xref: /dpdk/drivers/net/mlx5/mlx5.c (revision fdf7471cccb8be023037c218d1402c0549eb2c8e)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <assert.h>
10 #include <dlfcn.h>
11 #include <stdint.h>
12 #include <stdlib.h>
13 #include <errno.h>
14 #include <net/if.h>
15 #include <sys/mman.h>
16 #include <linux/rtnetlink.h>
17 
18 /* Verbs header. */
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27 
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
31 #include <rte_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_kvargs.h>
36 #include <rte_rwlock.h>
37 #include <rte_spinlock.h>
38 #include <rte_string_fns.h>
39 #include <rte_alarm.h>
40 
41 #include "mlx5.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_autoconf.h"
45 #include "mlx5_defs.h"
46 #include "mlx5_glue.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49 
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52 
53 /* Device parameter to enable RX completion entry padding to 128B. */
54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
55 
56 /* Device parameter to enable padding Rx packet to cacheline size. */
57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
58 
59 /* Device parameter to enable Multi-Packet Rx queue. */
60 #define MLX5_RX_MPRQ_EN "mprq_en"
61 
62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
64 
65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
67 
68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
70 
71 /* Device parameter to configure inline send. Deprecated, ignored.*/
72 #define MLX5_TXQ_INLINE "txq_inline"
73 
74 /* Device parameter to limit packet size to inline with ordinary SEND. */
75 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
76 
77 /* Device parameter to configure minimal data size to inline. */
78 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
79 
80 /* Device parameter to limit packet size to inline with Enhanced MPW. */
81 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
82 
83 /*
84  * Device parameter to configure the number of TX queues threshold for
85  * enabling inline send.
86  */
87 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
88 
89 /*
90  * Device parameter to configure the number of TX queues threshold for
91  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
92  */
93 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
94 
95 /* Device parameter to enable multi-packet send WQEs. */
96 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
97 
98 /*
99  * Device parameter to force doorbell register mapping
100  * to non-cahed region eliminating the extra write memory barrier.
101  */
102 #define MLX5_TX_DB_NC "tx_db_nc"
103 
104 /*
105  * Device parameter to include 2 dsegs in the title WQEBB.
106  * Deprecated, ignored.
107  */
108 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
109 
110 /*
111  * Device parameter to limit the size of inlining packet.
112  * Deprecated, ignored.
113  */
114 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
115 
116 /*
117  * Device parameter to enable hardware Tx vector.
118  * Deprecated, ignored (no vectorized Tx routines anymore).
119  */
120 #define MLX5_TX_VEC_EN "tx_vec_en"
121 
122 /* Device parameter to enable hardware Rx vector. */
123 #define MLX5_RX_VEC_EN "rx_vec_en"
124 
125 /* Allow L3 VXLAN flow creation. */
126 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
127 
128 /* Activate DV E-Switch flow steering. */
129 #define MLX5_DV_ESW_EN "dv_esw_en"
130 
131 /* Activate DV flow steering. */
132 #define MLX5_DV_FLOW_EN "dv_flow_en"
133 
134 /* Enable extensive flow metadata support. */
135 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
136 
137 /* Activate Netlink support in VF mode. */
138 #define MLX5_VF_NL_EN "vf_nl_en"
139 
140 /* Enable extending memsegs when creating a MR. */
141 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
142 
143 /* Select port representors to instantiate. */
144 #define MLX5_REPRESENTOR "representor"
145 
146 /* Device parameter to configure the maximum number of dump files per queue. */
147 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
148 
149 /* Configure timeout of LRO session (in microseconds). */
150 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
151 
152 #ifndef HAVE_IBV_MLX5_MOD_MPW
153 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
154 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
155 #endif
156 
157 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
158 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
159 #endif
160 
161 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
162 
163 /* Shared memory between primary and secondary processes. */
164 struct mlx5_shared_data *mlx5_shared_data;
165 
166 /* Spinlock for mlx5_shared_data allocation. */
167 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
168 
169 /* Process local data for secondary processes. */
170 static struct mlx5_local_data mlx5_local_data;
171 
172 /** Driver-specific log messages type. */
173 int mlx5_logtype;
174 
175 /** Data associated with devices to spawn. */
176 struct mlx5_dev_spawn_data {
177 	uint32_t ifindex; /**< Network interface index. */
178 	uint32_t max_port; /**< IB device maximal port index. */
179 	uint32_t ibv_port; /**< IB device physical port index. */
180 	int pf_bond; /**< bonding device PF index. < 0 - no bonding */
181 	struct mlx5_switch_info info; /**< Switch information. */
182 	struct ibv_device *ibv_dev; /**< Associated IB device. */
183 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
184 	struct rte_pci_device *pci_dev; /**< Backend PCI device. */
185 };
186 
187 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
188 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
189 
190 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
191 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
192 
193 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
194 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
195 
196 /**
197  * Allocate ID pool structure.
198  *
199  * @return
200  *   Pointer to pool object, NULL value otherwise.
201  */
202 struct mlx5_flow_id_pool *
203 mlx5_flow_id_pool_alloc(void)
204 {
205 	struct mlx5_flow_id_pool *pool;
206 	void *mem;
207 
208 	pool = rte_zmalloc("id pool allocation", sizeof(*pool),
209 			   RTE_CACHE_LINE_SIZE);
210 	if (!pool) {
211 		DRV_LOG(ERR, "can't allocate id pool");
212 		rte_errno  = ENOMEM;
213 		return NULL;
214 	}
215 	mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
216 			  RTE_CACHE_LINE_SIZE);
217 	if (!mem) {
218 		DRV_LOG(ERR, "can't allocate mem for id pool");
219 		rte_errno  = ENOMEM;
220 		goto error;
221 	}
222 	pool->free_arr = mem;
223 	pool->curr = pool->free_arr;
224 	pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
225 	pool->base_index = 0;
226 	return pool;
227 error:
228 	rte_free(pool);
229 	return NULL;
230 }
231 
232 /**
233  * Release ID pool structure.
234  *
235  * @param[in] pool
236  *   Pointer to flow id pool object to free.
237  */
238 void
239 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
240 {
241 	rte_free(pool->free_arr);
242 	rte_free(pool);
243 }
244 
245 /**
246  * Generate ID.
247  *
248  * @param[in] pool
249  *   Pointer to flow id pool.
250  * @param[out] id
251  *   The generated ID.
252  *
253  * @return
254  *   0 on success, error value otherwise.
255  */
256 uint32_t
257 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
258 {
259 	if (pool->curr == pool->free_arr) {
260 		if (pool->base_index == UINT32_MAX) {
261 			rte_errno  = ENOMEM;
262 			DRV_LOG(ERR, "no free id");
263 			return -rte_errno;
264 		}
265 		*id = ++pool->base_index;
266 		return 0;
267 	}
268 	*id = *(--pool->curr);
269 	return 0;
270 }
271 
272 /**
273  * Release ID.
274  *
275  * @param[in] pool
276  *   Pointer to flow id pool.
277  * @param[out] id
278  *   The generated ID.
279  *
280  * @return
281  *   0 on success, error value otherwise.
282  */
283 uint32_t
284 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
285 {
286 	uint32_t size;
287 	uint32_t size2;
288 	void *mem;
289 
290 	if (pool->curr == pool->last) {
291 		size = pool->curr - pool->free_arr;
292 		size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
293 		assert(size2 > size);
294 		mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
295 		if (!mem) {
296 			DRV_LOG(ERR, "can't allocate mem for id pool");
297 			rte_errno  = ENOMEM;
298 			return -rte_errno;
299 		}
300 		memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
301 		rte_free(pool->free_arr);
302 		pool->free_arr = mem;
303 		pool->curr = pool->free_arr + size;
304 		pool->last = pool->free_arr + size2;
305 	}
306 	*pool->curr = id;
307 	pool->curr++;
308 	return 0;
309 }
310 
311 /**
312  * Initialize the counters management structure.
313  *
314  * @param[in] sh
315  *   Pointer to mlx5_ibv_shared object to free
316  */
317 static void
318 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
319 {
320 	uint8_t i;
321 
322 	TAILQ_INIT(&sh->cmng.flow_counters);
323 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
324 		TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
325 }
326 
327 /**
328  * Destroy all the resources allocated for a counter memory management.
329  *
330  * @param[in] mng
331  *   Pointer to the memory management structure.
332  */
333 static void
334 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
335 {
336 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
337 
338 	LIST_REMOVE(mng, next);
339 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
340 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
341 	rte_free(mem);
342 }
343 
344 /**
345  * Close and release all the resources of the counters management.
346  *
347  * @param[in] sh
348  *   Pointer to mlx5_ibv_shared object to free.
349  */
350 static void
351 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
352 {
353 	struct mlx5_counter_stats_mem_mng *mng;
354 	uint8_t i;
355 	int j;
356 	int retries = 1024;
357 
358 	rte_errno = 0;
359 	while (--retries) {
360 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
361 		if (rte_errno != EINPROGRESS)
362 			break;
363 		rte_pause();
364 	}
365 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
366 		struct mlx5_flow_counter_pool *pool;
367 		uint32_t batch = !!(i % 2);
368 
369 		if (!sh->cmng.ccont[i].pools)
370 			continue;
371 		pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
372 		while (pool) {
373 			if (batch) {
374 				if (pool->min_dcs)
375 					claim_zero
376 					(mlx5_devx_cmd_destroy(pool->min_dcs));
377 			}
378 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
379 				if (pool->counters_raw[j].action)
380 					claim_zero
381 					(mlx5_glue->destroy_flow_action
382 					       (pool->counters_raw[j].action));
383 				if (!batch && pool->counters_raw[j].dcs)
384 					claim_zero(mlx5_devx_cmd_destroy
385 						  (pool->counters_raw[j].dcs));
386 			}
387 			TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
388 				     next);
389 			rte_free(pool);
390 			pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
391 		}
392 		rte_free(sh->cmng.ccont[i].pools);
393 	}
394 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
395 	while (mng) {
396 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
397 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
398 	}
399 	memset(&sh->cmng, 0, sizeof(sh->cmng));
400 }
401 
402 /**
403  * Extract pdn of PD object using DV API.
404  *
405  * @param[in] pd
406  *   Pointer to the verbs PD object.
407  * @param[out] pdn
408  *   Pointer to the PD object number variable.
409  *
410  * @return
411  *   0 on success, error value otherwise.
412  */
413 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
414 static int
415 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
416 {
417 	struct mlx5dv_obj obj;
418 	struct mlx5dv_pd pd_info;
419 	int ret = 0;
420 
421 	obj.pd.in = pd;
422 	obj.pd.out = &pd_info;
423 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
424 	if (ret) {
425 		DRV_LOG(DEBUG, "Fail to get PD object info");
426 		return ret;
427 	}
428 	*pdn = pd_info.pdn;
429 	return 0;
430 }
431 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
432 
433 static int
434 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
435 {
436 	char *env;
437 	int value;
438 
439 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
440 	/* Get environment variable to store. */
441 	env = getenv(MLX5_SHUT_UP_BF);
442 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
443 	if (config->dbnc == MLX5_ARG_UNSET)
444 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
445 	else
446 		setenv(MLX5_SHUT_UP_BF, config->dbnc ? "1" : "0", 1);
447 	return value;
448 }
449 
450 static void
451 mlx5_restore_doorbell_mapping_env(const struct mlx5_dev_config *config,
452 				  int value)
453 {
454 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
455 	if (config->dbnc == MLX5_ARG_UNSET)
456 		return;
457 	/* Restore the original environment variable state. */
458 	if (value == MLX5_ARG_UNSET)
459 		unsetenv(MLX5_SHUT_UP_BF);
460 	else
461 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
462 }
463 
464 /**
465  * Allocate shared IB device context. If there is multiport device the
466  * master and representors will share this context, if there is single
467  * port dedicated IB device, the context will be used by only given
468  * port due to unification.
469  *
470  * Routine first searches the context for the specified IB device name,
471  * if found the shared context assumed and reference counter is incremented.
472  * If no context found the new one is created and initialized with specified
473  * IB device context and parameters.
474  *
475  * @param[in] spawn
476  *   Pointer to the IB device attributes (name, port, etc).
477  * @param[in] config
478  *   Pointer to device configuration structure.
479  *
480  * @return
481  *   Pointer to mlx5_ibv_shared object on success,
482  *   otherwise NULL and rte_errno is set.
483  */
484 static struct mlx5_ibv_shared *
485 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
486 			const struct mlx5_dev_config *config)
487 {
488 	struct mlx5_ibv_shared *sh;
489 	int dbmap_env;
490 	int err = 0;
491 	uint32_t i;
492 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
493 	struct mlx5_devx_tis_attr tis_attr = { 0 };
494 #endif
495 
496 	assert(spawn);
497 	/* Secondary process should not create the shared context. */
498 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
499 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
500 	/* Search for IB context by device name. */
501 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
502 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
503 			sh->refcnt++;
504 			goto exit;
505 		}
506 	}
507 	/* No device found, we have to create new shared context. */
508 	assert(spawn->max_port);
509 	sh = rte_zmalloc("ethdev shared ib context",
510 			 sizeof(struct mlx5_ibv_shared) +
511 			 spawn->max_port *
512 			 sizeof(struct mlx5_ibv_shared_port),
513 			 RTE_CACHE_LINE_SIZE);
514 	if (!sh) {
515 		DRV_LOG(ERR, "shared context allocation failure");
516 		rte_errno  = ENOMEM;
517 		goto exit;
518 	}
519 	/*
520 	 * Configure environment variable "MLX5_BF_SHUT_UP"
521 	 * before the device creation. The rdma_core library
522 	 * checks the variable at device creation and
523 	 * stores the result internally.
524 	 */
525 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
526 	/* Try to open IB device with DV first, then usual Verbs. */
527 	errno = 0;
528 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
529 	if (sh->ctx) {
530 		sh->devx = 1;
531 		DRV_LOG(DEBUG, "DevX is supported");
532 		/* The device is created, no need for environment. */
533 		mlx5_restore_doorbell_mapping_env(config, dbmap_env);
534 	} else {
535 		/* The environment variable is still configured. */
536 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
537 		err = errno ? errno : ENODEV;
538 		/*
539 		 * The environment variable is not needed anymore,
540 		 * all device creation attempts are completed.
541 		 */
542 		mlx5_restore_doorbell_mapping_env(config, dbmap_env);
543 		if (!sh->ctx) {
544 			goto error;
545 		}
546 		DRV_LOG(DEBUG, "DevX is NOT supported");
547 	}
548 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
549 	if (err) {
550 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
551 		goto error;
552 	}
553 	sh->refcnt = 1;
554 	sh->max_port = spawn->max_port;
555 	strncpy(sh->ibdev_name, sh->ctx->device->name,
556 		sizeof(sh->ibdev_name));
557 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
558 		sizeof(sh->ibdev_path));
559 	pthread_mutex_init(&sh->intr_mutex, NULL);
560 	/*
561 	 * Setting port_id to max unallowed value means
562 	 * there is no interrupt subhandler installed for
563 	 * the given port index i.
564 	 */
565 	for (i = 0; i < sh->max_port; i++) {
566 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
567 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
568 	}
569 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
570 	if (sh->pd == NULL) {
571 		DRV_LOG(ERR, "PD allocation failure");
572 		err = ENOMEM;
573 		goto error;
574 	}
575 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
576 	if (sh->devx) {
577 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
578 		if (err) {
579 			DRV_LOG(ERR, "Fail to extract pdn from PD");
580 			goto error;
581 		}
582 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
583 		if (!sh->td) {
584 			DRV_LOG(ERR, "TD allocation failure");
585 			err = ENOMEM;
586 			goto error;
587 		}
588 		tis_attr.transport_domain = sh->td->id;
589 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
590 		if (!sh->tis) {
591 			DRV_LOG(ERR, "TIS allocation failure");
592 			err = ENOMEM;
593 			goto error;
594 		}
595 	}
596 	sh->flow_id_pool = mlx5_flow_id_pool_alloc();
597 	if (!sh->flow_id_pool) {
598 		DRV_LOG(ERR, "can't create flow id pool");
599 		err = ENOMEM;
600 		goto error;
601 	}
602 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
603 	/*
604 	 * Once the device is added to the list of memory event
605 	 * callback, its global MR cache table cannot be expanded
606 	 * on the fly because of deadlock. If it overflows, lookup
607 	 * should be done by searching MR list linearly, which is slow.
608 	 *
609 	 * At this point the device is not added to the memory
610 	 * event list yet, context is just being created.
611 	 */
612 	err = mlx5_mr_btree_init(&sh->mr.cache,
613 				 MLX5_MR_BTREE_CACHE_N * 2,
614 				 spawn->pci_dev->device.numa_node);
615 	if (err) {
616 		err = rte_errno;
617 		goto error;
618 	}
619 	mlx5_flow_counters_mng_init(sh);
620 	/* Add device to memory callback list. */
621 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
622 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
623 			 sh, mem_event_cb);
624 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
625 	/* Add context to the global device list. */
626 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
627 exit:
628 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
629 	return sh;
630 error:
631 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
632 	assert(sh);
633 	if (sh->tis)
634 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
635 	if (sh->td)
636 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
637 	if (sh->pd)
638 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
639 	if (sh->ctx)
640 		claim_zero(mlx5_glue->close_device(sh->ctx));
641 	if (sh->flow_id_pool)
642 		mlx5_flow_id_pool_release(sh->flow_id_pool);
643 	rte_free(sh);
644 	assert(err > 0);
645 	rte_errno = err;
646 	return NULL;
647 }
648 
649 /**
650  * Free shared IB device context. Decrement counter and if zero free
651  * all allocated resources and close handles.
652  *
653  * @param[in] sh
654  *   Pointer to mlx5_ibv_shared object to free
655  */
656 static void
657 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
658 {
659 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
660 #ifndef NDEBUG
661 	/* Check the object presence in the list. */
662 	struct mlx5_ibv_shared *lctx;
663 
664 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
665 		if (lctx == sh)
666 			break;
667 	assert(lctx);
668 	if (lctx != sh) {
669 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
670 		goto exit;
671 	}
672 #endif
673 	assert(sh);
674 	assert(sh->refcnt);
675 	/* Secondary process should not free the shared context. */
676 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
677 	if (--sh->refcnt)
678 		goto exit;
679 	/* Release created Memory Regions. */
680 	mlx5_mr_release(sh);
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 	/* Remove context from the global device list. */
686 	LIST_REMOVE(sh, next);
687 	/*
688 	 *  Ensure there is no async event handler installed.
689 	 *  Only primary process handles async device events.
690 	 **/
691 	mlx5_flow_counters_mng_close(sh);
692 	assert(!sh->intr_cnt);
693 	if (sh->intr_cnt)
694 		mlx5_intr_callback_unregister
695 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
696 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
697 	if (sh->devx_intr_cnt) {
698 		if (sh->intr_handle_devx.fd)
699 			rte_intr_callback_unregister(&sh->intr_handle_devx,
700 					  mlx5_dev_interrupt_handler_devx, sh);
701 		if (sh->devx_comp)
702 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
703 	}
704 #endif
705 	pthread_mutex_destroy(&sh->intr_mutex);
706 	if (sh->pd)
707 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
708 	if (sh->tis)
709 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
710 	if (sh->td)
711 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
712 	if (sh->ctx)
713 		claim_zero(mlx5_glue->close_device(sh->ctx));
714 	if (sh->flow_id_pool)
715 		mlx5_flow_id_pool_release(sh->flow_id_pool);
716 	rte_free(sh);
717 exit:
718 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
719 }
720 
721 /**
722  * Initialize DR related data within private structure.
723  * Routine checks the reference counter and does actual
724  * resources creation/initialization only if counter is zero.
725  *
726  * @param[in] priv
727  *   Pointer to the private device data structure.
728  *
729  * @return
730  *   Zero on success, positive error code otherwise.
731  */
732 static int
733 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
734 {
735 #ifdef HAVE_MLX5DV_DR
736 	struct mlx5_ibv_shared *sh = priv->sh;
737 	int err = 0;
738 	void *domain;
739 	char s[MLX5_HLIST_NAMESIZE];
740 
741 	assert(sh);
742 	if (sh->dv_refcnt) {
743 		/* Shared DV/DR structures is already initialized. */
744 		sh->dv_refcnt++;
745 		priv->dr_shared = 1;
746 		return 0;
747 	}
748 	/* Reference counter is zero, we should initialize structures. */
749 	domain = mlx5_glue->dr_create_domain(sh->ctx,
750 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
751 	if (!domain) {
752 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
753 		err = errno;
754 		goto error;
755 	}
756 	sh->rx_domain = domain;
757 	domain = mlx5_glue->dr_create_domain(sh->ctx,
758 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
759 	if (!domain) {
760 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
761 		err = errno;
762 		goto error;
763 	}
764 	pthread_mutex_init(&sh->dv_mutex, NULL);
765 	sh->tx_domain = domain;
766 #ifdef HAVE_MLX5DV_DR_ESWITCH
767 	if (priv->config.dv_esw_en) {
768 		domain  = mlx5_glue->dr_create_domain
769 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
770 		if (!domain) {
771 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
772 			err = errno;
773 			goto error;
774 		}
775 		sh->fdb_domain = domain;
776 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
777 	}
778 #endif
779 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
780 	sh->flow_tbls = mlx5_hlist_create(s,
781 					  MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
782 	if (!sh->flow_tbls) {
783 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
784 		err = -ENOMEM;
785 		goto error;
786 	}
787 	/* create tags hash list table. */
788 	snprintf(s, sizeof(s), "%s_tags", priv->sh->ibdev_name);
789 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
790 	if (!sh->flow_tbls) {
791 		DRV_LOG(ERR, "tags with hash creation failed.\n");
792 		err = -ENOMEM;
793 		goto error;
794 	}
795 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
796 	sh->dv_refcnt++;
797 	priv->dr_shared = 1;
798 	return 0;
799 
800 error:
801        /* Rollback the created objects. */
802 	if (sh->rx_domain) {
803 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
804 		sh->rx_domain = NULL;
805 	}
806 	if (sh->tx_domain) {
807 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
808 		sh->tx_domain = NULL;
809 	}
810 	if (sh->fdb_domain) {
811 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
812 		sh->fdb_domain = NULL;
813 	}
814 	if (sh->flow_tbls) {
815 		mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
816 		sh->flow_tbls = NULL;
817 	}
818 	if (sh->esw_drop_action) {
819 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
820 		sh->esw_drop_action = NULL;
821 	}
822 	if (sh->pop_vlan_action) {
823 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
824 		sh->pop_vlan_action = NULL;
825 	}
826 	return err;
827 #else
828 	(void)priv;
829 	return 0;
830 #endif
831 }
832 
833 /**
834  * Destroy DR related data within private structure.
835  *
836  * @param[in] priv
837  *   Pointer to the private device data structure.
838  */
839 static void
840 mlx5_free_shared_dr(struct mlx5_priv *priv)
841 {
842 #ifdef HAVE_MLX5DV_DR
843 	struct mlx5_ibv_shared *sh;
844 
845 	if (!priv->dr_shared)
846 		return;
847 	priv->dr_shared = 0;
848 	sh = priv->sh;
849 	assert(sh);
850 	assert(sh->dv_refcnt);
851 	if (sh->dv_refcnt && --sh->dv_refcnt)
852 		return;
853 	if (sh->flow_tbls) {
854 		/* flow table entries should be handled properly before. */
855 		mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
856 		sh->flow_tbls = NULL;
857 	}
858 	if (sh->tag_table) {
859 		/* tags should be destroyed with flow before. */
860 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
861 		sh->tag_table = NULL;
862 	}
863 	if (sh->rx_domain) {
864 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
865 		sh->rx_domain = NULL;
866 	}
867 	if (sh->tx_domain) {
868 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
869 		sh->tx_domain = NULL;
870 	}
871 #ifdef HAVE_MLX5DV_DR_ESWITCH
872 	if (sh->fdb_domain) {
873 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
874 		sh->fdb_domain = NULL;
875 	}
876 	if (sh->esw_drop_action) {
877 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
878 		sh->esw_drop_action = NULL;
879 	}
880 #endif
881 	if (sh->pop_vlan_action) {
882 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
883 		sh->pop_vlan_action = NULL;
884 	}
885 	pthread_mutex_destroy(&sh->dv_mutex);
886 #else
887 	(void)priv;
888 #endif
889 }
890 
891 /**
892  * Initialize shared data between primary and secondary process.
893  *
894  * A memzone is reserved by primary process and secondary processes attach to
895  * the memzone.
896  *
897  * @return
898  *   0 on success, a negative errno value otherwise and rte_errno is set.
899  */
900 static int
901 mlx5_init_shared_data(void)
902 {
903 	const struct rte_memzone *mz;
904 	int ret = 0;
905 
906 	rte_spinlock_lock(&mlx5_shared_data_lock);
907 	if (mlx5_shared_data == NULL) {
908 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
909 			/* Allocate shared memory. */
910 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
911 						 sizeof(*mlx5_shared_data),
912 						 SOCKET_ID_ANY, 0);
913 			if (mz == NULL) {
914 				DRV_LOG(ERR,
915 					"Cannot allocate mlx5 shared data");
916 				ret = -rte_errno;
917 				goto error;
918 			}
919 			mlx5_shared_data = mz->addr;
920 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
921 			rte_spinlock_init(&mlx5_shared_data->lock);
922 		} else {
923 			/* Lookup allocated shared memory. */
924 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
925 			if (mz == NULL) {
926 				DRV_LOG(ERR,
927 					"Cannot attach mlx5 shared data");
928 				ret = -rte_errno;
929 				goto error;
930 			}
931 			mlx5_shared_data = mz->addr;
932 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
933 		}
934 	}
935 error:
936 	rte_spinlock_unlock(&mlx5_shared_data_lock);
937 	return ret;
938 }
939 
940 /**
941  * Retrieve integer value from environment variable.
942  *
943  * @param[in] name
944  *   Environment variable name.
945  *
946  * @return
947  *   Integer value, 0 if the variable is not set.
948  */
949 int
950 mlx5_getenv_int(const char *name)
951 {
952 	const char *val = getenv(name);
953 
954 	if (val == NULL)
955 		return 0;
956 	return atoi(val);
957 }
958 
959 /**
960  * Verbs callback to allocate a memory. This function should allocate the space
961  * according to the size provided residing inside a huge page.
962  * Please note that all allocation must respect the alignment from libmlx5
963  * (i.e. currently sysconf(_SC_PAGESIZE)).
964  *
965  * @param[in] size
966  *   The size in bytes of the memory to allocate.
967  * @param[in] data
968  *   A pointer to the callback data.
969  *
970  * @return
971  *   Allocated buffer, NULL otherwise and rte_errno is set.
972  */
973 static void *
974 mlx5_alloc_verbs_buf(size_t size, void *data)
975 {
976 	struct mlx5_priv *priv = data;
977 	void *ret;
978 	size_t alignment = sysconf(_SC_PAGESIZE);
979 	unsigned int socket = SOCKET_ID_ANY;
980 
981 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
982 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
983 
984 		socket = ctrl->socket;
985 	} else if (priv->verbs_alloc_ctx.type ==
986 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
987 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
988 
989 		socket = ctrl->socket;
990 	}
991 	assert(data != NULL);
992 	ret = rte_malloc_socket(__func__, size, alignment, socket);
993 	if (!ret && size)
994 		rte_errno = ENOMEM;
995 	return ret;
996 }
997 
998 /**
999  * Verbs callback to free a memory.
1000  *
1001  * @param[in] ptr
1002  *   A pointer to the memory to free.
1003  * @param[in] data
1004  *   A pointer to the callback data.
1005  */
1006 static void
1007 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1008 {
1009 	assert(data != NULL);
1010 	rte_free(ptr);
1011 }
1012 
1013 /**
1014  * DPDK callback to add udp tunnel port
1015  *
1016  * @param[in] dev
1017  *   A pointer to eth_dev
1018  * @param[in] udp_tunnel
1019  *   A pointer to udp tunnel
1020  *
1021  * @return
1022  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1023  */
1024 int
1025 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1026 			 struct rte_eth_udp_tunnel *udp_tunnel)
1027 {
1028 	assert(udp_tunnel != NULL);
1029 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1030 	    udp_tunnel->udp_port == 4789)
1031 		return 0;
1032 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1033 	    udp_tunnel->udp_port == 4790)
1034 		return 0;
1035 	return -ENOTSUP;
1036 }
1037 
1038 /**
1039  * Initialize process private data structure.
1040  *
1041  * @param dev
1042  *   Pointer to Ethernet device structure.
1043  *
1044  * @return
1045  *   0 on success, a negative errno value otherwise and rte_errno is set.
1046  */
1047 int
1048 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1049 {
1050 	struct mlx5_priv *priv = dev->data->dev_private;
1051 	struct mlx5_proc_priv *ppriv;
1052 	size_t ppriv_size;
1053 
1054 	/*
1055 	 * UAR register table follows the process private structure. BlueFlame
1056 	 * registers for Tx queues are stored in the table.
1057 	 */
1058 	ppriv_size =
1059 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1060 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1061 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1062 	if (!ppriv) {
1063 		rte_errno = ENOMEM;
1064 		return -rte_errno;
1065 	}
1066 	ppriv->uar_table_sz = ppriv_size;
1067 	dev->process_private = ppriv;
1068 	return 0;
1069 }
1070 
1071 /**
1072  * Un-initialize process private data structure.
1073  *
1074  * @param dev
1075  *   Pointer to Ethernet device structure.
1076  */
1077 static void
1078 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1079 {
1080 	if (!dev->process_private)
1081 		return;
1082 	rte_free(dev->process_private);
1083 	dev->process_private = NULL;
1084 }
1085 
1086 /**
1087  * DPDK callback to close the device.
1088  *
1089  * Destroy all queues and objects, free memory.
1090  *
1091  * @param dev
1092  *   Pointer to Ethernet device structure.
1093  */
1094 static void
1095 mlx5_dev_close(struct rte_eth_dev *dev)
1096 {
1097 	struct mlx5_priv *priv = dev->data->dev_private;
1098 	unsigned int i;
1099 	int ret;
1100 
1101 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1102 		dev->data->port_id,
1103 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1104 	/* In case mlx5_dev_stop() has not been called. */
1105 	mlx5_dev_interrupt_handler_uninstall(dev);
1106 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1107 	mlx5_traffic_disable(dev);
1108 	mlx5_flow_flush(dev, NULL);
1109 	mlx5_flow_meter_flush(dev, NULL);
1110 	/* Prevent crashes when queues are still in use. */
1111 	dev->rx_pkt_burst = removed_rx_burst;
1112 	dev->tx_pkt_burst = removed_tx_burst;
1113 	rte_wmb();
1114 	/* Disable datapath on secondary process. */
1115 	mlx5_mp_req_stop_rxtx(dev);
1116 	if (priv->rxqs != NULL) {
1117 		/* XXX race condition if mlx5_rx_burst() is still running. */
1118 		usleep(1000);
1119 		for (i = 0; (i != priv->rxqs_n); ++i)
1120 			mlx5_rxq_release(dev, i);
1121 		priv->rxqs_n = 0;
1122 		priv->rxqs = NULL;
1123 	}
1124 	if (priv->txqs != NULL) {
1125 		/* XXX race condition if mlx5_tx_burst() is still running. */
1126 		usleep(1000);
1127 		for (i = 0; (i != priv->txqs_n); ++i)
1128 			mlx5_txq_release(dev, i);
1129 		priv->txqs_n = 0;
1130 		priv->txqs = NULL;
1131 	}
1132 	mlx5_proc_priv_uninit(dev);
1133 	if (priv->mreg_cp_tbl)
1134 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1135 	mlx5_mprq_free_mp(dev);
1136 	mlx5_free_shared_dr(priv);
1137 	if (priv->rss_conf.rss_key != NULL)
1138 		rte_free(priv->rss_conf.rss_key);
1139 	if (priv->reta_idx != NULL)
1140 		rte_free(priv->reta_idx);
1141 	if (priv->config.vf)
1142 		mlx5_nl_mac_addr_flush(dev);
1143 	if (priv->nl_socket_route >= 0)
1144 		close(priv->nl_socket_route);
1145 	if (priv->nl_socket_rdma >= 0)
1146 		close(priv->nl_socket_rdma);
1147 	if (priv->vmwa_context)
1148 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1149 	if (priv->sh) {
1150 		/*
1151 		 * Free the shared context in last turn, because the cleanup
1152 		 * routines above may use some shared fields, like
1153 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1154 		 * ifindex if Netlink fails.
1155 		 */
1156 		mlx5_free_shared_ibctx(priv->sh);
1157 		priv->sh = NULL;
1158 	}
1159 	ret = mlx5_hrxq_verify(dev);
1160 	if (ret)
1161 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1162 			dev->data->port_id);
1163 	ret = mlx5_ind_table_obj_verify(dev);
1164 	if (ret)
1165 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1166 			dev->data->port_id);
1167 	ret = mlx5_rxq_obj_verify(dev);
1168 	if (ret)
1169 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1170 			dev->data->port_id);
1171 	ret = mlx5_rxq_verify(dev);
1172 	if (ret)
1173 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1174 			dev->data->port_id);
1175 	ret = mlx5_txq_obj_verify(dev);
1176 	if (ret)
1177 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1178 			dev->data->port_id);
1179 	ret = mlx5_txq_verify(dev);
1180 	if (ret)
1181 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1182 			dev->data->port_id);
1183 	ret = mlx5_flow_verify(dev);
1184 	if (ret)
1185 		DRV_LOG(WARNING, "port %u some flows still remain",
1186 			dev->data->port_id);
1187 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1188 		unsigned int c = 0;
1189 		uint16_t port_id;
1190 
1191 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1192 			struct mlx5_priv *opriv =
1193 				rte_eth_devices[port_id].data->dev_private;
1194 
1195 			if (!opriv ||
1196 			    opriv->domain_id != priv->domain_id ||
1197 			    &rte_eth_devices[port_id] == dev)
1198 				continue;
1199 			++c;
1200 			break;
1201 		}
1202 		if (!c)
1203 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1204 	}
1205 	memset(priv, 0, sizeof(*priv));
1206 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1207 	/*
1208 	 * Reset mac_addrs to NULL such that it is not freed as part of
1209 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1210 	 * it is freed when dev_private is freed.
1211 	 */
1212 	dev->data->mac_addrs = NULL;
1213 }
1214 
1215 const struct eth_dev_ops mlx5_dev_ops = {
1216 	.dev_configure = mlx5_dev_configure,
1217 	.dev_start = mlx5_dev_start,
1218 	.dev_stop = mlx5_dev_stop,
1219 	.dev_set_link_down = mlx5_set_link_down,
1220 	.dev_set_link_up = mlx5_set_link_up,
1221 	.dev_close = mlx5_dev_close,
1222 	.promiscuous_enable = mlx5_promiscuous_enable,
1223 	.promiscuous_disable = mlx5_promiscuous_disable,
1224 	.allmulticast_enable = mlx5_allmulticast_enable,
1225 	.allmulticast_disable = mlx5_allmulticast_disable,
1226 	.link_update = mlx5_link_update,
1227 	.stats_get = mlx5_stats_get,
1228 	.stats_reset = mlx5_stats_reset,
1229 	.xstats_get = mlx5_xstats_get,
1230 	.xstats_reset = mlx5_xstats_reset,
1231 	.xstats_get_names = mlx5_xstats_get_names,
1232 	.fw_version_get = mlx5_fw_version_get,
1233 	.dev_infos_get = mlx5_dev_infos_get,
1234 	.read_clock = mlx5_read_clock,
1235 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1236 	.vlan_filter_set = mlx5_vlan_filter_set,
1237 	.rx_queue_setup = mlx5_rx_queue_setup,
1238 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1239 	.tx_queue_setup = mlx5_tx_queue_setup,
1240 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1241 	.rx_queue_release = mlx5_rx_queue_release,
1242 	.tx_queue_release = mlx5_tx_queue_release,
1243 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1244 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1245 	.mac_addr_remove = mlx5_mac_addr_remove,
1246 	.mac_addr_add = mlx5_mac_addr_add,
1247 	.mac_addr_set = mlx5_mac_addr_set,
1248 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1249 	.mtu_set = mlx5_dev_set_mtu,
1250 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1251 	.vlan_offload_set = mlx5_vlan_offload_set,
1252 	.reta_update = mlx5_dev_rss_reta_update,
1253 	.reta_query = mlx5_dev_rss_reta_query,
1254 	.rss_hash_update = mlx5_rss_hash_update,
1255 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1256 	.filter_ctrl = mlx5_dev_filter_ctrl,
1257 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1258 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1259 	.rx_queue_count = mlx5_rx_queue_count,
1260 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1261 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1262 	.is_removed = mlx5_is_removed,
1263 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1264 	.get_module_info = mlx5_get_module_info,
1265 	.get_module_eeprom = mlx5_get_module_eeprom,
1266 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1267 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1268 };
1269 
1270 /* Available operations from secondary process. */
1271 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1272 	.stats_get = mlx5_stats_get,
1273 	.stats_reset = mlx5_stats_reset,
1274 	.xstats_get = mlx5_xstats_get,
1275 	.xstats_reset = mlx5_xstats_reset,
1276 	.xstats_get_names = mlx5_xstats_get_names,
1277 	.fw_version_get = mlx5_fw_version_get,
1278 	.dev_infos_get = mlx5_dev_infos_get,
1279 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1280 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1281 	.get_module_info = mlx5_get_module_info,
1282 	.get_module_eeprom = mlx5_get_module_eeprom,
1283 };
1284 
1285 /* Available operations in flow isolated mode. */
1286 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1287 	.dev_configure = mlx5_dev_configure,
1288 	.dev_start = mlx5_dev_start,
1289 	.dev_stop = mlx5_dev_stop,
1290 	.dev_set_link_down = mlx5_set_link_down,
1291 	.dev_set_link_up = mlx5_set_link_up,
1292 	.dev_close = mlx5_dev_close,
1293 	.promiscuous_enable = mlx5_promiscuous_enable,
1294 	.promiscuous_disable = mlx5_promiscuous_disable,
1295 	.allmulticast_enable = mlx5_allmulticast_enable,
1296 	.allmulticast_disable = mlx5_allmulticast_disable,
1297 	.link_update = mlx5_link_update,
1298 	.stats_get = mlx5_stats_get,
1299 	.stats_reset = mlx5_stats_reset,
1300 	.xstats_get = mlx5_xstats_get,
1301 	.xstats_reset = mlx5_xstats_reset,
1302 	.xstats_get_names = mlx5_xstats_get_names,
1303 	.fw_version_get = mlx5_fw_version_get,
1304 	.dev_infos_get = mlx5_dev_infos_get,
1305 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1306 	.vlan_filter_set = mlx5_vlan_filter_set,
1307 	.rx_queue_setup = mlx5_rx_queue_setup,
1308 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1309 	.tx_queue_setup = mlx5_tx_queue_setup,
1310 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1311 	.rx_queue_release = mlx5_rx_queue_release,
1312 	.tx_queue_release = mlx5_tx_queue_release,
1313 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1314 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1315 	.mac_addr_remove = mlx5_mac_addr_remove,
1316 	.mac_addr_add = mlx5_mac_addr_add,
1317 	.mac_addr_set = mlx5_mac_addr_set,
1318 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1319 	.mtu_set = mlx5_dev_set_mtu,
1320 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1321 	.vlan_offload_set = mlx5_vlan_offload_set,
1322 	.filter_ctrl = mlx5_dev_filter_ctrl,
1323 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1324 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1325 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1326 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1327 	.is_removed = mlx5_is_removed,
1328 	.get_module_info = mlx5_get_module_info,
1329 	.get_module_eeprom = mlx5_get_module_eeprom,
1330 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1331 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1332 };
1333 
1334 /**
1335  * Verify and store value for device argument.
1336  *
1337  * @param[in] key
1338  *   Key argument to verify.
1339  * @param[in] val
1340  *   Value associated with key.
1341  * @param opaque
1342  *   User data.
1343  *
1344  * @return
1345  *   0 on success, a negative errno value otherwise and rte_errno is set.
1346  */
1347 static int
1348 mlx5_args_check(const char *key, const char *val, void *opaque)
1349 {
1350 	struct mlx5_dev_config *config = opaque;
1351 	unsigned long tmp;
1352 
1353 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1354 	if (!strcmp(MLX5_REPRESENTOR, key))
1355 		return 0;
1356 	errno = 0;
1357 	tmp = strtoul(val, NULL, 0);
1358 	if (errno) {
1359 		rte_errno = errno;
1360 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1361 		return -rte_errno;
1362 	}
1363 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1364 		config->cqe_comp = !!tmp;
1365 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1366 		config->cqe_pad = !!tmp;
1367 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1368 		config->hw_padding = !!tmp;
1369 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1370 		config->mprq.enabled = !!tmp;
1371 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1372 		config->mprq.stride_num_n = tmp;
1373 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1374 		config->mprq.max_memcpy_len = tmp;
1375 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1376 		config->mprq.min_rxqs_num = tmp;
1377 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1378 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1379 				 " converted to txq_inline_max", key);
1380 		config->txq_inline_max = tmp;
1381 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1382 		config->txq_inline_max = tmp;
1383 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1384 		config->txq_inline_min = tmp;
1385 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1386 		config->txq_inline_mpw = tmp;
1387 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1388 		config->txqs_inline = tmp;
1389 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1390 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1391 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1392 		config->mps = !!tmp;
1393 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1394 		config->dbnc = !!tmp;
1395 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1396 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1397 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1398 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1399 				 " converted to txq_inline_mpw", key);
1400 		config->txq_inline_mpw = tmp;
1401 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1402 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1403 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1404 		config->rx_vec_en = !!tmp;
1405 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1406 		config->l3_vxlan_en = !!tmp;
1407 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1408 		config->vf_nl_en = !!tmp;
1409 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1410 		config->dv_esw_en = !!tmp;
1411 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1412 		config->dv_flow_en = !!tmp;
1413 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1414 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1415 		    tmp != MLX5_XMETA_MODE_META16 &&
1416 		    tmp != MLX5_XMETA_MODE_META32) {
1417 			DRV_LOG(WARNING, "invalid extensive "
1418 					 "metadata parameter");
1419 			rte_errno = EINVAL;
1420 			return -rte_errno;
1421 		}
1422 		config->dv_xmeta_en = tmp;
1423 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1424 		config->mr_ext_memseg_en = !!tmp;
1425 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1426 		config->max_dump_files_num = tmp;
1427 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1428 		config->lro.timeout = tmp;
1429 	} else {
1430 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1431 		rte_errno = EINVAL;
1432 		return -rte_errno;
1433 	}
1434 	return 0;
1435 }
1436 
1437 /**
1438  * Parse device parameters.
1439  *
1440  * @param config
1441  *   Pointer to device configuration structure.
1442  * @param devargs
1443  *   Device arguments structure.
1444  *
1445  * @return
1446  *   0 on success, a negative errno value otherwise and rte_errno is set.
1447  */
1448 static int
1449 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1450 {
1451 	const char **params = (const char *[]){
1452 		MLX5_RXQ_CQE_COMP_EN,
1453 		MLX5_RXQ_CQE_PAD_EN,
1454 		MLX5_RXQ_PKT_PAD_EN,
1455 		MLX5_RX_MPRQ_EN,
1456 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1457 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1458 		MLX5_RXQS_MIN_MPRQ,
1459 		MLX5_TXQ_INLINE,
1460 		MLX5_TXQ_INLINE_MIN,
1461 		MLX5_TXQ_INLINE_MAX,
1462 		MLX5_TXQ_INLINE_MPW,
1463 		MLX5_TXQS_MIN_INLINE,
1464 		MLX5_TXQS_MAX_VEC,
1465 		MLX5_TXQ_MPW_EN,
1466 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1467 		MLX5_TXQ_MAX_INLINE_LEN,
1468 		MLX5_TX_DB_NC,
1469 		MLX5_TX_VEC_EN,
1470 		MLX5_RX_VEC_EN,
1471 		MLX5_L3_VXLAN_EN,
1472 		MLX5_VF_NL_EN,
1473 		MLX5_DV_ESW_EN,
1474 		MLX5_DV_FLOW_EN,
1475 		MLX5_DV_XMETA_EN,
1476 		MLX5_MR_EXT_MEMSEG_EN,
1477 		MLX5_REPRESENTOR,
1478 		MLX5_MAX_DUMP_FILES_NUM,
1479 		MLX5_LRO_TIMEOUT_USEC,
1480 		NULL,
1481 	};
1482 	struct rte_kvargs *kvlist;
1483 	int ret = 0;
1484 	int i;
1485 
1486 	if (devargs == NULL)
1487 		return 0;
1488 	/* Following UGLY cast is done to pass checkpatch. */
1489 	kvlist = rte_kvargs_parse(devargs->args, params);
1490 	if (kvlist == NULL) {
1491 		rte_errno = EINVAL;
1492 		return -rte_errno;
1493 	}
1494 	/* Process parameters. */
1495 	for (i = 0; (params[i] != NULL); ++i) {
1496 		if (rte_kvargs_count(kvlist, params[i])) {
1497 			ret = rte_kvargs_process(kvlist, params[i],
1498 						 mlx5_args_check, config);
1499 			if (ret) {
1500 				rte_errno = EINVAL;
1501 				rte_kvargs_free(kvlist);
1502 				return -rte_errno;
1503 			}
1504 		}
1505 	}
1506 	rte_kvargs_free(kvlist);
1507 	return 0;
1508 }
1509 
1510 static struct rte_pci_driver mlx5_driver;
1511 
1512 /**
1513  * PMD global initialization.
1514  *
1515  * Independent from individual device, this function initializes global
1516  * per-PMD data structures distinguishing primary and secondary processes.
1517  * Hence, each initialization is called once per a process.
1518  *
1519  * @return
1520  *   0 on success, a negative errno value otherwise and rte_errno is set.
1521  */
1522 static int
1523 mlx5_init_once(void)
1524 {
1525 	struct mlx5_shared_data *sd;
1526 	struct mlx5_local_data *ld = &mlx5_local_data;
1527 	int ret = 0;
1528 
1529 	if (mlx5_init_shared_data())
1530 		return -rte_errno;
1531 	sd = mlx5_shared_data;
1532 	assert(sd);
1533 	rte_spinlock_lock(&sd->lock);
1534 	switch (rte_eal_process_type()) {
1535 	case RTE_PROC_PRIMARY:
1536 		if (sd->init_done)
1537 			break;
1538 		LIST_INIT(&sd->mem_event_cb_list);
1539 		rte_rwlock_init(&sd->mem_event_rwlock);
1540 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1541 						mlx5_mr_mem_event_cb, NULL);
1542 		ret = mlx5_mp_init_primary();
1543 		if (ret)
1544 			goto out;
1545 		sd->init_done = true;
1546 		break;
1547 	case RTE_PROC_SECONDARY:
1548 		if (ld->init_done)
1549 			break;
1550 		ret = mlx5_mp_init_secondary();
1551 		if (ret)
1552 			goto out;
1553 		++sd->secondary_cnt;
1554 		ld->init_done = true;
1555 		break;
1556 	default:
1557 		break;
1558 	}
1559 out:
1560 	rte_spinlock_unlock(&sd->lock);
1561 	return ret;
1562 }
1563 
1564 /**
1565  * Configures the minimal amount of data to inline into WQE
1566  * while sending packets.
1567  *
1568  * - the txq_inline_min has the maximal priority, if this
1569  *   key is specified in devargs
1570  * - if DevX is enabled the inline mode is queried from the
1571  *   device (HCA attributes and NIC vport context if needed).
1572  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1573  *   and none (0 bytes) for other NICs
1574  *
1575  * @param spawn
1576  *   Verbs device parameters (name, port, switch_info) to spawn.
1577  * @param config
1578  *   Device configuration parameters.
1579  */
1580 static void
1581 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1582 		    struct mlx5_dev_config *config)
1583 {
1584 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1585 		/* Application defines size of inlined data explicitly. */
1586 		switch (spawn->pci_dev->id.device_id) {
1587 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1588 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1589 			if (config->txq_inline_min <
1590 				       (int)MLX5_INLINE_HSIZE_L2) {
1591 				DRV_LOG(DEBUG,
1592 					"txq_inline_mix aligned to minimal"
1593 					" ConnectX-4 required value %d",
1594 					(int)MLX5_INLINE_HSIZE_L2);
1595 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1596 			}
1597 			break;
1598 		}
1599 		goto exit;
1600 	}
1601 	if (config->hca_attr.eth_net_offloads) {
1602 		/* We have DevX enabled, inline mode queried successfully. */
1603 		switch (config->hca_attr.wqe_inline_mode) {
1604 		case MLX5_CAP_INLINE_MODE_L2:
1605 			/* outer L2 header must be inlined. */
1606 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1607 			goto exit;
1608 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1609 			/* No inline data are required by NIC. */
1610 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1611 			config->hw_vlan_insert =
1612 				config->hca_attr.wqe_vlan_insert;
1613 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1614 			goto exit;
1615 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1616 			/* inline mode is defined by NIC vport context. */
1617 			if (!config->hca_attr.eth_virt)
1618 				break;
1619 			switch (config->hca_attr.vport_inline_mode) {
1620 			case MLX5_INLINE_MODE_NONE:
1621 				config->txq_inline_min =
1622 					MLX5_INLINE_HSIZE_NONE;
1623 				goto exit;
1624 			case MLX5_INLINE_MODE_L2:
1625 				config->txq_inline_min =
1626 					MLX5_INLINE_HSIZE_L2;
1627 				goto exit;
1628 			case MLX5_INLINE_MODE_IP:
1629 				config->txq_inline_min =
1630 					MLX5_INLINE_HSIZE_L3;
1631 				goto exit;
1632 			case MLX5_INLINE_MODE_TCP_UDP:
1633 				config->txq_inline_min =
1634 					MLX5_INLINE_HSIZE_L4;
1635 				goto exit;
1636 			case MLX5_INLINE_MODE_INNER_L2:
1637 				config->txq_inline_min =
1638 					MLX5_INLINE_HSIZE_INNER_L2;
1639 				goto exit;
1640 			case MLX5_INLINE_MODE_INNER_IP:
1641 				config->txq_inline_min =
1642 					MLX5_INLINE_HSIZE_INNER_L3;
1643 				goto exit;
1644 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1645 				config->txq_inline_min =
1646 					MLX5_INLINE_HSIZE_INNER_L4;
1647 				goto exit;
1648 			}
1649 		}
1650 	}
1651 	/*
1652 	 * We get here if we are unable to deduce
1653 	 * inline data size with DevX. Try PCI ID
1654 	 * to determine old NICs.
1655 	 */
1656 	switch (spawn->pci_dev->id.device_id) {
1657 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1658 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1659 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1660 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1661 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1662 		config->hw_vlan_insert = 0;
1663 		break;
1664 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1665 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1666 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1667 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1668 		/*
1669 		 * These NICs support VLAN insertion from WQE and
1670 		 * report the wqe_vlan_insert flag. But there is the bug
1671 		 * and PFC control may be broken, so disable feature.
1672 		 */
1673 		config->hw_vlan_insert = 0;
1674 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1675 		break;
1676 	default:
1677 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1678 		break;
1679 	}
1680 exit:
1681 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1682 }
1683 
1684 /**
1685  * Configures the metadata mask fields in the shared context.
1686  *
1687  * @param [in] dev
1688  *   Pointer to Ethernet device.
1689  */
1690 static void
1691 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1692 {
1693 	struct mlx5_priv *priv = dev->data->dev_private;
1694 	struct mlx5_ibv_shared *sh = priv->sh;
1695 	uint32_t meta, mark, reg_c0;
1696 
1697 	reg_c0 = ~priv->vport_meta_mask;
1698 	switch (priv->config.dv_xmeta_en) {
1699 	case MLX5_XMETA_MODE_LEGACY:
1700 		meta = UINT32_MAX;
1701 		mark = MLX5_FLOW_MARK_MASK;
1702 		break;
1703 	case MLX5_XMETA_MODE_META16:
1704 		meta = reg_c0 >> rte_bsf32(reg_c0);
1705 		mark = MLX5_FLOW_MARK_MASK;
1706 		break;
1707 	case MLX5_XMETA_MODE_META32:
1708 		meta = UINT32_MAX;
1709 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1710 		break;
1711 	default:
1712 		meta = 0;
1713 		mark = 0;
1714 		assert(false);
1715 		break;
1716 	}
1717 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1718 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1719 				 sh->dv_mark_mask, mark);
1720 	else
1721 		sh->dv_mark_mask = mark;
1722 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1723 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1724 				 sh->dv_meta_mask, meta);
1725 	else
1726 		sh->dv_meta_mask = meta;
1727 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1728 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1729 				 sh->dv_meta_mask, reg_c0);
1730 	else
1731 		sh->dv_regc0_mask = reg_c0;
1732 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1733 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1734 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1735 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1736 }
1737 
1738 /**
1739  * Allocate page of door-bells and register it using DevX API.
1740  *
1741  * @param [in] dev
1742  *   Pointer to Ethernet device.
1743  *
1744  * @return
1745  *   Pointer to new page on success, NULL otherwise.
1746  */
1747 static struct mlx5_devx_dbr_page *
1748 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1749 {
1750 	struct mlx5_priv *priv = dev->data->dev_private;
1751 	struct mlx5_devx_dbr_page *page;
1752 
1753 	/* Allocate space for door-bell page and management data. */
1754 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1755 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1756 	if (!page) {
1757 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1758 			dev->data->port_id);
1759 		return NULL;
1760 	}
1761 	/* Register allocated memory. */
1762 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1763 					      MLX5_DBR_PAGE_SIZE, 0);
1764 	if (!page->umem) {
1765 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1766 			dev->data->port_id);
1767 		rte_free(page);
1768 		return NULL;
1769 	}
1770 	return page;
1771 }
1772 
1773 /**
1774  * Find the next available door-bell, allocate new page if needed.
1775  *
1776  * @param [in] dev
1777  *   Pointer to Ethernet device.
1778  * @param [out] dbr_page
1779  *   Door-bell page containing the page data.
1780  *
1781  * @return
1782  *   Door-bell address offset on success, a negative error value otherwise.
1783  */
1784 int64_t
1785 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1786 {
1787 	struct mlx5_priv *priv = dev->data->dev_private;
1788 	struct mlx5_devx_dbr_page *page = NULL;
1789 	uint32_t i, j;
1790 
1791 	LIST_FOREACH(page, &priv->dbrpgs, next)
1792 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1793 			break;
1794 	if (!page) { /* No page with free door-bell exists. */
1795 		page = mlx5_alloc_dbr_page(dev);
1796 		if (!page) /* Failed to allocate new page. */
1797 			return (-1);
1798 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1799 	}
1800 	/* Loop to find bitmap part with clear bit. */
1801 	for (i = 0;
1802 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1803 	     i++)
1804 		; /* Empty. */
1805 	/* Find the first clear bit. */
1806 	j = rte_bsf64(~page->dbr_bitmap[i]);
1807 	assert(i < (MLX5_DBR_PER_PAGE / 64));
1808 	page->dbr_bitmap[i] |= (1 << j);
1809 	page->dbr_count++;
1810 	*dbr_page = page;
1811 	return (((i * 64) + j) * sizeof(uint64_t));
1812 }
1813 
1814 /**
1815  * Release a door-bell record.
1816  *
1817  * @param [in] dev
1818  *   Pointer to Ethernet device.
1819  * @param [in] umem_id
1820  *   UMEM ID of page containing the door-bell record to release.
1821  * @param [in] offset
1822  *   Offset of door-bell record in page.
1823  *
1824  * @return
1825  *   0 on success, a negative error value otherwise.
1826  */
1827 int32_t
1828 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1829 {
1830 	struct mlx5_priv *priv = dev->data->dev_private;
1831 	struct mlx5_devx_dbr_page *page = NULL;
1832 	int ret = 0;
1833 
1834 	LIST_FOREACH(page, &priv->dbrpgs, next)
1835 		/* Find the page this address belongs to. */
1836 		if (page->umem->umem_id == umem_id)
1837 			break;
1838 	if (!page)
1839 		return -EINVAL;
1840 	page->dbr_count--;
1841 	if (!page->dbr_count) {
1842 		/* Page not used, free it and remove from list. */
1843 		LIST_REMOVE(page, next);
1844 		if (page->umem)
1845 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
1846 		rte_free(page);
1847 	} else {
1848 		/* Mark in bitmap that this door-bell is not in use. */
1849 		offset /= MLX5_DBR_SIZE;
1850 		int i = offset / 64;
1851 		int j = offset % 64;
1852 
1853 		page->dbr_bitmap[i] &= ~(1 << j);
1854 	}
1855 	return ret;
1856 }
1857 
1858 /**
1859  * Check sibling device configurations.
1860  *
1861  * Sibling devices sharing the Infiniband device context
1862  * should have compatible configurations. This regards
1863  * representors and bonding slaves.
1864  *
1865  * @param priv
1866  *   Private device descriptor.
1867  * @param config
1868  *   Configuration of the device is going to be created.
1869  *
1870  * @return
1871  *   0 on success, EINVAL otherwise
1872  */
1873 static int
1874 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1875 			      struct mlx5_dev_config *config)
1876 {
1877 	struct mlx5_ibv_shared *sh = priv->sh;
1878 	struct mlx5_dev_config *sh_conf = NULL;
1879 	uint16_t port_id;
1880 
1881 	assert(sh);
1882 	/* Nothing to compare for the single/first device. */
1883 	if (sh->refcnt == 1)
1884 		return 0;
1885 	/* Find the device with shared context. */
1886 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1887 		struct mlx5_priv *opriv =
1888 			rte_eth_devices[port_id].data->dev_private;
1889 
1890 		if (opriv && opriv != priv && opriv->sh == sh) {
1891 			sh_conf = &opriv->config;
1892 			break;
1893 		}
1894 	}
1895 	if (!sh_conf)
1896 		return 0;
1897 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1898 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1899 			     " for shared %s context", sh->ibdev_name);
1900 		rte_errno = EINVAL;
1901 		return rte_errno;
1902 	}
1903 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1904 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1905 			     " for shared %s context", sh->ibdev_name);
1906 		rte_errno = EINVAL;
1907 		return rte_errno;
1908 	}
1909 	return 0;
1910 }
1911 /**
1912  * Spawn an Ethernet device from Verbs information.
1913  *
1914  * @param dpdk_dev
1915  *   Backing DPDK device.
1916  * @param spawn
1917  *   Verbs device parameters (name, port, switch_info) to spawn.
1918  * @param config
1919  *   Device configuration parameters.
1920  *
1921  * @return
1922  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1923  *   is set. The following errors are defined:
1924  *
1925  *   EBUSY: device is not supposed to be spawned.
1926  *   EEXIST: device is already spawned
1927  */
1928 static struct rte_eth_dev *
1929 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1930 	       struct mlx5_dev_spawn_data *spawn,
1931 	       struct mlx5_dev_config config)
1932 {
1933 	const struct mlx5_switch_info *switch_info = &spawn->info;
1934 	struct mlx5_ibv_shared *sh = NULL;
1935 	struct ibv_port_attr port_attr;
1936 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1937 	struct rte_eth_dev *eth_dev = NULL;
1938 	struct mlx5_priv *priv = NULL;
1939 	int err = 0;
1940 	unsigned int hw_padding = 0;
1941 	unsigned int mps;
1942 	unsigned int cqe_comp;
1943 	unsigned int cqe_pad = 0;
1944 	unsigned int tunnel_en = 0;
1945 	unsigned int mpls_en = 0;
1946 	unsigned int swp = 0;
1947 	unsigned int mprq = 0;
1948 	unsigned int mprq_min_stride_size_n = 0;
1949 	unsigned int mprq_max_stride_size_n = 0;
1950 	unsigned int mprq_min_stride_num_n = 0;
1951 	unsigned int mprq_max_stride_num_n = 0;
1952 	struct rte_ether_addr mac;
1953 	char name[RTE_ETH_NAME_MAX_LEN];
1954 	int own_domain_id = 0;
1955 	uint16_t port_id;
1956 	unsigned int i;
1957 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1958 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
1959 #endif
1960 
1961 	/* Determine if this port representor is supposed to be spawned. */
1962 	if (switch_info->representor && dpdk_dev->devargs) {
1963 		struct rte_eth_devargs eth_da;
1964 
1965 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1966 		if (err) {
1967 			rte_errno = -err;
1968 			DRV_LOG(ERR, "failed to process device arguments: %s",
1969 				strerror(rte_errno));
1970 			return NULL;
1971 		}
1972 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
1973 			if (eth_da.representor_ports[i] ==
1974 			    (uint16_t)switch_info->port_name)
1975 				break;
1976 		if (i == eth_da.nb_representor_ports) {
1977 			rte_errno = EBUSY;
1978 			return NULL;
1979 		}
1980 	}
1981 	/* Build device name. */
1982 	if (spawn->pf_bond <  0) {
1983 		/* Single device. */
1984 		if (!switch_info->representor)
1985 			strlcpy(name, dpdk_dev->name, sizeof(name));
1986 		else
1987 			snprintf(name, sizeof(name), "%s_representor_%u",
1988 				 dpdk_dev->name, switch_info->port_name);
1989 	} else {
1990 		/* Bonding device. */
1991 		if (!switch_info->representor)
1992 			snprintf(name, sizeof(name), "%s_%s",
1993 				 dpdk_dev->name, spawn->ibv_dev->name);
1994 		else
1995 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
1996 				 dpdk_dev->name, spawn->ibv_dev->name,
1997 				 switch_info->port_name);
1998 	}
1999 	/* check if the device is already spawned */
2000 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2001 		rte_errno = EEXIST;
2002 		return NULL;
2003 	}
2004 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2005 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2006 		eth_dev = rte_eth_dev_attach_secondary(name);
2007 		if (eth_dev == NULL) {
2008 			DRV_LOG(ERR, "can not attach rte ethdev");
2009 			rte_errno = ENOMEM;
2010 			return NULL;
2011 		}
2012 		eth_dev->device = dpdk_dev;
2013 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2014 		err = mlx5_proc_priv_init(eth_dev);
2015 		if (err)
2016 			return NULL;
2017 		/* Receive command fd from primary process */
2018 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2019 		if (err < 0)
2020 			return NULL;
2021 		/* Remap UAR for Tx queues. */
2022 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2023 		if (err)
2024 			return NULL;
2025 		/*
2026 		 * Ethdev pointer is still required as input since
2027 		 * the primary device is not accessible from the
2028 		 * secondary process.
2029 		 */
2030 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2031 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2032 		return eth_dev;
2033 	}
2034 	/*
2035 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2036 	 * advance to create dv/verbs device context. We proceed the
2037 	 * devargs here to get ones, and later proceed devargs again
2038 	 * to override some hardware settings.
2039 	 */
2040 	err = mlx5_args(&config, dpdk_dev->devargs);
2041 	if (err) {
2042 		err = rte_errno;
2043 		DRV_LOG(ERR, "failed to process device arguments: %s",
2044 			strerror(rte_errno));
2045 		goto error;
2046 	}
2047 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2048 	if (!sh)
2049 		return NULL;
2050 	config.devx = sh->devx;
2051 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2052 	config.dest_tir = 1;
2053 #endif
2054 #ifdef HAVE_IBV_MLX5_MOD_SWP
2055 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2056 #endif
2057 	/*
2058 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2059 	 * as all ConnectX-5 devices.
2060 	 */
2061 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2062 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2063 #endif
2064 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2065 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2066 #endif
2067 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2068 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2069 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2070 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2071 			mps = MLX5_MPW_ENHANCED;
2072 		} else {
2073 			DRV_LOG(DEBUG, "MPW is supported");
2074 			mps = MLX5_MPW;
2075 		}
2076 	} else {
2077 		DRV_LOG(DEBUG, "MPW isn't supported");
2078 		mps = MLX5_MPW_DISABLED;
2079 	}
2080 #ifdef HAVE_IBV_MLX5_MOD_SWP
2081 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2082 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2083 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2084 #endif
2085 	config.swp = !!swp;
2086 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2087 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2088 		struct mlx5dv_striding_rq_caps mprq_caps =
2089 			dv_attr.striding_rq_caps;
2090 
2091 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2092 			mprq_caps.min_single_stride_log_num_of_bytes);
2093 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2094 			mprq_caps.max_single_stride_log_num_of_bytes);
2095 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2096 			mprq_caps.min_single_wqe_log_num_of_strides);
2097 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2098 			mprq_caps.max_single_wqe_log_num_of_strides);
2099 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2100 			mprq_caps.supported_qpts);
2101 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2102 		mprq = 1;
2103 		mprq_min_stride_size_n =
2104 			mprq_caps.min_single_stride_log_num_of_bytes;
2105 		mprq_max_stride_size_n =
2106 			mprq_caps.max_single_stride_log_num_of_bytes;
2107 		mprq_min_stride_num_n =
2108 			mprq_caps.min_single_wqe_log_num_of_strides;
2109 		mprq_max_stride_num_n =
2110 			mprq_caps.max_single_wqe_log_num_of_strides;
2111 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2112 						   mprq_min_stride_num_n);
2113 	}
2114 #endif
2115 	if (RTE_CACHE_LINE_SIZE == 128 &&
2116 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2117 		cqe_comp = 0;
2118 	else
2119 		cqe_comp = 1;
2120 	config.cqe_comp = cqe_comp;
2121 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2122 	/* Whether device supports 128B Rx CQE padding. */
2123 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2124 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2125 #endif
2126 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2127 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2128 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2129 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2130 			     (dv_attr.tunnel_offloads_caps &
2131 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
2132 	}
2133 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2134 		tunnel_en ? "" : "not ");
2135 #else
2136 	DRV_LOG(WARNING,
2137 		"tunnel offloading disabled due to old OFED/rdma-core version");
2138 #endif
2139 	config.tunnel_en = tunnel_en;
2140 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2141 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2142 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2143 		   (dv_attr.tunnel_offloads_caps &
2144 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2145 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2146 		mpls_en ? "" : "not ");
2147 #else
2148 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2149 		" old OFED/rdma-core version or firmware configuration");
2150 #endif
2151 	config.mpls_en = mpls_en;
2152 	/* Check port status. */
2153 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2154 	if (err) {
2155 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2156 		goto error;
2157 	}
2158 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2159 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2160 		err = EINVAL;
2161 		goto error;
2162 	}
2163 	if (port_attr.state != IBV_PORT_ACTIVE)
2164 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2165 			mlx5_glue->port_state_str(port_attr.state),
2166 			port_attr.state);
2167 	/* Allocate private eth device data. */
2168 	priv = rte_zmalloc("ethdev private structure",
2169 			   sizeof(*priv),
2170 			   RTE_CACHE_LINE_SIZE);
2171 	if (priv == NULL) {
2172 		DRV_LOG(ERR, "priv allocation failure");
2173 		err = ENOMEM;
2174 		goto error;
2175 	}
2176 	priv->sh = sh;
2177 	priv->ibv_port = spawn->ibv_port;
2178 	priv->pci_dev = spawn->pci_dev;
2179 	priv->mtu = RTE_ETHER_MTU;
2180 #ifndef RTE_ARCH_64
2181 	/* Initialize UAR access locks for 32bit implementations. */
2182 	rte_spinlock_init(&priv->uar_lock_cq);
2183 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2184 		rte_spinlock_init(&priv->uar_lock[i]);
2185 #endif
2186 	/* Some internal functions rely on Netlink sockets, open them now. */
2187 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2188 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2189 	priv->nl_sn = 0;
2190 	priv->representor = !!switch_info->representor;
2191 	priv->master = !!switch_info->master;
2192 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2193 	priv->vport_meta_tag = 0;
2194 	priv->vport_meta_mask = 0;
2195 	priv->pf_bond = spawn->pf_bond;
2196 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2197 	/*
2198 	 * The DevX port query API is implemented. E-Switch may use
2199 	 * either vport or reg_c[0] metadata register to match on
2200 	 * vport index. The engaged part of metadata register is
2201 	 * defined by mask.
2202 	 */
2203 	if (switch_info->representor || switch_info->master) {
2204 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2205 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2206 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2207 						 &devx_port);
2208 		if (err) {
2209 			DRV_LOG(WARNING,
2210 				"can't query devx port %d on device %s",
2211 				spawn->ibv_port, spawn->ibv_dev->name);
2212 			devx_port.comp_mask = 0;
2213 		}
2214 	}
2215 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2216 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2217 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2218 		if (!priv->vport_meta_mask) {
2219 			DRV_LOG(ERR, "vport zero mask for port %d"
2220 				     " on bonding device %s",
2221 				     spawn->ibv_port, spawn->ibv_dev->name);
2222 			err = ENOTSUP;
2223 			goto error;
2224 		}
2225 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2226 			DRV_LOG(ERR, "invalid vport tag for port %d"
2227 				     " on bonding device %s",
2228 				     spawn->ibv_port, spawn->ibv_dev->name);
2229 			err = ENOTSUP;
2230 			goto error;
2231 		}
2232 	} else if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2233 		priv->vport_id = devx_port.vport_num;
2234 	} else if (spawn->pf_bond >= 0) {
2235 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2236 			     " on bonding device %s",
2237 			     spawn->ibv_port, spawn->ibv_dev->name);
2238 		err = ENOTSUP;
2239 		goto error;
2240 	} else {
2241 		/* Suppose vport index in compatible way. */
2242 		priv->vport_id = switch_info->representor ?
2243 				 switch_info->port_name + 1 : -1;
2244 	}
2245 #else
2246 	/*
2247 	 * Kernel/rdma_core support single E-Switch per PF configurations
2248 	 * only and vport_id field contains the vport index for
2249 	 * associated VF, which is deduced from representor port name.
2250 	 * For example, let's have the IB device port 10, it has
2251 	 * attached network device eth0, which has port name attribute
2252 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2253 	 * as 3 (2+1). This assigning schema should be changed if the
2254 	 * multiple E-Switch instances per PF configurations or/and PCI
2255 	 * subfunctions are added.
2256 	 */
2257 	priv->vport_id = switch_info->representor ?
2258 			 switch_info->port_name + 1 : -1;
2259 #endif
2260 	/* representor_id field keeps the unmodified VF index. */
2261 	priv->representor_id = switch_info->representor ?
2262 			       switch_info->port_name : -1;
2263 	/*
2264 	 * Look for sibling devices in order to reuse their switch domain
2265 	 * if any, otherwise allocate one.
2266 	 */
2267 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2268 		const struct mlx5_priv *opriv =
2269 			rte_eth_devices[port_id].data->dev_private;
2270 
2271 		if (!opriv ||
2272 		    opriv->sh != priv->sh ||
2273 			opriv->domain_id ==
2274 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2275 			continue;
2276 		priv->domain_id = opriv->domain_id;
2277 		break;
2278 	}
2279 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2280 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2281 		if (err) {
2282 			err = rte_errno;
2283 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2284 				strerror(rte_errno));
2285 			goto error;
2286 		}
2287 		own_domain_id = 1;
2288 	}
2289 	/* Override some values set by hardware configuration. */
2290 	mlx5_args(&config, dpdk_dev->devargs);
2291 	err = mlx5_dev_check_sibling_config(priv, &config);
2292 	if (err)
2293 		goto error;
2294 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2295 			    IBV_DEVICE_RAW_IP_CSUM);
2296 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2297 		(config.hw_csum ? "" : "not "));
2298 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2299 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2300 	DRV_LOG(DEBUG, "counters are not supported");
2301 #endif
2302 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
2303 	if (config.dv_flow_en) {
2304 		DRV_LOG(WARNING, "DV flow is not supported");
2305 		config.dv_flow_en = 0;
2306 	}
2307 #endif
2308 	config.ind_table_max_size =
2309 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2310 	/*
2311 	 * Remove this check once DPDK supports larger/variable
2312 	 * indirection tables.
2313 	 */
2314 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2315 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2316 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2317 		config.ind_table_max_size);
2318 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2319 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2320 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2321 		(config.hw_vlan_strip ? "" : "not "));
2322 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2323 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2324 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2325 		(config.hw_fcs_strip ? "" : "not "));
2326 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2327 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2328 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2329 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2330 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2331 #endif
2332 	if (config.hw_padding && !hw_padding) {
2333 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2334 		config.hw_padding = 0;
2335 	} else if (config.hw_padding) {
2336 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2337 	}
2338 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2339 		      (sh->device_attr.tso_caps.supported_qpts &
2340 		       (1 << IBV_QPT_RAW_PACKET)));
2341 	if (config.tso)
2342 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2343 	/*
2344 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2345 	 * by default.
2346 	 */
2347 	if (config.mps == MLX5_ARG_UNSET)
2348 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2349 							  MLX5_MPW_DISABLED;
2350 	else
2351 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2352 	DRV_LOG(INFO, "%sMPS is %s",
2353 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
2354 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2355 	if (config.cqe_comp && !cqe_comp) {
2356 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2357 		config.cqe_comp = 0;
2358 	}
2359 	if (config.cqe_pad && !cqe_pad) {
2360 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2361 		config.cqe_pad = 0;
2362 	} else if (config.cqe_pad) {
2363 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2364 	}
2365 	if (config.devx) {
2366 		priv->counter_fallback = 0;
2367 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2368 		if (err) {
2369 			err = -err;
2370 			goto error;
2371 		}
2372 		if (!config.hca_attr.flow_counters_dump)
2373 			priv->counter_fallback = 1;
2374 #ifndef HAVE_IBV_DEVX_ASYNC
2375 		priv->counter_fallback = 1;
2376 #endif
2377 		if (priv->counter_fallback)
2378 			DRV_LOG(INFO, "Use fall-back DV counter management");
2379 		/* Check for LRO support. */
2380 		if (config.dest_tir && config.hca_attr.lro_cap &&
2381 		    config.dv_flow_en) {
2382 			/* TBD check tunnel lro caps. */
2383 			config.lro.supported = config.hca_attr.lro_cap;
2384 			DRV_LOG(DEBUG, "Device supports LRO");
2385 			/*
2386 			 * If LRO timeout is not configured by application,
2387 			 * use the minimal supported value.
2388 			 */
2389 			if (!config.lro.timeout)
2390 				config.lro.timeout =
2391 				config.hca_attr.lro_timer_supported_periods[0];
2392 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2393 				config.lro.timeout);
2394 		}
2395 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2396 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2397 		    config.dv_flow_en) {
2398 			uint8_t reg_c_mask =
2399 				config.hca_attr.qos.flow_meter_reg_c_ids;
2400 			/*
2401 			 * Meter needs two REG_C's for color match and pre-sfx
2402 			 * flow match. Here get the REG_C for color match.
2403 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2404 			 */
2405 			reg_c_mask &= 0xfc;
2406 			if (__builtin_popcount(reg_c_mask) < 1) {
2407 				priv->mtr_en = 0;
2408 				DRV_LOG(WARNING, "No available register for"
2409 					" meter.");
2410 			} else {
2411 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2412 						      REG_C_0;
2413 				priv->mtr_en = 1;
2414 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2415 					priv->mtr_color_reg);
2416 			}
2417 		}
2418 #endif
2419 	}
2420 	if (config.mprq.enabled && mprq) {
2421 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2422 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2423 			config.mprq.stride_num_n =
2424 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2425 					mprq_min_stride_num_n);
2426 			DRV_LOG(WARNING,
2427 				"the number of strides"
2428 				" for Multi-Packet RQ is out of range,"
2429 				" setting default value (%u)",
2430 				1 << config.mprq.stride_num_n);
2431 		}
2432 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2433 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2434 	} else if (config.mprq.enabled && !mprq) {
2435 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2436 		config.mprq.enabled = 0;
2437 	}
2438 	if (config.max_dump_files_num == 0)
2439 		config.max_dump_files_num = 128;
2440 	eth_dev = rte_eth_dev_allocate(name);
2441 	if (eth_dev == NULL) {
2442 		DRV_LOG(ERR, "can not allocate rte ethdev");
2443 		err = ENOMEM;
2444 		goto error;
2445 	}
2446 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2447 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2448 	if (priv->representor) {
2449 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2450 		eth_dev->data->representor_id = priv->representor_id;
2451 	}
2452 	/*
2453 	 * Store associated network device interface index. This index
2454 	 * is permanent throughout the lifetime of device. So, we may store
2455 	 * the ifindex here and use the cached value further.
2456 	 */
2457 	assert(spawn->ifindex);
2458 	priv->if_index = spawn->ifindex;
2459 	eth_dev->data->dev_private = priv;
2460 	priv->dev_data = eth_dev->data;
2461 	eth_dev->data->mac_addrs = priv->mac;
2462 	eth_dev->device = dpdk_dev;
2463 	/* Configure the first MAC address by default. */
2464 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2465 		DRV_LOG(ERR,
2466 			"port %u cannot get MAC address, is mlx5_en"
2467 			" loaded? (errno: %s)",
2468 			eth_dev->data->port_id, strerror(rte_errno));
2469 		err = ENODEV;
2470 		goto error;
2471 	}
2472 	DRV_LOG(INFO,
2473 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2474 		eth_dev->data->port_id,
2475 		mac.addr_bytes[0], mac.addr_bytes[1],
2476 		mac.addr_bytes[2], mac.addr_bytes[3],
2477 		mac.addr_bytes[4], mac.addr_bytes[5]);
2478 #ifndef NDEBUG
2479 	{
2480 		char ifname[IF_NAMESIZE];
2481 
2482 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2483 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2484 				eth_dev->data->port_id, ifname);
2485 		else
2486 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2487 				eth_dev->data->port_id);
2488 	}
2489 #endif
2490 	/* Get actual MTU if possible. */
2491 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2492 	if (err) {
2493 		err = rte_errno;
2494 		goto error;
2495 	}
2496 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2497 		priv->mtu);
2498 	/* Initialize burst functions to prevent crashes before link-up. */
2499 	eth_dev->rx_pkt_burst = removed_rx_burst;
2500 	eth_dev->tx_pkt_burst = removed_tx_burst;
2501 	eth_dev->dev_ops = &mlx5_dev_ops;
2502 	/* Register MAC address. */
2503 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2504 	if (config.vf && config.vf_nl_en)
2505 		mlx5_nl_mac_addr_sync(eth_dev);
2506 	TAILQ_INIT(&priv->flows);
2507 	TAILQ_INIT(&priv->ctrl_flows);
2508 	TAILQ_INIT(&priv->flow_meters);
2509 	TAILQ_INIT(&priv->flow_meter_profiles);
2510 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2511 	struct mlx5dv_ctx_allocators alctr = {
2512 		.alloc = &mlx5_alloc_verbs_buf,
2513 		.free = &mlx5_free_verbs_buf,
2514 		.data = priv,
2515 	};
2516 	mlx5_glue->dv_set_context_attr(sh->ctx,
2517 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2518 				       (void *)((uintptr_t)&alctr));
2519 	/* Bring Ethernet device up. */
2520 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2521 		eth_dev->data->port_id);
2522 	mlx5_set_link_up(eth_dev);
2523 	/*
2524 	 * Even though the interrupt handler is not installed yet,
2525 	 * interrupts will still trigger on the async_fd from
2526 	 * Verbs context returned by ibv_open_device().
2527 	 */
2528 	mlx5_link_update(eth_dev, 0);
2529 #ifdef HAVE_MLX5DV_DR_ESWITCH
2530 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2531 	      (switch_info->representor || switch_info->master)))
2532 		config.dv_esw_en = 0;
2533 #else
2534 	config.dv_esw_en = 0;
2535 #endif
2536 	/* Detect minimal data bytes to inline. */
2537 	mlx5_set_min_inline(spawn, &config);
2538 	/* Store device configuration on private structure. */
2539 	priv->config = config;
2540 	/* Create context for virtual machine VLAN workaround. */
2541 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2542 	if (config.dv_flow_en) {
2543 		err = mlx5_alloc_shared_dr(priv);
2544 		if (err)
2545 			goto error;
2546 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc();
2547 		if (!priv->qrss_id_pool) {
2548 			DRV_LOG(ERR, "can't create flow id pool");
2549 			err = ENOMEM;
2550 			goto error;
2551 		}
2552 	}
2553 	/* Supported Verbs flow priority number detection. */
2554 	err = mlx5_flow_discover_priorities(eth_dev);
2555 	if (err < 0) {
2556 		err = -err;
2557 		goto error;
2558 	}
2559 	priv->config.flow_prio = err;
2560 	if (!priv->config.dv_esw_en &&
2561 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2562 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2563 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2564 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2565 	}
2566 	mlx5_set_metadata_mask(eth_dev);
2567 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2568 	    !priv->sh->dv_regc0_mask) {
2569 		DRV_LOG(ERR, "metadata mode %u is not supported "
2570 			     "(no metadata reg_c[0] is available)",
2571 			     priv->config.dv_xmeta_en);
2572 			err = ENOTSUP;
2573 			goto error;
2574 	}
2575 	/* Query availibility of metadata reg_c's. */
2576 	err = mlx5_flow_discover_mreg_c(eth_dev);
2577 	if (err < 0) {
2578 		err = -err;
2579 		goto error;
2580 	}
2581 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2582 		DRV_LOG(DEBUG,
2583 			"port %u extensive metadata register is not supported",
2584 			eth_dev->data->port_id);
2585 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2586 			DRV_LOG(ERR, "metadata mode %u is not supported "
2587 				     "(no metadata registers available)",
2588 				     priv->config.dv_xmeta_en);
2589 			err = ENOTSUP;
2590 			goto error;
2591 		}
2592 	}
2593 	if (priv->config.dv_flow_en &&
2594 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2595 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
2596 	    priv->sh->dv_regc0_mask) {
2597 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2598 						      MLX5_FLOW_MREG_HTABLE_SZ);
2599 		if (!priv->mreg_cp_tbl) {
2600 			err = ENOMEM;
2601 			goto error;
2602 		}
2603 	}
2604 	return eth_dev;
2605 error:
2606 	if (priv) {
2607 		if (priv->mreg_cp_tbl)
2608 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2609 		if (priv->sh)
2610 			mlx5_free_shared_dr(priv);
2611 		if (priv->nl_socket_route >= 0)
2612 			close(priv->nl_socket_route);
2613 		if (priv->nl_socket_rdma >= 0)
2614 			close(priv->nl_socket_rdma);
2615 		if (priv->vmwa_context)
2616 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2617 		if (priv->qrss_id_pool)
2618 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
2619 		if (own_domain_id)
2620 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2621 		rte_free(priv);
2622 		if (eth_dev != NULL)
2623 			eth_dev->data->dev_private = NULL;
2624 	}
2625 	if (eth_dev != NULL) {
2626 		/* mac_addrs must not be freed alone because part of dev_private */
2627 		eth_dev->data->mac_addrs = NULL;
2628 		rte_eth_dev_release_port(eth_dev);
2629 	}
2630 	if (sh)
2631 		mlx5_free_shared_ibctx(sh);
2632 	assert(err > 0);
2633 	rte_errno = err;
2634 	return NULL;
2635 }
2636 
2637 /**
2638  * Comparison callback to sort device data.
2639  *
2640  * This is meant to be used with qsort().
2641  *
2642  * @param a[in]
2643  *   Pointer to pointer to first data object.
2644  * @param b[in]
2645  *   Pointer to pointer to second data object.
2646  *
2647  * @return
2648  *   0 if both objects are equal, less than 0 if the first argument is less
2649  *   than the second, greater than 0 otherwise.
2650  */
2651 static int
2652 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2653 {
2654 	const struct mlx5_switch_info *si_a =
2655 		&((const struct mlx5_dev_spawn_data *)a)->info;
2656 	const struct mlx5_switch_info *si_b =
2657 		&((const struct mlx5_dev_spawn_data *)b)->info;
2658 	int ret;
2659 
2660 	/* Master device first. */
2661 	ret = si_b->master - si_a->master;
2662 	if (ret)
2663 		return ret;
2664 	/* Then representor devices. */
2665 	ret = si_b->representor - si_a->representor;
2666 	if (ret)
2667 		return ret;
2668 	/* Unidentified devices come last in no specific order. */
2669 	if (!si_a->representor)
2670 		return 0;
2671 	/* Order representors by name. */
2672 	return si_a->port_name - si_b->port_name;
2673 }
2674 
2675 /**
2676  * Match PCI information for possible slaves of bonding device.
2677  *
2678  * @param[in] ibv_dev
2679  *   Pointer to Infiniband device structure.
2680  * @param[in] pci_dev
2681  *   Pointer to PCI device structure to match PCI address.
2682  * @param[in] nl_rdma
2683  *   Netlink RDMA group socket handle.
2684  *
2685  * @return
2686  *   negative value if no bonding device found, otherwise
2687  *   positive index of slave PF in bonding.
2688  */
2689 static int
2690 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2691 			   const struct rte_pci_device *pci_dev,
2692 			   int nl_rdma)
2693 {
2694 	char ifname[IF_NAMESIZE + 1];
2695 	unsigned int ifindex;
2696 	unsigned int np, i;
2697 	FILE *file = NULL;
2698 	int pf = -1;
2699 
2700 	/*
2701 	 * Try to get master device name. If something goes
2702 	 * wrong suppose the lack of kernel support and no
2703 	 * bonding devices.
2704 	 */
2705 	if (nl_rdma < 0)
2706 		return -1;
2707 	if (!strstr(ibv_dev->name, "bond"))
2708 		return -1;
2709 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2710 	if (!np)
2711 		return -1;
2712 	/*
2713 	 * The Master device might not be on the predefined
2714 	 * port (not on port index 1, it is not garanted),
2715 	 * we have to scan all Infiniband device port and
2716 	 * find master.
2717 	 */
2718 	for (i = 1; i <= np; ++i) {
2719 		/* Check whether Infiniband port is populated. */
2720 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2721 		if (!ifindex)
2722 			continue;
2723 		if (!if_indextoname(ifindex, ifname))
2724 			continue;
2725 		/* Try to read bonding slave names from sysfs. */
2726 		MKSTR(slaves,
2727 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2728 		file = fopen(slaves, "r");
2729 		if (file)
2730 			break;
2731 	}
2732 	if (!file)
2733 		return -1;
2734 	/* Use safe format to check maximal buffer length. */
2735 	assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2736 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2737 		char tmp_str[IF_NAMESIZE + 32];
2738 		struct rte_pci_addr pci_addr;
2739 		struct mlx5_switch_info	info;
2740 
2741 		/* Process slave interface names in the loop. */
2742 		snprintf(tmp_str, sizeof(tmp_str),
2743 			 "/sys/class/net/%s", ifname);
2744 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2745 			DRV_LOG(WARNING, "can not get PCI address"
2746 					 " for netdev \"%s\"", ifname);
2747 			continue;
2748 		}
2749 		if (pci_dev->addr.domain != pci_addr.domain ||
2750 		    pci_dev->addr.bus != pci_addr.bus ||
2751 		    pci_dev->addr.devid != pci_addr.devid ||
2752 		    pci_dev->addr.function != pci_addr.function)
2753 			continue;
2754 		/* Slave interface PCI address match found. */
2755 		fclose(file);
2756 		snprintf(tmp_str, sizeof(tmp_str),
2757 			 "/sys/class/net/%s/phys_port_name", ifname);
2758 		file = fopen(tmp_str, "rb");
2759 		if (!file)
2760 			break;
2761 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2762 		if (fscanf(file, "%32s", tmp_str) == 1)
2763 			mlx5_translate_port_name(tmp_str, &info);
2764 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2765 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2766 			pf = info.port_name;
2767 		break;
2768 	}
2769 	if (file)
2770 		fclose(file);
2771 	return pf;
2772 }
2773 
2774 /**
2775  * DPDK callback to register a PCI device.
2776  *
2777  * This function spawns Ethernet devices out of a given PCI device.
2778  *
2779  * @param[in] pci_drv
2780  *   PCI driver structure (mlx5_driver).
2781  * @param[in] pci_dev
2782  *   PCI device information.
2783  *
2784  * @return
2785  *   0 on success, a negative errno value otherwise and rte_errno is set.
2786  */
2787 static int
2788 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2789 	       struct rte_pci_device *pci_dev)
2790 {
2791 	struct ibv_device **ibv_list;
2792 	/*
2793 	 * Number of found IB Devices matching with requested PCI BDF.
2794 	 * nd != 1 means there are multiple IB devices over the same
2795 	 * PCI device and we have representors and master.
2796 	 */
2797 	unsigned int nd = 0;
2798 	/*
2799 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2800 	 * we have the single multiport IB device, and there may be
2801 	 * representors attached to some of found ports.
2802 	 */
2803 	unsigned int np = 0;
2804 	/*
2805 	 * Number of DPDK ethernet devices to Spawn - either over
2806 	 * multiple IB devices or multiple ports of single IB device.
2807 	 * Actually this is the number of iterations to spawn.
2808 	 */
2809 	unsigned int ns = 0;
2810 	/*
2811 	 * Bonding device
2812 	 *   < 0 - no bonding device (single one)
2813 	 *  >= 0 - bonding device (value is slave PF index)
2814 	 */
2815 	int bd = -1;
2816 	struct mlx5_dev_spawn_data *list = NULL;
2817 	struct mlx5_dev_config dev_config;
2818 	int ret;
2819 
2820 	ret = mlx5_init_once();
2821 	if (ret) {
2822 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2823 			strerror(rte_errno));
2824 		return -rte_errno;
2825 	}
2826 	assert(pci_drv == &mlx5_driver);
2827 	errno = 0;
2828 	ibv_list = mlx5_glue->get_device_list(&ret);
2829 	if (!ibv_list) {
2830 		rte_errno = errno ? errno : ENOSYS;
2831 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2832 		return -rte_errno;
2833 	}
2834 	/*
2835 	 * First scan the list of all Infiniband devices to find
2836 	 * matching ones, gathering into the list.
2837 	 */
2838 	struct ibv_device *ibv_match[ret + 1];
2839 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2840 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2841 	unsigned int i;
2842 
2843 	while (ret-- > 0) {
2844 		struct rte_pci_addr pci_addr;
2845 
2846 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2847 		bd = mlx5_device_bond_pci_match
2848 				(ibv_list[ret], pci_dev, nl_rdma);
2849 		if (bd >= 0) {
2850 			/*
2851 			 * Bonding device detected. Only one match is allowed,
2852 			 * the bonding is supported over multi-port IB device,
2853 			 * there should be no matches on representor PCI
2854 			 * functions or non VF LAG bonding devices with
2855 			 * specified address.
2856 			 */
2857 			if (nd) {
2858 				DRV_LOG(ERR,
2859 					"multiple PCI match on bonding device"
2860 					"\"%s\" found", ibv_list[ret]->name);
2861 				rte_errno = ENOENT;
2862 				ret = -rte_errno;
2863 				goto exit;
2864 			}
2865 			DRV_LOG(INFO, "PCI information matches for"
2866 				      " slave %d bonding device \"%s\"",
2867 				      bd, ibv_list[ret]->name);
2868 			ibv_match[nd++] = ibv_list[ret];
2869 			break;
2870 		}
2871 		if (mlx5_dev_to_pci_addr
2872 			(ibv_list[ret]->ibdev_path, &pci_addr))
2873 			continue;
2874 		if (pci_dev->addr.domain != pci_addr.domain ||
2875 		    pci_dev->addr.bus != pci_addr.bus ||
2876 		    pci_dev->addr.devid != pci_addr.devid ||
2877 		    pci_dev->addr.function != pci_addr.function)
2878 			continue;
2879 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2880 			ibv_list[ret]->name);
2881 		ibv_match[nd++] = ibv_list[ret];
2882 	}
2883 	ibv_match[nd] = NULL;
2884 	if (!nd) {
2885 		/* No device matches, just complain and bail out. */
2886 		DRV_LOG(WARNING,
2887 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2888 			" are kernel drivers loaded?",
2889 			pci_dev->addr.domain, pci_dev->addr.bus,
2890 			pci_dev->addr.devid, pci_dev->addr.function);
2891 		rte_errno = ENOENT;
2892 		ret = -rte_errno;
2893 		goto exit;
2894 	}
2895 	if (nd == 1) {
2896 		/*
2897 		 * Found single matching device may have multiple ports.
2898 		 * Each port may be representor, we have to check the port
2899 		 * number and check the representors existence.
2900 		 */
2901 		if (nl_rdma >= 0)
2902 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2903 		if (!np)
2904 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2905 					 " ports number", ibv_match[0]->name);
2906 		if (bd >= 0 && !np) {
2907 			DRV_LOG(ERR, "can not get ports"
2908 				     " for bonding device");
2909 			rte_errno = ENOENT;
2910 			ret = -rte_errno;
2911 			goto exit;
2912 		}
2913 	}
2914 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2915 	if (bd >= 0) {
2916 		/*
2917 		 * This may happen if there is VF LAG kernel support and
2918 		 * application is compiled with older rdma_core library.
2919 		 */
2920 		DRV_LOG(ERR,
2921 			"No kernel/verbs support for VF LAG bonding found.");
2922 		rte_errno = ENOTSUP;
2923 		ret = -rte_errno;
2924 		goto exit;
2925 	}
2926 #endif
2927 	/*
2928 	 * Now we can determine the maximal
2929 	 * amount of devices to be spawned.
2930 	 */
2931 	list = rte_zmalloc("device spawn data",
2932 			 sizeof(struct mlx5_dev_spawn_data) *
2933 			 (np ? np : nd),
2934 			 RTE_CACHE_LINE_SIZE);
2935 	if (!list) {
2936 		DRV_LOG(ERR, "spawn data array allocation failure");
2937 		rte_errno = ENOMEM;
2938 		ret = -rte_errno;
2939 		goto exit;
2940 	}
2941 	if (bd >= 0 || np > 1) {
2942 		/*
2943 		 * Single IB device with multiple ports found,
2944 		 * it may be E-Switch master device and representors.
2945 		 * We have to perform identification trough the ports.
2946 		 */
2947 		assert(nl_rdma >= 0);
2948 		assert(ns == 0);
2949 		assert(nd == 1);
2950 		assert(np);
2951 		for (i = 1; i <= np; ++i) {
2952 			list[ns].max_port = np;
2953 			list[ns].ibv_port = i;
2954 			list[ns].ibv_dev = ibv_match[0];
2955 			list[ns].eth_dev = NULL;
2956 			list[ns].pci_dev = pci_dev;
2957 			list[ns].pf_bond = bd;
2958 			list[ns].ifindex = mlx5_nl_ifindex
2959 					(nl_rdma, list[ns].ibv_dev->name, i);
2960 			if (!list[ns].ifindex) {
2961 				/*
2962 				 * No network interface index found for the
2963 				 * specified port, it means there is no
2964 				 * representor on this port. It's OK,
2965 				 * there can be disabled ports, for example
2966 				 * if sriov_numvfs < sriov_totalvfs.
2967 				 */
2968 				continue;
2969 			}
2970 			ret = -1;
2971 			if (nl_route >= 0)
2972 				ret = mlx5_nl_switch_info
2973 					       (nl_route,
2974 						list[ns].ifindex,
2975 						&list[ns].info);
2976 			if (ret || (!list[ns].info.representor &&
2977 				    !list[ns].info.master)) {
2978 				/*
2979 				 * We failed to recognize representors with
2980 				 * Netlink, let's try to perform the task
2981 				 * with sysfs.
2982 				 */
2983 				ret =  mlx5_sysfs_switch_info
2984 						(list[ns].ifindex,
2985 						 &list[ns].info);
2986 			}
2987 			if (!ret && bd >= 0) {
2988 				switch (list[ns].info.name_type) {
2989 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2990 					if (list[ns].info.port_name == bd)
2991 						ns++;
2992 					break;
2993 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2994 					if (list[ns].info.pf_num == bd)
2995 						ns++;
2996 					break;
2997 				default:
2998 					break;
2999 				}
3000 				continue;
3001 			}
3002 			if (!ret && (list[ns].info.representor ^
3003 				     list[ns].info.master))
3004 				ns++;
3005 		}
3006 		if (!ns) {
3007 			DRV_LOG(ERR,
3008 				"unable to recognize master/representors"
3009 				" on the IB device with multiple ports");
3010 			rte_errno = ENOENT;
3011 			ret = -rte_errno;
3012 			goto exit;
3013 		}
3014 	} else {
3015 		/*
3016 		 * The existence of several matching entries (nd > 1) means
3017 		 * port representors have been instantiated. No existing Verbs
3018 		 * call nor sysfs entries can tell them apart, this can only
3019 		 * be done through Netlink calls assuming kernel drivers are
3020 		 * recent enough to support them.
3021 		 *
3022 		 * In the event of identification failure through Netlink,
3023 		 * try again through sysfs, then:
3024 		 *
3025 		 * 1. A single IB device matches (nd == 1) with single
3026 		 *    port (np=0/1) and is not a representor, assume
3027 		 *    no switch support.
3028 		 *
3029 		 * 2. Otherwise no safe assumptions can be made;
3030 		 *    complain louder and bail out.
3031 		 */
3032 		np = 1;
3033 		for (i = 0; i != nd; ++i) {
3034 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3035 			list[ns].max_port = 1;
3036 			list[ns].ibv_port = 1;
3037 			list[ns].ibv_dev = ibv_match[i];
3038 			list[ns].eth_dev = NULL;
3039 			list[ns].pci_dev = pci_dev;
3040 			list[ns].pf_bond = -1;
3041 			list[ns].ifindex = 0;
3042 			if (nl_rdma >= 0)
3043 				list[ns].ifindex = mlx5_nl_ifindex
3044 					(nl_rdma, list[ns].ibv_dev->name, 1);
3045 			if (!list[ns].ifindex) {
3046 				char ifname[IF_NAMESIZE];
3047 
3048 				/*
3049 				 * Netlink failed, it may happen with old
3050 				 * ib_core kernel driver (before 4.16).
3051 				 * We can assume there is old driver because
3052 				 * here we are processing single ports IB
3053 				 * devices. Let's try sysfs to retrieve
3054 				 * the ifindex. The method works for
3055 				 * master device only.
3056 				 */
3057 				if (nd > 1) {
3058 					/*
3059 					 * Multiple devices found, assume
3060 					 * representors, can not distinguish
3061 					 * master/representor and retrieve
3062 					 * ifindex via sysfs.
3063 					 */
3064 					continue;
3065 				}
3066 				ret = mlx5_get_master_ifname
3067 					(ibv_match[i]->ibdev_path, &ifname);
3068 				if (!ret)
3069 					list[ns].ifindex =
3070 						if_nametoindex(ifname);
3071 				if (!list[ns].ifindex) {
3072 					/*
3073 					 * No network interface index found
3074 					 * for the specified device, it means
3075 					 * there it is neither representor
3076 					 * nor master.
3077 					 */
3078 					continue;
3079 				}
3080 			}
3081 			ret = -1;
3082 			if (nl_route >= 0)
3083 				ret = mlx5_nl_switch_info
3084 					       (nl_route,
3085 						list[ns].ifindex,
3086 						&list[ns].info);
3087 			if (ret || (!list[ns].info.representor &&
3088 				    !list[ns].info.master)) {
3089 				/*
3090 				 * We failed to recognize representors with
3091 				 * Netlink, let's try to perform the task
3092 				 * with sysfs.
3093 				 */
3094 				ret =  mlx5_sysfs_switch_info
3095 						(list[ns].ifindex,
3096 						 &list[ns].info);
3097 			}
3098 			if (!ret && (list[ns].info.representor ^
3099 				     list[ns].info.master)) {
3100 				ns++;
3101 			} else if ((nd == 1) &&
3102 				   !list[ns].info.representor &&
3103 				   !list[ns].info.master) {
3104 				/*
3105 				 * Single IB device with
3106 				 * one physical port and
3107 				 * attached network device.
3108 				 * May be SRIOV is not enabled
3109 				 * or there is no representors.
3110 				 */
3111 				DRV_LOG(INFO, "no E-Switch support detected");
3112 				ns++;
3113 				break;
3114 			}
3115 		}
3116 		if (!ns) {
3117 			DRV_LOG(ERR,
3118 				"unable to recognize master/representors"
3119 				" on the multiple IB devices");
3120 			rte_errno = ENOENT;
3121 			ret = -rte_errno;
3122 			goto exit;
3123 		}
3124 	}
3125 	assert(ns);
3126 	/*
3127 	 * Sort list to probe devices in natural order for users convenience
3128 	 * (i.e. master first, then representors from lowest to highest ID).
3129 	 */
3130 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3131 	/* Default configuration. */
3132 	dev_config = (struct mlx5_dev_config){
3133 		.hw_padding = 0,
3134 		.mps = MLX5_ARG_UNSET,
3135 		.dbnc = MLX5_ARG_UNSET,
3136 		.rx_vec_en = 1,
3137 		.txq_inline_max = MLX5_ARG_UNSET,
3138 		.txq_inline_min = MLX5_ARG_UNSET,
3139 		.txq_inline_mpw = MLX5_ARG_UNSET,
3140 		.txqs_inline = MLX5_ARG_UNSET,
3141 		.vf_nl_en = 1,
3142 		.mr_ext_memseg_en = 1,
3143 		.mprq = {
3144 			.enabled = 0, /* Disabled by default. */
3145 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3146 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3147 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3148 		},
3149 		.dv_esw_en = 1,
3150 	};
3151 	/* Device specific configuration. */
3152 	switch (pci_dev->id.device_id) {
3153 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3154 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3155 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3156 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3157 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3158 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3159 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3160 		dev_config.vf = 1;
3161 		break;
3162 	default:
3163 		break;
3164 	}
3165 	for (i = 0; i != ns; ++i) {
3166 		uint32_t restore;
3167 
3168 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3169 						 &list[i],
3170 						 dev_config);
3171 		if (!list[i].eth_dev) {
3172 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3173 				break;
3174 			/* Device is disabled or already spawned. Ignore it. */
3175 			continue;
3176 		}
3177 		restore = list[i].eth_dev->data->dev_flags;
3178 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3179 		/* Restore non-PCI flags cleared by the above call. */
3180 		list[i].eth_dev->data->dev_flags |= restore;
3181 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3182 		rte_eth_dev_probing_finish(list[i].eth_dev);
3183 	}
3184 	if (i != ns) {
3185 		DRV_LOG(ERR,
3186 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3187 			" encountering an error: %s",
3188 			pci_dev->addr.domain, pci_dev->addr.bus,
3189 			pci_dev->addr.devid, pci_dev->addr.function,
3190 			strerror(rte_errno));
3191 		ret = -rte_errno;
3192 		/* Roll back. */
3193 		while (i--) {
3194 			if (!list[i].eth_dev)
3195 				continue;
3196 			mlx5_dev_close(list[i].eth_dev);
3197 			/* mac_addrs must not be freed because in dev_private */
3198 			list[i].eth_dev->data->mac_addrs = NULL;
3199 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3200 		}
3201 		/* Restore original error. */
3202 		rte_errno = -ret;
3203 	} else {
3204 		ret = 0;
3205 	}
3206 exit:
3207 	/*
3208 	 * Do the routine cleanup:
3209 	 * - close opened Netlink sockets
3210 	 * - free allocated spawn data array
3211 	 * - free the Infiniband device list
3212 	 */
3213 	if (nl_rdma >= 0)
3214 		close(nl_rdma);
3215 	if (nl_route >= 0)
3216 		close(nl_route);
3217 	if (list)
3218 		rte_free(list);
3219 	assert(ibv_list);
3220 	mlx5_glue->free_device_list(ibv_list);
3221 	return ret;
3222 }
3223 
3224 /**
3225  * Look for the ethernet device belonging to mlx5 driver.
3226  *
3227  * @param[in] port_id
3228  *   port_id to start looking for device.
3229  * @param[in] pci_dev
3230  *   Pointer to the hint PCI device. When device is being probed
3231  *   the its siblings (master and preceding representors might
3232  *   not have assigned driver yet (because the mlx5_pci_probe()
3233  *   is not completed yet, for this case match on hint PCI
3234  *   device may be used to detect sibling device.
3235  *
3236  * @return
3237  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3238  */
3239 uint16_t
3240 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3241 {
3242 	while (port_id < RTE_MAX_ETHPORTS) {
3243 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3244 
3245 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3246 		    dev->device &&
3247 		    (dev->device == &pci_dev->device ||
3248 		     (dev->device->driver &&
3249 		     dev->device->driver->name &&
3250 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3251 			break;
3252 		port_id++;
3253 	}
3254 	if (port_id >= RTE_MAX_ETHPORTS)
3255 		return RTE_MAX_ETHPORTS;
3256 	return port_id;
3257 }
3258 
3259 /**
3260  * DPDK callback to remove a PCI device.
3261  *
3262  * This function removes all Ethernet devices belong to a given PCI device.
3263  *
3264  * @param[in] pci_dev
3265  *   Pointer to the PCI device.
3266  *
3267  * @return
3268  *   0 on success, the function cannot fail.
3269  */
3270 static int
3271 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3272 {
3273 	uint16_t port_id;
3274 
3275 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3276 		rte_eth_dev_close(port_id);
3277 	return 0;
3278 }
3279 
3280 static const struct rte_pci_id mlx5_pci_id_map[] = {
3281 	{
3282 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3283 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3284 	},
3285 	{
3286 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3287 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3288 	},
3289 	{
3290 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3291 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3292 	},
3293 	{
3294 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3295 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3296 	},
3297 	{
3298 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3299 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3300 	},
3301 	{
3302 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3303 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3304 	},
3305 	{
3306 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3307 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3308 	},
3309 	{
3310 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3311 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3312 	},
3313 	{
3314 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3315 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3316 	},
3317 	{
3318 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3319 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3320 	},
3321 	{
3322 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3323 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3324 	},
3325 	{
3326 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3327 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3328 	},
3329 	{
3330 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3331 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3332 	},
3333 	{
3334 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3335 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3336 	},
3337 	{
3338 		.vendor_id = 0
3339 	}
3340 };
3341 
3342 static struct rte_pci_driver mlx5_driver = {
3343 	.driver = {
3344 		.name = MLX5_DRIVER_NAME
3345 	},
3346 	.id_table = mlx5_pci_id_map,
3347 	.probe = mlx5_pci_probe,
3348 	.remove = mlx5_pci_remove,
3349 	.dma_map = mlx5_dma_map,
3350 	.dma_unmap = mlx5_dma_unmap,
3351 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3352 		     RTE_PCI_DRV_PROBE_AGAIN,
3353 };
3354 
3355 #ifdef RTE_IBVERBS_LINK_DLOPEN
3356 
3357 /**
3358  * Suffix RTE_EAL_PMD_PATH with "-glue".
3359  *
3360  * This function performs a sanity check on RTE_EAL_PMD_PATH before
3361  * suffixing its last component.
3362  *
3363  * @param buf[out]
3364  *   Output buffer, should be large enough otherwise NULL is returned.
3365  * @param size
3366  *   Size of @p out.
3367  *
3368  * @return
3369  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
3370  */
3371 static char *
3372 mlx5_glue_path(char *buf, size_t size)
3373 {
3374 	static const char *const bad[] = { "/", ".", "..", NULL };
3375 	const char *path = RTE_EAL_PMD_PATH;
3376 	size_t len = strlen(path);
3377 	size_t off;
3378 	int i;
3379 
3380 	while (len && path[len - 1] == '/')
3381 		--len;
3382 	for (off = len; off && path[off - 1] != '/'; --off)
3383 		;
3384 	for (i = 0; bad[i]; ++i)
3385 		if (!strncmp(path + off, bad[i], (int)(len - off)))
3386 			goto error;
3387 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
3388 	if (i == -1 || (size_t)i >= size)
3389 		goto error;
3390 	return buf;
3391 error:
3392 	DRV_LOG(ERR,
3393 		"unable to append \"-glue\" to last component of"
3394 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
3395 		" please re-configure DPDK");
3396 	return NULL;
3397 }
3398 
3399 /**
3400  * Initialization routine for run-time dependency on rdma-core.
3401  */
3402 static int
3403 mlx5_glue_init(void)
3404 {
3405 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
3406 	const char *path[] = {
3407 		/*
3408 		 * A basic security check is necessary before trusting
3409 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
3410 		 */
3411 		(geteuid() == getuid() && getegid() == getgid() ?
3412 		 getenv("MLX5_GLUE_PATH") : NULL),
3413 		/*
3414 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
3415 		 * variant, otherwise let dlopen() look up libraries on its
3416 		 * own.
3417 		 */
3418 		(*RTE_EAL_PMD_PATH ?
3419 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
3420 	};
3421 	unsigned int i = 0;
3422 	void *handle = NULL;
3423 	void **sym;
3424 	const char *dlmsg;
3425 
3426 	while (!handle && i != RTE_DIM(path)) {
3427 		const char *end;
3428 		size_t len;
3429 		int ret;
3430 
3431 		if (!path[i]) {
3432 			++i;
3433 			continue;
3434 		}
3435 		end = strpbrk(path[i], ":;");
3436 		if (!end)
3437 			end = path[i] + strlen(path[i]);
3438 		len = end - path[i];
3439 		ret = 0;
3440 		do {
3441 			char name[ret + 1];
3442 
3443 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
3444 				       (int)len, path[i],
3445 				       (!len || *(end - 1) == '/') ? "" : "/");
3446 			if (ret == -1)
3447 				break;
3448 			if (sizeof(name) != (size_t)ret + 1)
3449 				continue;
3450 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
3451 				name);
3452 			handle = dlopen(name, RTLD_LAZY);
3453 			break;
3454 		} while (1);
3455 		path[i] = end + 1;
3456 		if (!*end)
3457 			++i;
3458 	}
3459 	if (!handle) {
3460 		rte_errno = EINVAL;
3461 		dlmsg = dlerror();
3462 		if (dlmsg)
3463 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
3464 		goto glue_error;
3465 	}
3466 	sym = dlsym(handle, "mlx5_glue");
3467 	if (!sym || !*sym) {
3468 		rte_errno = EINVAL;
3469 		dlmsg = dlerror();
3470 		if (dlmsg)
3471 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
3472 		goto glue_error;
3473 	}
3474 	mlx5_glue = *sym;
3475 	return 0;
3476 glue_error:
3477 	if (handle)
3478 		dlclose(handle);
3479 	DRV_LOG(WARNING,
3480 		"cannot initialize PMD due to missing run-time dependency on"
3481 		" rdma-core libraries (libibverbs, libmlx5)");
3482 	return -rte_errno;
3483 }
3484 
3485 #endif
3486 
3487 /**
3488  * Driver initialization routine.
3489  */
3490 RTE_INIT(rte_mlx5_pmd_init)
3491 {
3492 	/* Initialize driver log type. */
3493 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3494 	if (mlx5_logtype >= 0)
3495 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3496 
3497 	/* Build the static tables for Verbs conversion. */
3498 	mlx5_set_ptype_table();
3499 	mlx5_set_cksum_table();
3500 	mlx5_set_swp_types_table();
3501 	/*
3502 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3503 	 * huge pages. Calling ibv_fork_init() during init allows
3504 	 * applications to use fork() safely for purposes other than
3505 	 * using this PMD, which is not supported in forked processes.
3506 	 */
3507 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3508 	/* Match the size of Rx completion entry to the size of a cacheline. */
3509 	if (RTE_CACHE_LINE_SIZE == 128)
3510 		setenv("MLX5_CQE_SIZE", "128", 0);
3511 	/*
3512 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
3513 	 * cleanup all the Verbs resources even when the device was removed.
3514 	 */
3515 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
3516 #ifdef RTE_IBVERBS_LINK_DLOPEN
3517 	if (mlx5_glue_init())
3518 		return;
3519 	assert(mlx5_glue);
3520 #endif
3521 #ifndef NDEBUG
3522 	/* Glue structure must not contain any NULL pointers. */
3523 	{
3524 		unsigned int i;
3525 
3526 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
3527 			assert(((const void *const *)mlx5_glue)[i]);
3528 	}
3529 #endif
3530 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
3531 		DRV_LOG(ERR,
3532 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
3533 			mlx5_glue->version, MLX5_GLUE_VERSION);
3534 		return;
3535 	}
3536 	mlx5_glue->fork_init();
3537 	rte_pci_register(&mlx5_driver);
3538 }
3539 
3540 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3541 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3542 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3543