHED &&
1540 		    tmp != MLX5_TXDB_HEURISTIC) {
1541 			DRV_LOG(ERR, "invalid Tx doorbell "
1542 				     "mapping parameter");
1543 			rte_errno = EINVAL;
1544 			return -rte_errno;
1545 		}
1546 		config->dbnc = tmp;
1547 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1548 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1549 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1550 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1551 				 " converted to txq_inline_mpw", key);
1552 		config->txq_inline_mpw = tmp;
1553 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1554 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1555 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1556 		config->rx_vec_en = !!tmp;
1557 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1558 		config->l3_vxlan_en = !!tmp;
1559 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1560 		config->vf_nl_en = !!tmp;
1561 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1562 		config->dv_esw_en = !!tmp;
1563 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1564 		config->dv_flow_en = !!tmp;
1565 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1566 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1567 		    tmp != MLX5_XMETA_MODE_META16 &&
1568 		    tmp != MLX5_XMETA_MODE_META32) {
1569 			DRV_LOG(ERR, "invalid extensive "
1570 				     "metadata parameter");
1571 			rte_errno = EINVAL;
1572 			return -rte_errno;
1573 		}
1574 		config->dv_xmeta_en = tmp;
1575 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1576 		config->mr_ext_memseg_en = !!tmp;
1577 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1578 		config->max_dump_files_num = tmp;
1579 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1580 		config->lro.timeout = tmp;
1581 	} else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1582 		DRV_LOG(DEBUG, "class argument is %s.", val);
1583 	} else {
1584 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1585 		rte_errno = EINVAL;
1586 		return -rte_errno;
1587 	}
1588 	return 0;
1589 }
1590 
1591 /**
1592  * Parse device parameters.
1593  *
1594  * @param config
1595  *   Pointer to device configuration structure.
1596  * @param devargs
1597  *   Device arguments structure.
1598  *
1599  * @return
1600  *   0 on success, a negative errno value otherwise and rte_errno is set.
1601  */
1602 static int
1603 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1604 {
1605 	const char **params = (const char *[]){
1606 		MLX5_RXQ_CQE_COMP_EN,
1607 		MLX5_RXQ_CQE_PAD_EN,
1608 		MLX5_RXQ_PKT_PAD_EN,
1609 		MLX5_RX_MPRQ_EN,
1610 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1611 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1612 		MLX5_RXQS_MIN_MPRQ,
1613 		MLX5_TXQ_INLINE,
1614 		MLX5_TXQ_INLINE_MIN,
1615 		MLX5_TXQ_INLINE_MAX,
1616 		MLX5_TXQ_INLINE_MPW,
1617 		MLX5_TXQS_MIN_INLINE,
1618 		MLX5_TXQS_MAX_VEC,
1619 		MLX5_TXQ_MPW_EN,
1620 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1621 		MLX5_TXQ_MAX_INLINE_LEN,
1622 		MLX5_TX_DB_NC,
1623 		MLX5_TX_VEC_EN,
1624 		MLX5_RX_VEC_EN,
1625 		MLX5_L3_VXLAN_EN,
1626 		MLX5_VF_NL_EN,
1627 		MLX5_DV_ESW_EN,
1628 		MLX5_DV_FLOW_EN,
1629 		MLX5_DV_XMETA_EN,
1630 		MLX5_MR_EXT_MEMSEG_EN,
1631 		MLX5_REPRESENTOR,
1632 		MLX5_MAX_DUMP_FILES_NUM,
1633 		MLX5_LRO_TIMEOUT_USEC,
1634 		MLX5_CLASS_ARG_NAME,
1635 		NULL,
1636 	};
1637 	struct rte_kvargs *kvlist;
1638 	int ret = 0;
1639 	int i;
1640 
1641 	if (devargs == NULL)
1642 		return 0;
1643 	/* Following UGLY cast is done to pass checkpatch. */
1644 	kvlist = rte_kvargs_parse(devargs->args, params);
1645 	if (kvlist == NULL) {
1646 		rte_errno = EINVAL;
1647 		return -rte_errno;
1648 	}
1649 	/* Process parameters. */
1650 	for (i = 0; (params[i] != NULL); ++i) {
1651 		if (rte_kvargs_count(kvlist, params[i])) {
1652 			ret = rte_kvargs_process(kvlist, params[i],
1653 						 mlx5_args_check, config);
1654 			if (ret) {
1655 				rte_errno = EINVAL;
1656 				rte_kvargs_free(kvlist);
1657 				return -rte_errno;
1658 			}
1659 		}
1660 	}
1661 	rte_kvargs_free(kvlist);
1662 	return 0;
1663 }
1664 
1665 static struct rte_pci_driver mlx5_driver;
1666 
1667 /**
1668  * PMD global initialization.
1669  *
1670  * Independent from individual device, this function initializes global
1671  * per-PMD data structures distinguishing primary and secondary processes.
1672  * Hence, each initialization is called once per a process.
1673  *
1674  * @return
1675  *   0 on success, a negative errno value otherwise and rte_errno is set.
1676  */
1677 static int
1678 mlx5_init_once(void)
1679 {
1680 	struct mlx5_shared_data *sd;
1681 	struct mlx5_local_data *ld = &mlx5_local_data;
1682 	int ret = 0;
1683 
1684 	if (mlx5_init_shared_data())
1685 		return -rte_errno;
1686 	sd = mlx5_shared_data;
1687 	MLX5_ASSERT(sd);
1688 	rte_spinlock_lock(&sd->lock);
1689 	switch (rte_eal_process_type()) {
1690 	case RTE_PROC_PRIMARY:
1691 		if (sd->init_done)
1692 			break;
1693 		LIST_INIT(&sd->mem_event_cb_list);
1694 		rte_rwlock_init(&sd->mem_event_rwlock);
1695 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1696 						mlx5_mr_mem_event_cb, NULL);
1697 		ret = mlx5_mp_init_primary();
1698 		if (ret)
1699 			goto out;
1700 		sd->init_done = true;
1701 		break;
1702 	case RTE_PROC_SECONDARY:
1703 		if (ld->init_done)
1704 			break;
1705 		ret = mlx5_mp_init_secondary();
1706 		if (ret)
1707 			goto out;
1708 		++sd->secondary_cnt;
1709 		ld->init_done = true;
1710 		break;
1711 	default:
1712 		break;
1713 	}
1714 out:
1715 	rte_spinlock_unlock(&sd->lock);
1716 	return ret;
1717 }
1718 
1719 /**
1720  * Configures the minimal amount of data to inline into WQE
1721  * while sending packets.
1722  *
1723  * - the txq_inline_min has the maximal priority, if this
1724  *   key is specified in devargs
1725  * - if DevX is enabled the inline mode is queried from the
1726  *   device (HCA attributes and NIC vport context if needed).
1727  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1728  *   and none (0 bytes) for other NICs
1729  *
1730  * @param spawn
1731  *   Verbs device parameters (name, port, switch_info) to spawn.
1732  * @param config
1733  *   Device configuration parameters.
1734  */
1735 static void
1736 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1737 		    struct mlx5_dev_config *config)
1738 {
1739 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1740 		/* Application defines size of inlined data explicitly. */
1741 		switch (spawn->pci_dev->id.device_id) {
1742 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1743 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1744 			if (config->txq_inline_min <
1745 				       (int)MLX5_INLINE_HSIZE_L2) {
1746 				DRV_LOG(DEBUG,
1747 					"txq_inline_mix aligned to minimal"
1748 					" ConnectX-4 required value %d",
1749 					(int)MLX5_INLINE_HSIZE_L2);
1750 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1751 			}
1752 			break;
1753 		}
1754 		goto exit;
1755 	}
1756 	if (config->hca_attr.eth_net_offloads) {
1757 		/* We have DevX enabled, inline mode queried successfully. */
1758 		switch (config->hca_attr.wqe_inline_mode) {
1759 		case MLX5_CAP_INLINE_MODE_L2:
1760 			/* outer L2 header must be inlined. */
1761 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1762 			goto exit;
1763 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1764 			/* No inline data are required by NIC. */
1765 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1766 			config->hw_vlan_insert =
1767 				config->hca_attr.wqe_vlan_insert;
1768 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1769 			goto exit;
1770 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1771 			/* inline mode is defined by NIC vport context. */
1772 			if (!config->hca_attr.eth_virt)
1773 				break;
1774 			switch (config->hca_attr.vport_inline_mode) {
1775 			case MLX5_INLINE_MODE_NONE:
1776 				config->txq_inline_min =
1777 					MLX5_INLINE_HSIZE_NONE;
1778 				goto exit;
1779 			case MLX5_INLINE_MODE_L2:
1780 				config->txq_inline_min =
1781 					MLX5_INLINE_HSIZE_L2;
1782 				goto exit;
1783 			case MLX5_INLINE_MODE_IP:
1784 				config->txq_inline_min =
1785 					MLX5_INLINE_HSIZE_L3;
1786 				goto exit;
1787 			case MLX5_INLINE_MODE_TCP_UDP:
1788 				config->txq_inline_min =
1789 					MLX5_INLINE_HSIZE_L4;
1790 				goto exit;
1791 			case MLX5_INLINE_MODE_INNER_L2:
1792 				config->txq_inline_min =
1793 					MLX5_INLINE_HSIZE_INNER_L2;
1794 				goto exit;
1795 			case MLX5_INLINE_MODE_INNER_IP:
1796 				config->txq_inline_min =
1797 					MLX5_INLINE_HSIZE_INNER_L3;
1798 				goto exit;
1799 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1800 				config->txq_inline_min =
1801 					MLX5_INLINE_HSIZE_INNER_L4;
1802 				goto exit;
1803 			}
1804 		}
1805 	}
1806 	/*
1807 	 * We get here if we are unable to deduce
1808 	 * inline data size with DevX. Try PCI ID
1809 	 * to determine old NICs.
1810 	 */
1811 	switch (spawn->pci_dev->id.device_id) {
1812 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1813 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1814 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1815 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1816 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1817 		config->hw_vlan_insert = 0;
1818 		break;
1819 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1820 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1821 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1822 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1823 		/*
1824 		 * These NICs support VLAN insertion from WQE and
1825 		 * report the wqe_vlan_insert flag. But there is the bug
1826 		 * and PFC control may be broken, so disable feature.
1827 		 */
1828 		config->hw_vlan_insert = 0;
1829 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1830 		break;
1831 	default:
1832 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1833 		break;
1834 	}
1835 exit:
1836 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1837 }
1838 
1839 /**
1840  * Configures the metadata mask fields in the shared context.
1841  *
1842  * @param [in] dev
1843  *   Pointer to Ethernet device.
1844  */
1845 static void
1846 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1847 {
1848 	struct mlx5_priv *priv = dev->data->dev_private;
1849 	struct mlx5_ibv_shared *sh = priv->sh;
1850 	uint32_t meta, mark, reg_c0;
1851 
1852 	reg_c0 = ~priv->vport_meta_mask;
1853 	switch (priv->config.dv_xmeta_en) {
1854 	case MLX5_XMETA_MODE_LEGACY:
1855 		meta = UINT32_MAX;
1856 		mark = MLX5_FLOW_MARK_MASK;
1857 		break;
1858 	case MLX5_XMETA_MODE_META16:
1859 		meta = reg_c0 >> rte_bsf32(reg_c0);
1860 		mark = MLX5_FLOW_MARK_MASK;
1861 		break;
1862 	case MLX5_XMETA_MODE_META32:
1863 		meta = UINT32_MAX;
1864 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1865 		break;
1866 	default:
1867 		meta = 0;
1868 		mark = 0;
1869 		MLX5_ASSERT(false);
1870 		break;
1871 	}
1872 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1873 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1874 				 sh->dv_mark_mask, mark);
1875 	else
1876 		sh->dv_mark_mask = mark;
1877 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1878 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1879 				 sh->dv_meta_mask, meta);
1880 	else
1881 		sh->dv_meta_mask = meta;
1882 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1883 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1884 				 sh->dv_meta_mask, reg_c0);
1885 	else
1886 		sh->dv_regc0_mask = reg_c0;
1887 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1888 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1889 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1890 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1891 }
1892 
1893 /**
1894  * Allocate page of door-bells and register it using DevX API.
1895  *
1896  * @param [in] dev
1897  *   Pointer to Ethernet device.
1898  *
1899  * @return
1900  *   Pointer to new page on success, NULL otherwise.
1901  */
1902 static struct mlx5_devx_dbr_page *
1903 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1904 {
1905 	struct mlx5_priv *priv = dev->data->dev_private;
1906 	struct mlx5_devx_dbr_page *page;
1907 
1908 	/* Allocate space for door-bell page and management data. */
1909 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1910 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1911 	if (!page) {
1912 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1913 			dev->data->port_id);
1914 		return NULL;
1915 	}
1916 	/* Register allocated memory. */
1917 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1918 					      MLX5_DBR_PAGE_SIZE, 0);
1919 	if (!page->umem) {
1920 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1921 			dev->data->port_id);
1922 		rte_free(page);
1923 		return NULL;
1924 	}
1925 	return page;
1926 }
1927 
1928 /**
1929  * Find the next available door-bell, allocate new page if needed.
1930  *
1931  * @param [in] dev
1932  *   Pointer to Ethernet device.
1933  * @param [out] dbr_page
1934  *   Door-bell page containing the page data.
1935  *
1936  * @return
1937  *   Door-bell address offset on success, a negative error value otherwise.
1938  */
1939 int64_t
1940 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1941 {
1942 	struct mlx5_priv *priv = dev->data->dev_private;
1943 	struct mlx5_devx_dbr_page *page = NULL;
1944 	uint32_t i, j;
1945 
1946 	LIST_FOREACH(page, &priv->dbrpgs, next)
1947 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1948 			break;
1949 	if (!page) { /* No page with free door-bell exists. */
1950 		page = mlx5_alloc_dbr_page(dev);
1951 		if (!page) /* Failed to allocate new page. */
1952 			return (-1);
1953 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1954 	}
1955 	/* Loop to find bitmap part with clear bit. */
1956 	for (i = 0;
1957 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1958 	     i++)
1959 		; /* Empty. */
1960 	/* Find the first clear bit. */
1961 	j = rte_bsf64(~page->dbr_bitmap[i]);
1962 	MLX5_ASSERT(i < (MLX5_DBR_PER_PAGE / 64));
1963 	page->dbr_bitmap[i] |= (1 << j);
1964 	page->dbr_count++;
1965 	*dbr_page = page;
1966 	return (((i * 64) + j) * sizeof(uint64_t));
1967 }
1968 
1969 /**
1970  * Release a door-bell record.
1971  *
1972  * @param [in] dev
1973  *   Pointer to Ethernet device.
1974  * @param [in] umem_id
1975  *   UMEM ID of page containing the door-bell record to release.
1976  * @param [in] offset
1977  *   Offset of door-bell record in page.
1978  *
1979  * @return
1980  *   0 on success, a negative error value otherwise.
1981  */
1982 int32_t
1983 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1984 {
1985 	struct mlx5_priv *priv = dev->data->dev_private;
1986 	struct mlx5_devx_dbr_page *page = NULL;
1987 	int ret = 0;
1988 
1989 	LIST_FOREACH(page, &priv->dbrpgs, next)
1990 		/* Find the page this address belongs to. */
1991 		if (page->umem->umem_id == umem_id)
1992 			break;
1993 	if (!page)
1994 		return -EINVAL;
1995 	page->dbr_count--;
1996 	if (!page->dbr_count) {
1997 		/* Page not used, free it and remove from list. */
1998 		LIST_REMOVE(page, next);
1999 		if (page->umem)
2000 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
2001 		rte_free(page);
2002 	} else {
2003 		/* Mark in bitmap that this door-bell is not in use. */
2004 		offset /= MLX5_DBR_SIZE;
2005 		int i = offset / 64;
2006 		int j = offset % 64;
2007 
2008 		page->dbr_bitmap[i] &= ~(1 << j);
2009 	}
2010 	return ret;
2011 }
2012 
2013 int
2014 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2015 {
2016 	static const char *const dynf_names[] = {
2017 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2018 		RTE_MBUF_DYNFLAG_METADATA_NAME
2019 	};
2020 	unsigned int i;
2021 
2022 	if (n < RTE_DIM(dynf_names))
2023 		return -ENOMEM;
2024 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2025 		if (names[i] == NULL)
2026 			return -EINVAL;
2027 		strcpy(names[i], dynf_names[i]);
2028 	}
2029 	return RTE_DIM(dynf_names);
2030 }
2031 
2032 /**
2033  * Check sibling device configurations.
2034  *
2035  * Sibling devices sharing the Infiniband device context
2036  * should have compatible configurations. This regards
2037  * representors and bonding slaves.
2038  *
2039  * @param priv
2040  *   Private device descriptor.
2041  * @param config
2042  *   Configuration of the device is going to be created.
2043  *
2044  * @return
2045  *   0 on success, EINVAL otherwise
2046  */
2047 static int
2048 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2049 			      struct mlx5_dev_config *config)
2050 {
2051 	struct mlx5_ibv_shared *sh = priv->sh;
2052 	struct mlx5_dev_config *sh_conf = NULL;
2053 	uint16_t port_id;
2054 
2055 	MLX5_ASSERT(sh);
2056 	/* Nothing to compare for the single/first device. */
2057 	if (sh->refcnt == 1)
2058 		return 0;
2059 	/* Find the device with shared context. */
2060 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2061 		struct mlx5_priv *opriv =
2062 			rte_eth_devices[port_id].data->dev_private;
2063 
2064 		if (opriv && opriv != priv && opriv->sh == sh) {
2065 			sh_conf = &opriv->config;
2066 			break;
2067 		}
2068 	}
2069 	if (!sh_conf)
2070 		return 0;
2071 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2072 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2073 			     " for shared %s context", sh->ibdev_name);
2074 		rte_errno = EINVAL;
2075 		return rte_errno;
2076 	}
2077 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2078 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2079 			     " for shared %s context", sh->ibdev_name);
2080 		rte_errno = EINVAL;
2081 		return rte_errno;
2082 	}
2083 	return 0;
2084 }
2085 /**
2086  * Spawn an Ethernet device from Verbs information.
2087  *
2088  * @param dpdk_dev
2089  *   Backing DPDK device.
2090  * @param spawn
2091  *   Verbs device parameters (name, port, switch_info) to spawn.
2092  * @param config
2093  *   Device configuration parameters.
2094  *
2095  * @return
2096  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2097  *   is set. The following errors are defined:
2098  *
2099  *   EBUSY: device is not supposed to be spawned.
2100  *   EEXIST: device is already spawned
2101  */
2102 static struct rte_eth_dev *
2103 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2104 	       struct mlx5_dev_spawn_data *spawn,
2105 	       struct mlx5_dev_config config)
2106 {
2107 	const struct mlx5_switch_info *switch_info = &spawn->info;
2108 	struct mlx5_ibv_shared *sh = NULL;
2109 	struct ibv_port_attr port_attr;
2110 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2111 	struct rte_eth_dev *eth_dev = NULL;
2112 	struct mlx5_priv *priv = NULL;
2113 	int err = 0;
2114 	unsigned int hw_padding = 0;
2115 	unsigned int mps;
2116 	unsigned int cqe_comp;
2117 	unsigned int cqe_pad = 0;
2118 	unsigned int tunnel_en = 0;
2119 	unsigned int mpls_en = 0;
2120 	unsigned int swp = 0;
2121 	unsigned int mprq = 0;
2122 	unsigned int mprq_min_stride_size_n = 0;
2123 	unsigned int mprq_max_stride_size_n = 0;
2124 	unsigned int mprq_min_stride_num_n = 0;
2125 	unsigned int mprq_max_stride_num_n = 0;
2126 	struct rte_ether_addr mac;
2127 	char name[RTE_ETH_NAME_MAX_LEN];
2128 	int own_domain_id = 0;
2129 	uint16_t port_id;
2130 	unsigned int i;
2131 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2132 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2133 #endif
2134 
2135 	/* Determine if this port representor is supposed to be spawned. */
2136 	if (switch_info->representor && dpdk_dev->devargs) {
2137 		struct rte_eth_devargs eth_da;
2138 
2139 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2140 		if (err) {
2141 			rte_errno = -err;
2142 			DRV_LOG(ERR, "failed to process device arguments: %s",
2143 				strerror(rte_errno));
2144 			return NULL;
2145 		}
2146 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
2147 			if (eth_da.representor_ports[i] ==
2148 			    (uint16_t)switch_info->port_name)
2149 				break;
2150 		if (i == eth_da.nb_representor_ports) {
2151 			rte_errno = EBUSY;
2152 			return NULL;
2153 		}
2154 	}
2155 	/* Build device name. */
2156 	if (spawn->pf_bond <  0) {
2157 		/* Single device. */
2158 		if (!switch_info->representor)
2159 			strlcpy(name, dpdk_dev->name, sizeof(name));
2160 		else
2161 			snprintf(name, sizeof(name), "%s_representor_%u",
2162 				 dpdk_dev->name, switch_info->port_name);
2163 	} else {
2164 		/* Bonding device. */
2165 		if (!switch_info->representor)
2166 			snprintf(name, sizeof(name), "%s_%s",
2167 				 dpdk_dev->name, spawn->ibv_dev->name);
2168 		else
2169 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
2170 				 dpdk_dev->name, spawn->ibv_dev->name,
2171 				 switch_info->port_name);
2172 	}
2173 	/* check if the device is already spawned */
2174 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2175 		rte_errno = EEXIST;
2176 		return NULL;
2177 	}
2178 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2179 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2180 		eth_dev = rte_eth_dev_attach_secondary(name);
2181 		if (eth_dev == NULL) {
2182 			DRV_LOG(ERR, "can not attach rte ethdev");
2183 			rte_errno = ENOMEM;
2184 			return NULL;
2185 		}
2186 		eth_dev->device = dpdk_dev;
2187 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2188 		err = mlx5_proc_priv_init(eth_dev);
2189 		if (err)
2190 			return NULL;
2191 		/* Receive command fd from primary process */
2192 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2193 		if (err < 0)
2194 			return NULL;
2195 		/* Remap UAR for Tx queues. */
2196 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2197 		if (err)
2198 			return NULL;
2199 		/*
2200 		 * Ethdev pointer is still required as input since
2201 		 * the primary device is not accessible from the
2202 		 * secondary process.
2203 		 */
2204 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2205 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2206 		return eth_dev;
2207 	}
2208 	/*
2209 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2210 	 * advance to create dv/verbs device context. We proceed the
2211 	 * devargs here to get ones, and later proceed devargs again
2212 	 * to override some hardware settings.
2213 	 */
2214 	err = mlx5_args(&config, dpdk_dev->devargs);
2215 	if (err) {
2216 		err = rte_errno;
2217 		DRV_LOG(ERR, "failed to process device arguments: %s",
2218 			strerror(rte_errno));
2219 		goto error;
2220 	}
2221 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2222 	if (!sh)
2223 		return NULL;
2224 	config.devx = sh->devx;
2225 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2226 	config.dest_tir = 1;
2227 #endif
2228 #ifdef HAVE_IBV_MLX5_MOD_SWP
2229 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2230 #endif
2231 	/*
2232 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2233 	 * as all ConnectX-5 devices.
2234 	 */
2235 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2236 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2237 #endif
2238 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2239 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2240 #endif
2241 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2242 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2243 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2244 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2245 			mps = MLX5_MPW_ENHANCED;
2246 		} else {
2247 			DRV_LOG(DEBUG, "MPW is supported");
2248 			mps = MLX5_MPW;
2249 		}
2250 	} else {
2251 		DRV_LOG(DEBUG, "MPW isn't supported");
2252 		mps = MLX5_MPW_DISABLED;
2253 	}
2254 #ifdef HAVE_IBV_MLX5_MOD_SWP
2255 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2256 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2257 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2258 #endif
2259 	config.swp = !!swp;
2260 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2261 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2262 		struct mlx5dv_striding_rq_caps mprq_caps =
2263 			dv_attr.striding_rq_caps;
2264 
2265 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2266 			mprq_caps.min_single_stride_log_num_of_bytes);
2267 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2268 			mprq_caps.max_single_stride_log_num_of_bytes);
2269 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2270 			mprq_caps.min_single_wqe_log_num_of_strides);
2271 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2272 			mprq_caps.max_single_wqe_log_num_of_strides);
2273 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2274 			mprq_caps.supported_qpts);
2275 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2276 		mprq = 1;
2277 		mprq_min_stride_size_n =
2278 			mprq_caps.min_single_stride_log_num_of_bytes;
2279 		mprq_max_stride_size_n =
2280 			mprq_caps.max_single_stride_log_num_of_bytes;
2281 		mprq_min_stride_num_n =
2282 			mprq_caps.min_single_wqe_log_num_of_strides;
2283 		mprq_max_stride_num_n =
2284 			mprq_caps.max_single_wqe_log_num_of_strides;
2285 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2286 						   mprq_min_stride_num_n);
2287 	}
2288 #endif
2289 	if (RTE_CACHE_LINE_SIZE == 128 &&
2290 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2291 		cqe_comp = 0;
2292 	else
2293 		cqe_comp = 1;
2294 	config.cqe_comp = cqe_comp;
2295 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2296 	/* Whether device supports 128B Rx CQE padding. */
2297 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2298 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2299 #endif
2300 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2301 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2302 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2303 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2304 			     (dv_attr.tunnel_offloads_caps &
2305 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2306 			     (dv_attr.tunnel_offloads_caps &
2307 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2308 	}
2309 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2310 		tunnel_en ? "" : "not ");
2311 #else
2312 	DRV_LOG(WARNING,
2313 		"tunnel offloading disabled due to old OFED/rdma-core version");
2314 #endif
2315 	config.tunnel_en = tunnel_en;
2316 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2317 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2318 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2319 		   (dv_attr.tunnel_offloads_caps &
2320 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2321 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2322 		mpls_en ? "" : "not ");
2323 #else
2324 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2325 		" old OFED/rdma-core version or firmware configuration");
2326 #endif
2327 	config.mpls_en = mpls_en;
2328 	/* Check port status. */
2329 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2330 	if (err) {
2331 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2332 		goto error;
2333 	}
2334 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2335 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2336 		err = EINVAL;
2337 		goto error;
2338 	}
2339 	if (port_attr.state != IBV_PORT_ACTIVE)
2340 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2341 			mlx5_glue->port_state_str(port_attr.state),
2342 			port_attr.state);
2343 	/* Allocate private eth device data. */
2344 	priv = rte_zmalloc("ethdev private structure",
2345 			   sizeof(*priv),
2346 			   RTE_CACHE_LINE_SIZE);
2347 	if (priv == NULL) {
2348 		DRV_LOG(ERR, "priv allocation failure");
2349 		err = ENOMEM;
2350 		goto error;
2351 	}
2352 	priv->sh = sh;
2353 	priv->ibv_port = spawn->ibv_port;
2354 	priv->pci_dev = spawn->pci_dev;
2355 	priv->mtu = RTE_ETHER_MTU;
2356 #ifndef RTE_ARCH_64
2357 	/* Initialize UAR access locks for 32bit implementations. */
2358 	rte_spinlock_init(&priv->uar_lock_cq);
2359 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2360 		rte_spinlock_init(&priv->uar_lock[i]);
2361 #endif
2362 	/* Some internal functions rely on Netlink sockets, open them now. */
2363 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2364 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2365 	priv->representor = !!switch_info->representor;
2366 	priv->master = !!switch_info->master;
2367 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2368 	priv->vport_meta_tag = 0;
2369 	priv->vport_meta_mask = 0;
2370 	priv->pf_bond = spawn->pf_bond;
2371 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2372 	/*
2373 	 * The DevX port query API is implemented. E-Switch may use
2374 	 * either vport or reg_c[0] metadata register to match on
2375 	 * vport index. The engaged part of metadata register is
2376 	 * defined by mask.
2377 	 */
2378 	if (switch_info->representor || switch_info->master) {
2379 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2380 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2381 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2382 						 &devx_port);
2383 		if (err) {
2384 			DRV_LOG(WARNING,
2385 				"can't query devx port %d on device %s",
2386 				spawn->ibv_port, spawn->ibv_dev->name);
2387 			devx_port.comp_mask = 0;
2388 		}
2389 	}
2390 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2391 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2392 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2393 		if (!priv->vport_meta_mask) {
2394 			DRV_LOG(ERR, "vport zero mask for port %d"
2395 				     " on bonding device %s",
2396 				     spawn->ibv_port, spawn->ibv_dev->name);
2397 			err = ENOTSUP;
2398 			goto error;
2399 		}
2400 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2401 			DRV_LOG(ERR, "invalid vport tag for port %d"
2402 				     " on bonding device %s",
2403 				     spawn->ibv_port, spawn->ibv_dev->name);
2404 			err = ENOTSUP;
2405 			goto error;
2406 		}
2407 	}
2408 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2409 		priv->vport_id = devx_port.vport_num;
2410 	} else if (spawn->pf_bond >= 0) {
2411 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2412 			     " on bonding device %s",
2413 			     spawn->ibv_port, spawn->ibv_dev->name);
2414 		err = ENOTSUP;
2415 		goto error;
2416 	} else {
2417 		/* Suppose vport index in compatible way. */
2418 		priv->vport_id = switch_info->representor ?
2419 				 switch_info->port_name + 1 : -1;
2420 	}
2421 #else
2422 	/*
2423 	 * Kernel/rdma_core support single E-Switch per PF configurations
2424 	 * only and vport_id field contains the vport index for
2425 	 * associated VF, which is deduced from representor port name.
2426 	 * For example, let's have the IB device port 10, it has
2427 	 * attached network device eth0, which has port name attribute
2428 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2429 	 * as 3 (2+1). This assigning schema should be changed if the
2430 	 * multiple E-Switch instances per PF configurations or/and PCI
2431 	 * subfunctions are added.
2432 	 */
2433 	priv->vport_id = switch_info->representor ?
2434 			 switch_info->port_name + 1 : -1;
2435 #endif
2436 	/* representor_id field keeps the unmodified VF index. */
2437 	priv->representor_id = switch_info->representor ?
2438 			       switch_info->port_name : -1;
2439 	/*
2440 	 * Look for sibling devices in order to reuse their switch domain
2441 	 * if any, otherwise allocate one.
2442 	 */
2443 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2444 		const struct mlx5_priv *opriv =
2445 			rte_eth_devices[port_id].data->dev_private;
2446 
2447 		if (!opriv ||
2448 		    opriv->sh != priv->sh ||
2449 			opriv->domain_id ==
2450 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2451 			continue;
2452 		priv->domain_id = opriv->domain_id;
2453 		break;
2454 	}
2455 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2456 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2457 		if (err) {
2458 			err = rte_errno;
2459 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2460 				strerror(rte_errno));
2461 			goto error;
2462 		}
2463 		own_domain_id = 1;
2464 	}
2465 	/* Override some values set by hardware configuration. */
2466 	mlx5_args(&config, dpdk_dev->devargs);
2467 	err = mlx5_dev_check_sibling_config(priv, &config);
2468 	if (err)
2469 		goto error;
2470 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2471 			    IBV_DEVICE_RAW_IP_CSUM);
2472 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2473 		(config.hw_csum ? "" : "not "));
2474 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2475 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2476 	DRV_LOG(DEBUG, "counters are not supported");
2477 #endif
2478 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2479 	if (config.dv_flow_en) {
2480 		DRV_LOG(WARNING, "DV flow is not supported");
2481 		config.dv_flow_en = 0;
2482 	}
2483 #endif
2484 	config.ind_table_max_size =
2485 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2486 	/*
2487 	 * Remove this check once DPDK supports larger/variable
2488 	 * indirection tables.
2489 	 */
2490 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2491 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2492 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2493 		config.ind_table_max_size);
2494 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2495 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2496 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2497 		(config.hw_vlan_strip ? "" : "not "));
2498 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2499 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2500 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2501 		(config.hw_fcs_strip ? "" : "not "));
2502 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2503 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2504 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2505 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2506 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2507 #endif
2508 	if (config.hw_padding && !hw_padding) {
2509 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2510 		config.hw_padding = 0;
2511 	} else if (config.hw_padding) {
2512 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2513 	}
2514 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2515 		      (sh->device_attr.tso_caps.supported_qpts &
2516 		       (1 << IBV_QPT_RAW_PACKET)));
2517 	if (config.tso)
2518 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2519 	/*
2520 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2521 	 * by default.
2522 	 */
2523 	if (config.mps == MLX5_ARG_UNSET)
2524 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2525 							  MLX5_MPW_DISABLED;
2526 	else
2527 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2528 	DRV_LOG(INFO, "%sMPS is %s",
2529 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2530 		config.mps == MLX5_MPW ? "legacy " : "",
2531 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2532 	if (config.cqe_comp && !cqe_comp) {
2533 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2534 		config.cqe_comp = 0;
2535 	}
2536 	if (config.cqe_pad && !cqe_pad) {
2537 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2538 		config.cqe_pad = 0;
2539 	} else if (config.cqe_pad) {
2540 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2541 	}
2542 	if (config.devx) {
2543 		priv->counter_fallback = 0;
2544 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2545 		if (err) {
2546 			err = -err;
2547 			goto error;
2548 		}
2549 		if (!config.hca_attr.flow_counters_dump)
2550 			priv->counter_fallback = 1;
2551 #ifndef HAVE_IBV_DEVX_ASYNC
2552 		priv->counter_fallback = 1;
2553 #endif
2554 		if (priv->counter_fallback)
2555 			DRV_LOG(INFO, "Use fall-back DV counter management");
2556 		/* Check for LRO support. */
2557 		if (config.dest_tir && config.hca_attr.lro_cap &&
2558 		    config.dv_flow_en) {
2559 			/* TBD check tunnel lro caps. */
2560 			config.lro.supported = config.hca_attr.lro_cap;
2561 			DRV_LOG(DEBUG, "Device supports LRO");
2562 			/*
2563 			 * If LRO timeout is not configured by application,
2564 			 * use the minimal supported value.
2565 			 */
2566 			if (!config.lro.timeout)
2567 				config.lro.timeout =
2568 				config.hca_attr.lro_timer_supported_periods[0];
2569 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2570 				config.lro.timeout);
2571 		}
2572 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2573 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2574 		    config.dv_flow_en) {
2575 			uint8_t reg_c_mask =
2576 				config.hca_attr.qos.flow_meter_reg_c_ids;
2577 			/*
2578 			 * Meter needs two REG_C's for color match and pre-sfx
2579 			 * flow match. Here get the REG_C for color match.
2580 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2581 			 */
2582 			reg_c_mask &= 0xfc;
2583 			if (__builtin_popcount(reg_c_mask) < 1) {
2584 				priv->mtr_en = 0;
2585 				DRV_LOG(WARNING, "No available register for"
2586 					" meter.");
2587 			} else {
2588 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2589 						      REG_C_0;
2590 				priv->mtr_en = 1;
2591 				priv->mtr_reg_share =
2592 				      config.hca_attr.qos.flow_meter_reg_share;
2593 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2594 					priv->mtr_color_reg);
2595 			}
2596 		}
2597 #endif
2598 	}
2599 	if (config.mprq.enabled && mprq) {
2600 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2601 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2602 			config.mprq.stride_num_n =
2603 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2604 					mprq_min_stride_num_n);
2605 			DRV_LOG(WARNING,
2606 				"the number of strides"
2607 				" for Multi-Packet RQ is out of range,"
2608 				" setting default value (%u)",
2609 				1 << config.mprq.stride_num_n);
2610 		}
2611 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2612 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2613 	} else if (config.mprq.enabled && !mprq) {
2614 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2615 		config.mprq.enabled = 0;
2616 	}
2617 	if (config.max_dump_files_num == 0)
2618 		config.max_dump_files_num = 128;
2619 	eth_dev = rte_eth_dev_allocate(name);
2620 	if (eth_dev == NULL) {
2621 		DRV_LOG(ERR, "can not allocate rte ethdev");
2622 		err = ENOMEM;
2623 		goto error;
2624 	}
2625 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2626 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2627 	if (priv->representor) {
2628 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2629 		eth_dev->data->representor_id = priv->representor_id;
2630 	}
2631 	/*
2632 	 * Store associated network device interface index. This index
2633 	 * is permanent throughout the lifetime of device. So, we may store
2634 	 * the ifindex here and use the cached value further.
2635 	 */
2636 	MLX5_ASSERT(spawn->ifindex);
2637 	priv->if_index = spawn->ifindex;
2638 	eth_dev->data->dev_private = priv;
2639 	priv->dev_data = eth_dev->data;
2640 	eth_dev->data->mac_addrs = priv->mac;
2641 	eth_dev->device = dpdk_dev;
2642 	/* Configure the first MAC address by default. */
2643 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2644 		DRV_LOG(ERR,
2645 			"port %u cannot get MAC address, is mlx5_en"
2646 			" loaded? (errno: %s)",
2647 			eth_dev->data->port_id, strerror(rte_errno));
2648 		err = ENODEV;
2649 		goto error;
2650 	}
2651 	DRV_LOG(INFO,
2652 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2653 		eth_dev->data->port_id,
2654 		mac.addr_bytes[0], mac.addr_bytes[1],
2655 		mac.addr_bytes[2], mac.addr_bytes[3],
2656 		mac.addr_bytes[4], mac.addr_bytes[5]);
2657 #ifdef RTE_LIBRTE_MLX5_DEBUG
2658 	{
2659 		char ifname[IF_NAMESIZE];
2660 
2661 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2662 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2663 				eth_dev->data->port_id, ifname);
2664 		else
2665 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2666 				eth_dev->data->port_id);
2667 	}
2668 #endif
2669 	/* Get actual MTU if possible. */
2670 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2671 	if (err) {
2672 		err = rte_errno;
2673 		goto error;
2674 	}
2675 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2676 		priv->mtu);
2677 	/* Initialize burst functions to prevent crashes before link-up. */
2678 	eth_dev->rx_pkt_burst = removed_rx_burst;
2679 	eth_dev->tx_pkt_burst = removed_tx_burst;
2680 	eth_dev->dev_ops = &mlx5_dev_ops;
2681 	/* Register MAC address. */
2682 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2683 	if (config.vf && config.vf_nl_en)
2684 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2685 				      mlx5_ifindex(eth_dev),
2686 				      eth_dev->data->mac_addrs,
2687 				      MLX5_MAX_MAC_ADDRESSES);
2688 	TAILQ_INIT(&priv->flows);
2689 	TAILQ_INIT(&priv->ctrl_flows);
2690 	TAILQ_INIT(&priv->flow_meters);
2691 	TAILQ_INIT(&priv->flow_meter_profiles);
2692 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2693 	struct mlx5dv_ctx_allocators alctr = {
2694 		.alloc = &mlx5_alloc_verbs_buf,
2695 		.free = &mlx5_free_verbs_buf,
2696 		.data = priv,
2697 	};
2698 	mlx5_glue->dv_set_context_attr(sh->ctx,
2699 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2700 				       (void *)((uintptr_t)&alctr));
2701 	/* Bring Ethernet device up. */
2702 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2703 		eth_dev->data->port_id);
2704 	mlx5_set_link_up(eth_dev);
2705 	/*
2706 	 * Even though the interrupt handler is not installed yet,
2707 	 * interrupts will still trigger on the async_fd from
2708 	 * Verbs context returned by ibv_open_device().
2709 	 */
2710 	mlx5_link_update(eth_dev, 0);
2711 #ifdef HAVE_MLX5DV_DR_ESWITCH
2712 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2713 	      (switch_info->representor || switch_info->master)))
2714 		config.dv_esw_en = 0;
2715 #else
2716 	config.dv_esw_en = 0;
2717 #endif
2718 	/* Detect minimal data bytes to inline. */
2719 	mlx5_set_min_inline(spawn, &config);
2720 	/* Store device configuration on private structure. */
2721 	priv->config = config;
2722 	/* Create context for virtual machine VLAN workaround. */
2723 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2724 	if (config.dv_flow_en) {
2725 		err = mlx5_alloc_shared_dr(priv);
2726 		if (err)
2727 			goto error;
2728 		/*
2729 		 * RSS id is shared with meter flow id. Meter flow id can only
2730 		 * use the 24 MSB of the register.
2731 		 */
2732 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2733 				     MLX5_MTR_COLOR_BITS);
2734 		if (!priv->qrss_id_pool) {
2735 			DRV_LOG(ERR, "can't create flow id pool");
2736 			err = ENOMEM;
2737 			goto error;
2738 		}
2739 	}
2740 	/* Supported Verbs flow priority number detection. */
2741 	err = mlx5_flow_discover_priorities(eth_dev);
2742 	if (err < 0) {
2743 		err = -err;
2744 		goto error;
2745 	}
2746 	priv->config.flow_prio = err;
2747 	if (!priv->config.dv_esw_en &&
2748 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2749 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2750 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2751 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2752 	}
2753 	mlx5_set_metadata_mask(eth_dev);
2754 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2755 	    !priv->sh->dv_regc0_mask) {
2756 		DRV_LOG(ERR, "metadata mode %u is not supported "
2757 			     "(no metadata reg_c[0] is available)",
2758 			     priv->config.dv_xmeta_en);
2759 			err = ENOTSUP;
2760 			goto error;
2761 	}
2762 	/* Query availibility of metadata reg_c's. */
2763 	err = mlx5_flow_discover_mreg_c(eth_dev);
2764 	if (err < 0) {
2765 		err = -err;
2766 		goto error;
2767 	}
2768 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2769 		DRV_LOG(DEBUG,
2770 			"port %u extensive metadata register is not supported",
2771 			eth_dev->data->port_id);
2772 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2773 			DRV_LOG(ERR, "metadata mode %u is not supported "
2774 				     "(no metadata registers available)",
2775 				     priv->config.dv_xmeta_en);
2776 			err = ENOTSUP;
2777 			goto error;
2778 		}
2779 	}
2780 	if (priv->config.dv_flow_en &&
2781 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2782 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
2783 	    priv->sh->dv_regc0_mask) {
2784 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2785 						      MLX5_FLOW_MREG_HTABLE_SZ);
2786 		if (!priv->mreg_cp_tbl) {
2787 			err = ENOMEM;
2788 			goto error;
2789 		}
2790 	}
2791 	return eth_dev;
2792 error:
2793 	if (priv) {
2794 		if (priv->mreg_cp_tbl)
2795 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2796 		if (priv->sh)
2797 			mlx5_free_shared_dr(priv);
2798 		if (priv->nl_socket_route >= 0)
2799 			close(priv->nl_socket_route);
2800 		if (priv->nl_socket_rdma >= 0)
2801 			close(priv->nl_socket_rdma);
2802 		if (priv->vmwa_context)
2803 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2804 		if (priv->qrss_id_pool)
2805 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
2806 		if (own_domain_id)
2807 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2808 		rte_free(priv);
2809 		if (eth_dev != NULL)
2810 			eth_dev->data->dev_private = NULL;
2811 	}
2812 	if (eth_dev != NULL) {
2813 		/* mac_addrs must not be freed alone because part of dev_private */
2814 		eth_dev->data->mac_addrs = NULL;
2815 		rte_eth_dev_release_port(eth_dev);
2816 	}
2817 	if (sh)
2818 		mlx5_free_shared_ibctx(sh);
2819 	MLX5_ASSERT(err > 0);
2820 	rte_errno = err;
2821 	return NULL;
2822 }
2823 
2824 /**
2825  * Comparison callback to sort device data.
2826  *
2827  * This is meant to be used with qsort().
2828  *
2829  * @param a[in]
2830  *   Pointer to pointer to first data object.
2831  * @param b[in]
2832  *   Pointer to pointer to second data object.
2833  *
2834  * @return
2835  *   0 if both objects are equal, less than 0 if the first argument is less
2836  *   than the second, greater than 0 otherwise.
2837  */
2838 static int
2839 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2840 {
2841 	const struct mlx5_switch_info *si_a =
2842 		&((const struct mlx5_dev_spawn_data *)a)->info;
2843 	const struct mlx5_switch_info *si_b =
2844 		&((const struct mlx5_dev_spawn_data *)b)->info;
2845 	int ret;
2846 
2847 	/* Master device first. */
2848 	ret = si_b->master - si_a->master;
2849 	if (ret)
2850 		return ret;
2851 	/* Then representor devices. */
2852 	ret = si_b->representor - si_a->representor;
2853 	if (ret)
2854 		return ret;
2855 	/* Unidentified devices come last in no specific order. */
2856 	if (!si_a->representor)
2857 		return 0;
2858 	/* Order representors by name. */
2859 	return si_a->port_name - si_b->port_name;
2860 }
2861 
2862 /**
2863  * Match PCI information for possible slaves of bonding device.
2864  *
2865  * @param[in] ibv_dev
2866  *   Pointer to Infiniband device structure.
2867  * @param[in] pci_dev
2868  *   Pointer to PCI device structure to match PCI address.
2869  * @param[in] nl_rdma
2870  *   Netlink RDMA group socket handle.
2871  *
2872  * @return
2873  *   negative value if no bonding device found, otherwise
2874  *   positive index of slave PF in bonding.
2875  */
2876 static int
2877 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2878 			   const struct rte_pci_device *pci_dev,
2879 			   int nl_rdma)
2880 {
2881 	char ifname[IF_NAMESIZE + 1];
2882 	unsigned int ifindex;
2883 	unsigned int np, i;
2884 	FILE *file = NULL;
2885 	int pf = -1;
2886 
2887 	/*
2888 	 * Try to get master device name. If something goes
2889 	 * wrong suppose the lack of kernel support and no
2890 	 * bonding devices.
2891 	 */
2892 	if (nl_rdma < 0)
2893 		return -1;
2894 	if (!strstr(ibv_dev->name, "bond"))
2895 		return -1;
2896 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2897 	if (!np)
2898 		return -1;
2899 	/*
2900 	 * The Master device might not be on the predefined
2901 	 * port (not on port index 1, it is not garanted),
2902 	 * we have to scan all Infiniband device port and
2903 	 * find master.
2904 	 */
2905 	for (i = 1; i <= np; ++i) {
2906 		/* Check whether Infiniband port is populated. */
2907 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2908 		if (!ifindex)
2909 			continue;
2910 		if (!if_indextoname(ifindex, ifname))
2911 			continue;
2912 		/* Try to read bonding slave names from sysfs. */
2913 		MKSTR(slaves,
2914 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2915 		file = fopen(slaves, "r");
2916 		if (file)
2917 			break;
2918 	}
2919 	if (!file)
2920 		return -1;
2921 	/* Use safe format to check maximal buffer length. */
2922 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2923 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2924 		char tmp_str[IF_NAMESIZE + 32];
2925 		struct rte_pci_addr pci_addr;
2926 		struct mlx5_switch_info	info;
2927 
2928 		/* Process slave interface names in the loop. */
2929 		snprintf(tmp_str, sizeof(tmp_str),
2930 			 "/sys/class/net/%s", ifname);
2931 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2932 			DRV_LOG(WARNING, "can not get PCI address"
2933 					 " for netdev \"%s\"", ifname);
2934 			continue;
2935 		}
2936 		if (pci_dev->addr.domain != pci_addr.domain ||
2937 		    pci_dev->addr.bus != pci_addr.bus ||
2938 		    pci_dev->addr.devid != pci_addr.devid ||
2939 		    pci_dev->addr.function != pci_addr.function)
2940 			continue;
2941 		/* Slave interface PCI address match found. */
2942 		fclose(file);
2943 		snprintf(tmp_str, sizeof(tmp_str),
2944 			 "/sys/class/net/%s/phys_port_name", ifname);
2945 		file = fopen(tmp_str, "rb");
2946 		if (!file)
2947 			break;
2948 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2949 		if (fscanf(file, "%32s", tmp_str) == 1)
2950 			mlx5_translate_port_name(tmp_str, &info);
2951 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2952 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2953 			pf = info.port_name;
2954 		break;
2955 	}
2956 	if (file)
2957 		fclose(file);
2958 	return pf;
2959 }
2960 
2961 /**
2962  * DPDK callback to register a PCI device.
2963  *
2964  * This function spawns Ethernet devices out of a given PCI device.
2965  *
2966  * @param[in] pci_drv
2967  *   PCI driver structure (mlx5_driver).
2968  * @param[in] pci_dev
2969  *   PCI device information.
2970  *
2971  * @return
2972  *   0 on success, a negative errno value otherwise and rte_errno is set.
2973  */
2974 static int
2975 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2976 	       struct rte_pci_device *pci_dev)
2977 {
2978 	struct ibv_device **ibv_list;
2979 	/*
2980 	 * Number of found IB Devices matching with requested PCI BDF.
2981 	 * nd != 1 means there are multiple IB devices over the same
2982 	 * PCI device and we have representors and master.
2983 	 */
2984 	unsigned int nd = 0;
2985 	/*
2986 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2987 	 * we have the single multiport IB device, and there may be
2988 	 * representors attached to some of found ports.
2989 	 */
2990 	unsigned int np = 0;
2991 	/*
2992 	 * Number of DPDK ethernet devices to Spawn - either over
2993 	 * multiple IB devices or multiple ports of single IB device.
2994 	 * Actually this is the number of iterations to spawn.
2995 	 */
2996 	unsigned int ns = 0;
2997 	/*
2998 	 * Bonding device
2999 	 *   < 0 - no bonding device (single one)
3000 	 *  >= 0 - bonding device (value is slave PF index)
3001 	 */
3002 	int bd = -1;
3003 	struct mlx5_dev_spawn_data *list = NULL;
3004 	struct mlx5_dev_config dev_config;
3005 	int ret;
3006 
3007 	if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
3008 		DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
3009 			" driver.");
3010 		return 1;
3011 	}
3012 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
3013 		mlx5_pmd_socket_init();
3014 	ret = mlx5_init_once();
3015 	if (ret) {
3016 		DRV_LOG(ERR, "unable to init PMD global data: %s",
3017 			strerror(rte_errno));
3018 		return -rte_errno;
3019 	}
3020 	MLX5_ASSERT(pci_drv == &mlx5_driver);
3021 	errno = 0;
3022 	ibv_list = mlx5_glue->get_device_list(&ret);
3023 	if (!ibv_list) {
3024 		rte_errno = errno ? errno : ENOSYS;
3025 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
3026 		return -rte_errno;
3027 	}
3028 	/*
3029 	 * First scan the list of all Infiniband devices to find
3030 	 * matching ones, gathering into the list.
3031 	 */
3032 	struct ibv_device *ibv_match[ret + 1];
3033 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3034 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3035 	unsigned int i;
3036 
3037 	while (ret-- > 0) {
3038 		struct rte_pci_addr pci_addr;
3039 
3040 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3041 		bd = mlx5_device_bond_pci_match
3042 				(ibv_list[ret], pci_dev, nl_rdma);
3043 		if (bd >= 0) {
3044 			/*
3045 			 * Bonding device detected. Only one match is allowed,
3046 			 * the bonding is supported over multi-port IB device,
3047 			 * there should be no matches on representor PCI
3048 			 * functions or non VF LAG bonding devices with
3049 			 * specified address.
3050 			 */
3051 			if (nd) {
3052 				DRV_LOG(ERR,
3053 					"multiple PCI match on bonding device"
3054 					"\"%s\" found", ibv_list[ret]->name);
3055 				rte_errno = ENOENT;
3056 				ret = -rte_errno;
3057 				goto exit;
3058 			}
3059 			DRV_LOG(INFO, "PCI information matches for"
3060 				      " slave %d bonding device \"%s\"",
3061 				      bd, ibv_list[ret]->name);
3062 			ibv_match[nd++] = ibv_list[ret];
3063 			break;
3064 		}
3065 		if (mlx5_dev_to_pci_addr
3066 			(ibv_list[ret]->ibdev_path, &pci_addr))
3067 			continue;
3068 		if (pci_dev->addr.domain != pci_addr.domain ||
3069 		    pci_dev->addr.bus != pci_addr.bus ||
3070 		    pci_dev->addr.devid != pci_addr.devid ||
3071 		    pci_dev->addr.function != pci_addr.function)
3072 			continue;
3073 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3074 			ibv_list[ret]->name);
3075 		ibv_match[nd++] = ibv_list[ret];
3076 	}
3077 	ibv_match[nd] = NULL;
3078 	if (!nd) {
3079 		/* No device matches, just complain and bail out. */
3080 		DRV_LOG(WARNING,
3081 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
3082 			" are kernel drivers loaded?",
3083 			pci_dev->addr.domain, pci_dev->addr.bus,
3084 			pci_dev->addr.devid, pci_dev->addr.function);
3085 		rte_errno = ENOENT;
3086 		ret = -rte_errno;
3087 		goto exit;
3088 	}
3089 	if (nd == 1) {
3090 		/*
3091 		 * Found single matching device may have multiple ports.
3092 		 * Each port may be representor, we have to check the port
3093 		 * number and check the representors existence.
3094 		 */
3095 		if (nl_rdma >= 0)
3096 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3097 		if (!np)
3098 			DRV_LOG(WARNING, "can not get IB device \"%s\""
3099 					 " ports number", ibv_match[0]->name);
3100 		if (bd >= 0 && !np) {
3101 			DRV_LOG(ERR, "can not get ports"
3102 				     " for bonding device");
3103 			rte_errno = ENOENT;
3104 			ret = -rte_errno;
3105 			goto exit;
3106 		}
3107 	}
3108 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3109 	if (bd >= 0) {
3110 		/*
3111 		 * This may happen if there is VF LAG kernel support and
3112 		 * application is compiled with older rdma_core library.
3113 		 */
3114 		DRV_LOG(ERR,
3115 			"No kernel/verbs support for VF LAG bonding found.");
3116 		rte_errno = ENOTSUP;
3117 		ret = -rte_errno;
3118 		goto exit;
3119 	}
3120 #endif
3121 	/*
3122 	 * Now we can determine the maximal
3123 	 * amount of devices to be spawned.
3124 	 */
3125 	list = rte_zmalloc("device spawn data",
3126 			 sizeof(struct mlx5_dev_spawn_data) *
3127 			 (np ? np : nd),
3128 			 RTE_CACHE_LINE_SIZE);
3129 	if (!list) {
3130 		DRV_LOG(ERR, "spawn data array allocation failure");
3131 		rte_errno = ENOMEM;
3132 		ret = -rte_errno;
3133 		goto exit;
3134 	}
3135 	if (bd >= 0 || np > 1) {
3136 		/*
3137 		 * Single IB device with multiple ports found,
3138 		 * it may be E-Switch master device and representors.
3139 		 * We have to perform identification trough the ports.
3140 		 */
3141 		MLX5_ASSERT(nl_rdma >= 0);
3142 		MLX5_ASSERT(ns == 0);
3143 		MLX5_ASSERT(nd == 1);
3144 		MLX5_ASSERT(np);
3145 		for (i = 1; i <= np; ++i) {
3146 			list[ns].max_port = np;
3147 			list[ns].ibv_port = i;
3148 			list[ns].ibv_dev = ibv_match[0];
3149 			list[ns].eth_dev = NULL;
3150 			list[ns].pci_dev = pci_dev;
3151 			list[ns].pf_bond = bd;
3152 			list[ns].ifindex = mlx5_nl_ifindex
3153 					(nl_rdma, list[ns].ibv_dev->name, i);
3154 			if (!list[ns].ifindex) {
3155 				/*
3156 				 * No network interface index found for the
3157 				 * specified port, it means there is no
3158 				 * representor on this port. It's OK,
3159 				 * there can be disabled ports, for example
3160 				 * if sriov_numvfs < sriov_totalvfs.
3161 				 */
3162 				continue;
3163 			}
3164 			ret = -1;
3165 			if (nl_route >= 0)
3166 				ret = mlx5_nl_switch_info
3167 					       (nl_route,
3168 						list[ns].ifindex,
3169 						&list[ns].info);
3170 			if (ret || (!list[ns].info.representor &&
3171 				    !list[ns].info.master)) {
3172 				/*
3173 				 * We failed to recognize representors with
3174 				 * Netlink, let's try to perform the task
3175 				 * with sysfs.
3176 				 */
3177 				ret =  mlx5_sysfs_switch_info
3178 						(list[ns].ifindex,
3179 						 &list[ns].info);
3180 			}
3181 			if (!ret && bd >= 0) {
3182 				switch (list[ns].info.name_type) {
3183 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3184 					if (list[ns].info.port_name == bd)
3185 						ns++;
3186 					break;
3187 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3188 					if (list[ns].info.pf_num == bd)
3189 						ns++;
3190 					break;
3191 				default:
3192 					break;
3193 				}
3194 				continue;
3195 			}
3196 			if (!ret && (list[ns].info.representor ^
3197 				     list[ns].info.master))
3198 				ns++;
3199 		}
3200 		if (!ns) {
3201 			DRV_LOG(ERR,
3202 				"unable to recognize master/representors"
3203 				" on the IB device with multiple ports");
3204 			rte_errno = ENOENT;
3205 			ret = -rte_errno;
3206 			goto exit;
3207 		}
3208 	} else {
3209 		/*
3210 		 * The existence of several matching entries (nd > 1) means
3211 		 * port representors have been instantiated. No existing Verbs
3212 		 * call nor sysfs entries can tell them apart, this can only
3213 		 * be done through Netlink calls assuming kernel drivers are
3214 		 * recent enough to support them.
3215 		 *
3216 		 * In the event of identification failure through Netlink,
3217 		 * try again through sysfs, then:
3218 		 *
3219 		 * 1. A single IB device matches (nd == 1) with single
3220 		 *    port (np=0/1) and is not a representor, assume
3221 		 *    no switch support.
3222 		 *
3223 		 * 2. Otherwise no safe assumptions can be made;
3224 		 *    complain louder and bail out.
3225 		 */
3226 		np = 1;
3227 		for (i = 0; i != nd; ++i) {
3228 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3229 			list[ns].max_port = 1;
3230 			list[ns].ibv_port = 1;
3231 			list[ns].ibv_dev = ibv_match[i];
3232 			list[ns].eth_dev = NULL;
3233 			list[ns].pci_dev = pci_dev;
3234 			list[ns].pf_bond = -1;
3235 			list[ns].ifindex = 0;
3236 			if (nl_rdma >= 0)
3237 				list[ns].ifindex = mlx5_nl_ifindex
3238 					(nl_rdma, list[ns].ibv_dev->name, 1);
3239 			if (!list[ns].ifindex) {
3240 				char ifname[IF_NAMESIZE];
3241 
3242 				/*
3243 				 * Netlink failed, it may happen with old
3244 				 * ib_core kernel driver (before 4.16).
3245 				 * We can assume there is old driver because
3246 				 * here we are processing single ports IB
3247 				 * devices. Let's try sysfs to retrieve
3248 				 * the ifindex. The method works for
3249 				 * master device only.
3250 				 */
3251 				if (nd > 1) {
3252 					/*
3253 					 * Multiple devices found, assume
3254 					 * representors, can not distinguish
3255 					 * master/representor and retrieve
3256 					 * ifindex via sysfs.
3257 					 */
3258 					continue;
3259 				}
3260 				ret = mlx5_get_master_ifname
3261 					(ibv_match[i]->ibdev_path, &ifname);
3262 				if (!ret)
3263 					list[ns].ifindex =
3264 						if_nametoindex(ifname);
3265 				if (!list[ns].ifindex) {
3266 					/*
3267 					 * No network interface index found
3268 					 * for the specified device, it means
3269 					 * there it is neither representor
3270 					 * nor master.
3271 					 */
3272 					continue;
3273 				}
3274 			}
3275 			ret = -1;
3276 			if (nl_route >= 0)
3277 				ret = mlx5_nl_switch_info
3278 					       (nl_route,
3279 						list[ns].ifindex,
3280 						&list[ns].info);
3281 			if (ret || (!list[ns].info.representor &&
3282 				    !list[ns].info.master)) {
3283 				/*
3284 				 * We failed to recognize representors with
3285 				 * Netlink, let's try to perform the task
3286 				 * with sysfs.
3287 				 */
3288 				ret =  mlx5_sysfs_switch_info
3289 						(list[ns].ifindex,
3290 						 &list[ns].info);
3291 			}
3292 			if (!ret && (list[ns].info.representor ^
3293 				     list[ns].info.master)) {
3294 				ns++;
3295 			} else if ((nd == 1) &&
3296 				   !list[ns].info.representor &&
3297 				   !list[ns].info.master) {
3298 				/*
3299 				 * Single IB device with
3300 				 * one physical port and
3301 				 * attached network device.
3302 				 * May be SRIOV is not enabled
3303 				 * or there is no representors.
3304 				 */
3305 				DRV_LOG(INFO, "no E-Switch support detected");
3306 				ns++;
3307 				break;
3308 			}
3309 		}
3310 		if (!ns) {
3311 			DRV_LOG(ERR,
3312 				"unable to recognize master/representors"
3313 				" on the multiple IB devices");
3314 			rte_errno = ENOENT;
3315 			ret = -rte_errno;
3316 			goto exit;
3317 		}
3318 	}
3319 	MLX5_ASSERT(ns);
3320 	/*
3321 	 * Sort list to probe devices in natural order for users convenience
3322 	 * (i.e. master first, then representors from lowest to highest ID).
3323 	 */
3324 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3325 	/* Default configuration. */
3326 	dev_config = (struct mlx5_dev_config){
3327 		.hw_padding = 0,
3328 		.mps = MLX5_ARG_UNSET,
3329 		.dbnc = MLX5_ARG_UNSET,
3330 		.rx_vec_en = 1,
3331 		.txq_inline_max = MLX5_ARG_UNSET,
3332 		.txq_inline_min = MLX5_ARG_UNSET,
3333 		.txq_inline_mpw = MLX5_ARG_UNSET,
3334 		.txqs_inline = MLX5_ARG_UNSET,
3335 		.vf_nl_en = 1,
3336 		.mr_ext_memseg_en = 1,
3337 		.mprq = {
3338 			.enabled = 0, /* Disabled by default. */
3339 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3340 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3341 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3342 		},
3343 		.dv_esw_en = 1,
3344 		.dv_flow_en = 1,
3345 	};
3346 	/* Device specific configuration. */
3347 	switch (pci_dev->id.device_id) {
3348 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3349 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3350 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3351 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3352 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3353 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3354 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3355 		dev_config.vf = 1;
3356 		break;
3357 	default:
3358 		break;
3359 	}
3360 	for (i = 0; i != ns; ++i) {
3361 		uint32_t restore;
3362 
3363 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3364 						 &list[i],
3365 						 dev_config);
3366 		if (!list[i].eth_dev) {
3367 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3368 				break;
3369 			/* Device is disabled or already spawned. Ignore it. */
3370 			continue;
3371 		}
3372 		restore = list[i].eth_dev->data->dev_flags;
3373 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3374 		/* Restore non-PCI flags cleared by the above call. */
3375 		list[i].eth_dev->data->dev_flags |= restore;
3376 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3377 		rte_eth_dev_probing_finish(list[i].eth_dev);
3378 	}
3379 	if (i != ns) {
3380 		DRV_LOG(ERR,
3381 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3382 			" encountering an error: %s",
3383 			pci_dev->addr.domain, pci_dev->addr.bus,
3384 			pci_dev->addr.devid, pci_dev->addr.function,
3385 			strerror(rte_errno));
3386 		ret = -rte_errno;
3387 		/* Roll back. */
3388 		while (i--) {
3389 			if (!list[i].eth_dev)
3390 				continue;
3391 			mlx5_dev_close(list[i].eth_dev);
3392 			/* mac_addrs must not be freed because in dev_private */
3393 			list[i].eth_dev->data->mac_addrs = NULL;
3394 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3395 		}
3396 		/* Restore original error. */
3397 		rte_errno = -ret;
3398 	} else {
3399 		ret = 0;
3400 	}
3401 exit:
3402 	/*
3403 	 * Do the routine cleanup:
3404 	 * - close opened Netlink sockets
3405 	 * - free allocated spawn data array
3406 	 * - free the Infiniband device list
3407 	 */
3408 	if (nl_rdma >= 0)
3409 		close(nl_rdma);
3410 	if (nl_route >= 0)
3411 		close(nl_route);
3412 	if (list)
3413 		rte_free(list);
3414 	MLX5_ASSERT(ibv_list);
3415 	mlx5_glue->free_device_list(ibv_list);
3416 	return ret;
3417 }
3418 
3419 /**
3420  * Look for the ethernet device belonging to mlx5 driver.
3421  *
3422  * @param[in] port_id
3423  *   port_id to start looking for device.
3424  * @param[in] pci_dev
3425  *   Pointer to the hint PCI device. When device is being probed
3426  *   the its siblings (master and preceding representors might
3427  *   not have assigned driver yet (because the mlx5_pci_probe()
3428  *   is not completed yet, for this case match on hint PCI
3429  *   device may be used to detect sibling device.
3430  *
3431  * @return
3432  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3433  */
3434 uint16_t
3435 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3436 {
3437 	while (port_id < RTE_MAX_ETHPORTS) {
3438 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3439 
3440 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3441 		    dev->device &&
3442 		    (dev->device == &pci_dev->device ||
3443 		     (dev->device->driver &&
3444 		     dev->device->driver->name &&
3445 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3446 			break;
3447 		port_id++;
3448 	}
3449 	if (port_id >= RTE_MAX_ETHPORTS)
3450 		return RTE_MAX_ETHPORTS;
3451 	return port_id;
3452 }
3453 
3454 /**
3455  * DPDK callback to remove a PCI device.
3456  *
3457  * This function removes all Ethernet devices belong to a given PCI device.
3458  *
3459  * @param[in] pci_dev
3460  *   Pointer to the PCI device.
3461  *
3462  * @return
3463  *   0 on success, the function cannot fail.
3464  */
3465 static int
3466 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3467 {
3468 	uint16_t port_id;
3469 
3470 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3471 		rte_eth_dev_close(port_id);
3472 	return 0;
3473 }
3474 
3475 static const struct rte_pci_id mlx5_pci_id_map[] = {
3476 	{
3477 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3478 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3479 	},
3480 	{
3481 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3482 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3483 	},
3484 	{
3485 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3486 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3487 	},
3488 	{
3489 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3490 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3491 	},
3492 	{
3493 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3494 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3495 	},
3496 	{
3497 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3498 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3499 	},
3500 	{
3501 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3502 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3503 	},
3504 	{
3505 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3506 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3507 	},
3508 	{
3509 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3510 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3511 	},
3512 	{
3513 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3514 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3515 	},
3516 	{
3517 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3518 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3519 	},
3520 	{
3521 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3522 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3523 	},
3524 	{
3525 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3526 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3527 	},
3528 	{
3529 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3530 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3531 	},
3532 	{
3533 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3534 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
3535 	},
3536 	{
3537 		.vendor_id = 0
3538 	}
3539 };
3540 
3541 static struct rte_pci_driver mlx5_driver = {
3542 	.driver = {
3543 		.name = MLX5_DRIVER_NAME
3544 	},
3545 	.id_table = mlx5_pci_id_map,
3546 	.probe = mlx5_pci_probe,
3547 	.remove = mlx5_pci_remove,
3548 	.dma_map = mlx5_dma_map,
3549 	.dma_unmap = mlx5_dma_unmap,
3550 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3551 		     RTE_PCI_DRV_PROBE_AGAIN,
3552 };
3553 
3554 /**
3555  * Driver initialization routine.
3556  */
3557 RTE_INIT(rte_mlx5_pmd_init)
3558 {
3559 	/* Initialize driver log type. */
3560 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3561 	if (mlx5_logtype >= 0)
3562 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3563 
3564 	/* Build the static tables for Verbs conversion. */
3565 	mlx5_set_ptype_table();
3566 	mlx5_set_cksum_table();
3567 	mlx5_set_swp_types_table();
3568 	if (mlx5_glue)
3569 		rte_pci_register(&mlx5_driver);
3570 }
3571 
3572 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3573 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3574 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3575