xref: /dpdk/drivers/net/mlx5/mlx5.c (revision d56ec3dcad056c47cef4e837d5191d04c936d87e)
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(int value)
452 {
453 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
454 	/* Restore the original environment variable state. */
455 	if (value == MLX5_ARG_UNSET)
456 		unsetenv(MLX5_SHUT_UP_BF);
457 	else
458 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
459 }
460 
461 /**
462  * Allocate shared IB device context. If there is multiport device the
463  * master and representors will share this context, if there is single
464  * port dedicated IB device, the context will be used by only given
465  * port due to unification.
466  *
467  * Routine first searches the context for the specified IB device name,
468  * if found the shared context assumed and reference counter is incremented.
469  * If no context found the new one is created and initialized with specified
470  * IB device context and parameters.
471  *
472  * @param[in] spawn
473  *   Pointer to the IB device attributes (name, port, etc).
474  * @param[in] config
475  *   Pointer to device configuration structure.
476  *
477  * @return
478  *   Pointer to mlx5_ibv_shared object on success,
479  *   otherwise NULL and rte_errno is set.
480  */
481 static struct mlx5_ibv_shared *
482 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
483 			const struct mlx5_dev_config *config)
484 {
485 	struct mlx5_ibv_shared *sh;
486 	int dbmap_env;
487 	int err = 0;
488 	uint32_t i;
489 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
490 	struct mlx5_devx_tis_attr tis_attr = { 0 };
491 #endif
492 
493 	assert(spawn);
494 	/* Secondary process should not create the shared context. */
495 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
496 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
497 	/* Search for IB context by device name. */
498 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
499 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
500 			sh->refcnt++;
501 			goto exit;
502 		}
503 	}
504 	/* No device found, we have to create new shared context. */
505 	assert(spawn->max_port);
506 	sh = rte_zmalloc("ethdev shared ib context",
507 			 sizeof(struct mlx5_ibv_shared) +
508 			 spawn->max_port *
509 			 sizeof(struct mlx5_ibv_shared_port),
510 			 RTE_CACHE_LINE_SIZE);
511 	if (!sh) {
512 		DRV_LOG(ERR, "shared context allocation failure");
513 		rte_errno  = ENOMEM;
514 		goto exit;
515 	}
516 	/*
517 	 * Configure environment variable "MLX5_BF_SHUT_UP"
518 	 * before the device creation. The rdma_core library
519 	 * checks the variable at device creation and
520 	 * stores the result internally.
521 	 */
522 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
523 	/* Try to open IB device with DV first, then usual Verbs. */
524 	errno = 0;
525 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
526 	if (sh->ctx) {
527 		sh->devx = 1;
528 		DRV_LOG(DEBUG, "DevX is supported");
529 		/* The device is created, no need for environment. */
530 		mlx5_restore_doorbell_mapping_env(dbmap_env);
531 	} else {
532 		/* The environment variable is still configured. */
533 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
534 		err = errno ? errno : ENODEV;
535 		/*
536 		 * The environment variable is not needed anymore,
537 		 * all device creation attempts are completed.
538 		 */
539 		mlx5_restore_doorbell_mapping_env(dbmap_env);
540 		if (!sh->ctx)
541 			goto error;
542 		DRV_LOG(DEBUG, "DevX is NOT supported");
543 	}
544 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
545 	if (err) {
546 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
547 		goto error;
548 	}
549 	sh->refcnt = 1;
550 	sh->max_port = spawn->max_port;
551 	strncpy(sh->ibdev_name, sh->ctx->device->name,
552 		sizeof(sh->ibdev_name));
553 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
554 		sizeof(sh->ibdev_path));
555 	pthread_mutex_init(&sh->intr_mutex, NULL);
556 	/*
557 	 * Setting port_id to max unallowed value means
558 	 * there is no interrupt subhandler installed for
559 	 * the given port index i.
560 	 */
561 	for (i = 0; i < sh->max_port; i++) {
562 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
563 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
564 	}
565 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
566 	if (sh->pd == NULL) {
567 		DRV_LOG(ERR, "PD allocation failure");
568 		err = ENOMEM;
569 		goto error;
570 	}
571 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
572 	if (sh->devx) {
573 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
574 		if (err) {
575 			DRV_LOG(ERR, "Fail to extract pdn from PD");
576 			goto error;
577 		}
578 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
579 		if (!sh->td) {
580 			DRV_LOG(ERR, "TD allocation failure");
581 			err = ENOMEM;
582 			goto error;
583 		}
584 		tis_attr.transport_domain = sh->td->id;
585 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
586 		if (!sh->tis) {
587 			DRV_LOG(ERR, "TIS allocation failure");
588 			err = ENOMEM;
589 			goto error;
590 		}
591 	}
592 	sh->flow_id_pool = mlx5_flow_id_pool_alloc();
593 	if (!sh->flow_id_pool) {
594 		DRV_LOG(ERR, "can't create flow id pool");
595 		err = ENOMEM;
596 		goto error;
597 	}
598 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
599 	/*
600 	 * Once the device is added to the list of memory event
601 	 * callback, its global MR cache table cannot be expanded
602 	 * on the fly because of deadlock. If it overflows, lookup
603 	 * should be done by searching MR list linearly, which is slow.
604 	 *
605 	 * At this point the device is not added to the memory
606 	 * event list yet, context is just being created.
607 	 */
608 	err = mlx5_mr_btree_init(&sh->mr.cache,
609 				 MLX5_MR_BTREE_CACHE_N * 2,
610 				 spawn->pci_dev->device.numa_node);
611 	if (err) {
612 		err = rte_errno;
613 		goto error;
614 	}
615 	mlx5_flow_counters_mng_init(sh);
616 	/* Add device to memory callback list. */
617 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
618 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
619 			 sh, mem_event_cb);
620 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
621 	/* Add context to the global device list. */
622 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
623 exit:
624 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
625 	return sh;
626 error:
627 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
628 	assert(sh);
629 	if (sh->tis)
630 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
631 	if (sh->td)
632 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
633 	if (sh->pd)
634 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
635 	if (sh->ctx)
636 		claim_zero(mlx5_glue->close_device(sh->ctx));
637 	if (sh->flow_id_pool)
638 		mlx5_flow_id_pool_release(sh->flow_id_pool);
639 	rte_free(sh);
640 	assert(err > 0);
641 	rte_errno = err;
642 	return NULL;
643 }
644 
645 /**
646  * Free shared IB device context. Decrement counter and if zero free
647  * all allocated resources and close handles.
648  *
649  * @param[in] sh
650  *   Pointer to mlx5_ibv_shared object to free
651  */
652 static void
653 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
654 {
655 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
656 #ifndef NDEBUG
657 	/* Check the object presence in the list. */
658 	struct mlx5_ibv_shared *lctx;
659 
660 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
661 		if (lctx == sh)
662 			break;
663 	assert(lctx);
664 	if (lctx != sh) {
665 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
666 		goto exit;
667 	}
668 #endif
669 	assert(sh);
670 	assert(sh->refcnt);
671 	/* Secondary process should not free the shared context. */
672 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
673 	if (--sh->refcnt)
674 		goto exit;
675 	/* Release created Memory Regions. */
676 	mlx5_mr_release(sh);
677 	/* Remove from memory callback device list. */
678 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
679 	LIST_REMOVE(sh, mem_event_cb);
680 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
681 	/* Remove context from the global device list. */
682 	LIST_REMOVE(sh, next);
683 	/*
684 	 *  Ensure there is no async event handler installed.
685 	 *  Only primary process handles async device events.
686 	 **/
687 	mlx5_flow_counters_mng_close(sh);
688 	assert(!sh->intr_cnt);
689 	if (sh->intr_cnt)
690 		mlx5_intr_callback_unregister
691 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
692 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
693 	if (sh->devx_intr_cnt) {
694 		if (sh->intr_handle_devx.fd)
695 			rte_intr_callback_unregister(&sh->intr_handle_devx,
696 					  mlx5_dev_interrupt_handler_devx, sh);
697 		if (sh->devx_comp)
698 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
699 	}
700 #endif
701 	pthread_mutex_destroy(&sh->intr_mutex);
702 	if (sh->pd)
703 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
704 	if (sh->tis)
705 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
706 	if (sh->td)
707 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
708 	if (sh->ctx)
709 		claim_zero(mlx5_glue->close_device(sh->ctx));
710 	if (sh->flow_id_pool)
711 		mlx5_flow_id_pool_release(sh->flow_id_pool);
712 	rte_free(sh);
713 exit:
714 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
715 }
716 
717 /**
718  * Initialize DR related data within private structure.
719  * Routine checks the reference counter and does actual
720  * resources creation/initialization only if counter is zero.
721  *
722  * @param[in] priv
723  *   Pointer to the private device data structure.
724  *
725  * @return
726  *   Zero on success, positive error code otherwise.
727  */
728 static int
729 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
730 {
731 #ifdef HAVE_MLX5DV_DR
732 	struct mlx5_ibv_shared *sh = priv->sh;
733 	int err = 0;
734 	void *domain;
735 	char s[MLX5_HLIST_NAMESIZE];
736 
737 	assert(sh);
738 	if (sh->dv_refcnt) {
739 		/* Shared DV/DR structures is already initialized. */
740 		sh->dv_refcnt++;
741 		priv->dr_shared = 1;
742 		return 0;
743 	}
744 	/* Reference counter is zero, we should initialize structures. */
745 	domain = mlx5_glue->dr_create_domain(sh->ctx,
746 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
747 	if (!domain) {
748 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
749 		err = errno;
750 		goto error;
751 	}
752 	sh->rx_domain = domain;
753 	domain = mlx5_glue->dr_create_domain(sh->ctx,
754 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
755 	if (!domain) {
756 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
757 		err = errno;
758 		goto error;
759 	}
760 	pthread_mutex_init(&sh->dv_mutex, NULL);
761 	sh->tx_domain = domain;
762 #ifdef HAVE_MLX5DV_DR_ESWITCH
763 	if (priv->config.dv_esw_en) {
764 		domain  = mlx5_glue->dr_create_domain
765 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
766 		if (!domain) {
767 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
768 			err = errno;
769 			goto error;
770 		}
771 		sh->fdb_domain = domain;
772 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
773 	}
774 #endif
775 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
776 	sh->flow_tbls = mlx5_hlist_create(s,
777 					  MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
778 	if (!sh->flow_tbls) {
779 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
780 		err = -ENOMEM;
781 		goto error;
782 	}
783 	/* create tags hash list table. */
784 	snprintf(s, sizeof(s), "%s_tags", priv->sh->ibdev_name);
785 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
786 	if (!sh->flow_tbls) {
787 		DRV_LOG(ERR, "tags with hash creation failed.\n");
788 		err = -ENOMEM;
789 		goto error;
790 	}
791 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
792 	sh->dv_refcnt++;
793 	priv->dr_shared = 1;
794 	return 0;
795 
796 error:
797        /* Rollback the created objects. */
798 	if (sh->rx_domain) {
799 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
800 		sh->rx_domain = NULL;
801 	}
802 	if (sh->tx_domain) {
803 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
804 		sh->tx_domain = NULL;
805 	}
806 	if (sh->fdb_domain) {
807 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
808 		sh->fdb_domain = NULL;
809 	}
810 	if (sh->flow_tbls) {
811 		mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
812 		sh->flow_tbls = NULL;
813 	}
814 	if (sh->esw_drop_action) {
815 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
816 		sh->esw_drop_action = NULL;
817 	}
818 	if (sh->pop_vlan_action) {
819 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
820 		sh->pop_vlan_action = NULL;
821 	}
822 	return err;
823 #else
824 	(void)priv;
825 	return 0;
826 #endif
827 }
828 
829 /**
830  * Destroy DR related data within private structure.
831  *
832  * @param[in] priv
833  *   Pointer to the private device data structure.
834  */
835 static void
836 mlx5_free_shared_dr(struct mlx5_priv *priv)
837 {
838 #ifdef HAVE_MLX5DV_DR
839 	struct mlx5_ibv_shared *sh;
840 
841 	if (!priv->dr_shared)
842 		return;
843 	priv->dr_shared = 0;
844 	sh = priv->sh;
845 	assert(sh);
846 	assert(sh->dv_refcnt);
847 	if (sh->dv_refcnt && --sh->dv_refcnt)
848 		return;
849 	if (sh->flow_tbls) {
850 		/* flow table entries should be handled properly before. */
851 		mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
852 		sh->flow_tbls = NULL;
853 	}
854 	if (sh->tag_table) {
855 		/* tags should be destroyed with flow before. */
856 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
857 		sh->tag_table = NULL;
858 	}
859 	if (sh->rx_domain) {
860 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
861 		sh->rx_domain = NULL;
862 	}
863 	if (sh->tx_domain) {
864 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
865 		sh->tx_domain = NULL;
866 	}
867 #ifdef HAVE_MLX5DV_DR_ESWITCH
868 	if (sh->fdb_domain) {
869 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
870 		sh->fdb_domain = NULL;
871 	}
872 	if (sh->esw_drop_action) {
873 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
874 		sh->esw_drop_action = NULL;
875 	}
876 #endif
877 	if (sh->pop_vlan_action) {
878 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
879 		sh->pop_vlan_action = NULL;
880 	}
881 	pthread_mutex_destroy(&sh->dv_mutex);
882 #else
883 	(void)priv;
884 #endif
885 }
886 
887 /**
888  * Initialize shared data between primary and secondary process.
889  *
890  * A memzone is reserved by primary process and secondary processes attach to
891  * the memzone.
892  *
893  * @return
894  *   0 on success, a negative errno value otherwise and rte_errno is set.
895  */
896 static int
897 mlx5_init_shared_data(void)
898 {
899 	const struct rte_memzone *mz;
900 	int ret = 0;
901 
902 	rte_spinlock_lock(&mlx5_shared_data_lock);
903 	if (mlx5_shared_data == NULL) {
904 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
905 			/* Allocate shared memory. */
906 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
907 						 sizeof(*mlx5_shared_data),
908 						 SOCKET_ID_ANY, 0);
909 			if (mz == NULL) {
910 				DRV_LOG(ERR,
911 					"Cannot allocate mlx5 shared data");
912 				ret = -rte_errno;
913 				goto error;
914 			}
915 			mlx5_shared_data = mz->addr;
916 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
917 			rte_spinlock_init(&mlx5_shared_data->lock);
918 		} else {
919 			/* Lookup allocated shared memory. */
920 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
921 			if (mz == NULL) {
922 				DRV_LOG(ERR,
923 					"Cannot attach mlx5 shared data");
924 				ret = -rte_errno;
925 				goto error;
926 			}
927 			mlx5_shared_data = mz->addr;
928 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
929 		}
930 	}
931 error:
932 	rte_spinlock_unlock(&mlx5_shared_data_lock);
933 	return ret;
934 }
935 
936 /**
937  * Retrieve integer value from environment variable.
938  *
939  * @param[in] name
940  *   Environment variable name.
941  *
942  * @return
943  *   Integer value, 0 if the variable is not set.
944  */
945 int
946 mlx5_getenv_int(const char *name)
947 {
948 	const char *val = getenv(name);
949 
950 	if (val == NULL)
951 		return 0;
952 	return atoi(val);
953 }
954 
955 /**
956  * Verbs callback to allocate a memory. This function should allocate the space
957  * according to the size provided residing inside a huge page.
958  * Please note that all allocation must respect the alignment from libmlx5
959  * (i.e. currently sysconf(_SC_PAGESIZE)).
960  *
961  * @param[in] size
962  *   The size in bytes of the memory to allocate.
963  * @param[in] data
964  *   A pointer to the callback data.
965  *
966  * @return
967  *   Allocated buffer, NULL otherwise and rte_errno is set.
968  */
969 static void *
970 mlx5_alloc_verbs_buf(size_t size, void *data)
971 {
972 	struct mlx5_priv *priv = data;
973 	void *ret;
974 	size_t alignment = sysconf(_SC_PAGESIZE);
975 	unsigned int socket = SOCKET_ID_ANY;
976 
977 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
978 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
979 
980 		socket = ctrl->socket;
981 	} else if (priv->verbs_alloc_ctx.type ==
982 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
983 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
984 
985 		socket = ctrl->socket;
986 	}
987 	assert(data != NULL);
988 	ret = rte_malloc_socket(__func__, size, alignment, socket);
989 	if (!ret && size)
990 		rte_errno = ENOMEM;
991 	return ret;
992 }
993 
994 /**
995  * Verbs callback to free a memory.
996  *
997  * @param[in] ptr
998  *   A pointer to the memory to free.
999  * @param[in] data
1000  *   A pointer to the callback data.
1001  */
1002 static void
1003 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1004 {
1005 	assert(data != NULL);
1006 	rte_free(ptr);
1007 }
1008 
1009 /**
1010  * DPDK callback to add udp tunnel port
1011  *
1012  * @param[in] dev
1013  *   A pointer to eth_dev
1014  * @param[in] udp_tunnel
1015  *   A pointer to udp tunnel
1016  *
1017  * @return
1018  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1019  */
1020 int
1021 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1022 			 struct rte_eth_udp_tunnel *udp_tunnel)
1023 {
1024 	assert(udp_tunnel != NULL);
1025 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1026 	    udp_tunnel->udp_port == 4789)
1027 		return 0;
1028 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1029 	    udp_tunnel->udp_port == 4790)
1030 		return 0;
1031 	return -ENOTSUP;
1032 }
1033 
1034 /**
1035  * Initialize process private data structure.
1036  *
1037  * @param dev
1038  *   Pointer to Ethernet device structure.
1039  *
1040  * @return
1041  *   0 on success, a negative errno value otherwise and rte_errno is set.
1042  */
1043 int
1044 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1045 {
1046 	struct mlx5_priv *priv = dev->data->dev_private;
1047 	struct mlx5_proc_priv *ppriv;
1048 	size_t ppriv_size;
1049 
1050 	/*
1051 	 * UAR register table follows the process private structure. BlueFlame
1052 	 * registers for Tx queues are stored in the table.
1053 	 */
1054 	ppriv_size =
1055 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1056 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1057 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1058 	if (!ppriv) {
1059 		rte_errno = ENOMEM;
1060 		return -rte_errno;
1061 	}
1062 	ppriv->uar_table_sz = ppriv_size;
1063 	dev->process_private = ppriv;
1064 	return 0;
1065 }
1066 
1067 /**
1068  * Un-initialize process private data structure.
1069  *
1070  * @param dev
1071  *   Pointer to Ethernet device structure.
1072  */
1073 static void
1074 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1075 {
1076 	if (!dev->process_private)
1077 		return;
1078 	rte_free(dev->process_private);
1079 	dev->process_private = NULL;
1080 }
1081 
1082 /**
1083  * DPDK callback to close the device.
1084  *
1085  * Destroy all queues and objects, free memory.
1086  *
1087  * @param dev
1088  *   Pointer to Ethernet device structure.
1089  */
1090 static void
1091 mlx5_dev_close(struct rte_eth_dev *dev)
1092 {
1093 	struct mlx5_priv *priv = dev->data->dev_private;
1094 	unsigned int i;
1095 	int ret;
1096 
1097 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1098 		dev->data->port_id,
1099 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1100 	/* In case mlx5_dev_stop() has not been called. */
1101 	mlx5_dev_interrupt_handler_uninstall(dev);
1102 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1103 	mlx5_traffic_disable(dev);
1104 	mlx5_flow_flush(dev, NULL);
1105 	mlx5_flow_meter_flush(dev, NULL);
1106 	/* Prevent crashes when queues are still in use. */
1107 	dev->rx_pkt_burst = removed_rx_burst;
1108 	dev->tx_pkt_burst = removed_tx_burst;
1109 	rte_wmb();
1110 	/* Disable datapath on secondary process. */
1111 	mlx5_mp_req_stop_rxtx(dev);
1112 	if (priv->rxqs != NULL) {
1113 		/* XXX race condition if mlx5_rx_burst() is still running. */
1114 		usleep(1000);
1115 		for (i = 0; (i != priv->rxqs_n); ++i)
1116 			mlx5_rxq_release(dev, i);
1117 		priv->rxqs_n = 0;
1118 		priv->rxqs = NULL;
1119 	}
1120 	if (priv->txqs != NULL) {
1121 		/* XXX race condition if mlx5_tx_burst() is still running. */
1122 		usleep(1000);
1123 		for (i = 0; (i != priv->txqs_n); ++i)
1124 			mlx5_txq_release(dev, i);
1125 		priv->txqs_n = 0;
1126 		priv->txqs = NULL;
1127 	}
1128 	mlx5_proc_priv_uninit(dev);
1129 	if (priv->mreg_cp_tbl)
1130 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1131 	mlx5_mprq_free_mp(dev);
1132 	mlx5_free_shared_dr(priv);
1133 	if (priv->rss_conf.rss_key != NULL)
1134 		rte_free(priv->rss_conf.rss_key);
1135 	if (priv->reta_idx != NULL)
1136 		rte_free(priv->reta_idx);
1137 	if (priv->config.vf)
1138 		mlx5_nl_mac_addr_flush(dev);
1139 	if (priv->nl_socket_route >= 0)
1140 		close(priv->nl_socket_route);
1141 	if (priv->nl_socket_rdma >= 0)
1142 		close(priv->nl_socket_rdma);
1143 	if (priv->vmwa_context)
1144 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1145 	if (priv->sh) {
1146 		/*
1147 		 * Free the shared context in last turn, because the cleanup
1148 		 * routines above may use some shared fields, like
1149 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1150 		 * ifindex if Netlink fails.
1151 		 */
1152 		mlx5_free_shared_ibctx(priv->sh);
1153 		priv->sh = NULL;
1154 	}
1155 	ret = mlx5_hrxq_verify(dev);
1156 	if (ret)
1157 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1158 			dev->data->port_id);
1159 	ret = mlx5_ind_table_obj_verify(dev);
1160 	if (ret)
1161 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1162 			dev->data->port_id);
1163 	ret = mlx5_rxq_obj_verify(dev);
1164 	if (ret)
1165 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1166 			dev->data->port_id);
1167 	ret = mlx5_rxq_verify(dev);
1168 	if (ret)
1169 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1170 			dev->data->port_id);
1171 	ret = mlx5_txq_obj_verify(dev);
1172 	if (ret)
1173 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1174 			dev->data->port_id);
1175 	ret = mlx5_txq_verify(dev);
1176 	if (ret)
1177 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1178 			dev->data->port_id);
1179 	ret = mlx5_flow_verify(dev);
1180 	if (ret)
1181 		DRV_LOG(WARNING, "port %u some flows still remain",
1182 			dev->data->port_id);
1183 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1184 		unsigned int c = 0;
1185 		uint16_t port_id;
1186 
1187 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1188 			struct mlx5_priv *opriv =
1189 				rte_eth_devices[port_id].data->dev_private;
1190 
1191 			if (!opriv ||
1192 			    opriv->domain_id != priv->domain_id ||
1193 			    &rte_eth_devices[port_id] == dev)
1194 				continue;
1195 			++c;
1196 			break;
1197 		}
1198 		if (!c)
1199 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1200 	}
1201 	memset(priv, 0, sizeof(*priv));
1202 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1203 	/*
1204 	 * Reset mac_addrs to NULL such that it is not freed as part of
1205 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1206 	 * it is freed when dev_private is freed.
1207 	 */
1208 	dev->data->mac_addrs = NULL;
1209 }
1210 
1211 const struct eth_dev_ops mlx5_dev_ops = {
1212 	.dev_configure = mlx5_dev_configure,
1213 	.dev_start = mlx5_dev_start,
1214 	.dev_stop = mlx5_dev_stop,
1215 	.dev_set_link_down = mlx5_set_link_down,
1216 	.dev_set_link_up = mlx5_set_link_up,
1217 	.dev_close = mlx5_dev_close,
1218 	.promiscuous_enable = mlx5_promiscuous_enable,
1219 	.promiscuous_disable = mlx5_promiscuous_disable,
1220 	.allmulticast_enable = mlx5_allmulticast_enable,
1221 	.allmulticast_disable = mlx5_allmulticast_disable,
1222 	.link_update = mlx5_link_update,
1223 	.stats_get = mlx5_stats_get,
1224 	.stats_reset = mlx5_stats_reset,
1225 	.xstats_get = mlx5_xstats_get,
1226 	.xstats_reset = mlx5_xstats_reset,
1227 	.xstats_get_names = mlx5_xstats_get_names,
1228 	.fw_version_get = mlx5_fw_version_get,
1229 	.dev_infos_get = mlx5_dev_infos_get,
1230 	.read_clock = mlx5_read_clock,
1231 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1232 	.vlan_filter_set = mlx5_vlan_filter_set,
1233 	.rx_queue_setup = mlx5_rx_queue_setup,
1234 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1235 	.tx_queue_setup = mlx5_tx_queue_setup,
1236 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1237 	.rx_queue_release = mlx5_rx_queue_release,
1238 	.tx_queue_release = mlx5_tx_queue_release,
1239 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1240 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1241 	.mac_addr_remove = mlx5_mac_addr_remove,
1242 	.mac_addr_add = mlx5_mac_addr_add,
1243 	.mac_addr_set = mlx5_mac_addr_set,
1244 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1245 	.mtu_set = mlx5_dev_set_mtu,
1246 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1247 	.vlan_offload_set = mlx5_vlan_offload_set,
1248 	.reta_update = mlx5_dev_rss_reta_update,
1249 	.reta_query = mlx5_dev_rss_reta_query,
1250 	.rss_hash_update = mlx5_rss_hash_update,
1251 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1252 	.filter_ctrl = mlx5_dev_filter_ctrl,
1253 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1254 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1255 	.rx_queue_count = mlx5_rx_queue_count,
1256 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1257 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1258 	.is_removed = mlx5_is_removed,
1259 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1260 	.get_module_info = mlx5_get_module_info,
1261 	.get_module_eeprom = mlx5_get_module_eeprom,
1262 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1263 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1264 };
1265 
1266 /* Available operations from secondary process. */
1267 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1268 	.stats_get = mlx5_stats_get,
1269 	.stats_reset = mlx5_stats_reset,
1270 	.xstats_get = mlx5_xstats_get,
1271 	.xstats_reset = mlx5_xstats_reset,
1272 	.xstats_get_names = mlx5_xstats_get_names,
1273 	.fw_version_get = mlx5_fw_version_get,
1274 	.dev_infos_get = mlx5_dev_infos_get,
1275 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1276 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1277 	.get_module_info = mlx5_get_module_info,
1278 	.get_module_eeprom = mlx5_get_module_eeprom,
1279 };
1280 
1281 /* Available operations in flow isolated mode. */
1282 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1283 	.dev_configure = mlx5_dev_configure,
1284 	.dev_start = mlx5_dev_start,
1285 	.dev_stop = mlx5_dev_stop,
1286 	.dev_set_link_down = mlx5_set_link_down,
1287 	.dev_set_link_up = mlx5_set_link_up,
1288 	.dev_close = mlx5_dev_close,
1289 	.promiscuous_enable = mlx5_promiscuous_enable,
1290 	.promiscuous_disable = mlx5_promiscuous_disable,
1291 	.allmulticast_enable = mlx5_allmulticast_enable,
1292 	.allmulticast_disable = mlx5_allmulticast_disable,
1293 	.link_update = mlx5_link_update,
1294 	.stats_get = mlx5_stats_get,
1295 	.stats_reset = mlx5_stats_reset,
1296 	.xstats_get = mlx5_xstats_get,
1297 	.xstats_reset = mlx5_xstats_reset,
1298 	.xstats_get_names = mlx5_xstats_get_names,
1299 	.fw_version_get = mlx5_fw_version_get,
1300 	.dev_infos_get = mlx5_dev_infos_get,
1301 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1302 	.vlan_filter_set = mlx5_vlan_filter_set,
1303 	.rx_queue_setup = mlx5_rx_queue_setup,
1304 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1305 	.tx_queue_setup = mlx5_tx_queue_setup,
1306 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1307 	.rx_queue_release = mlx5_rx_queue_release,
1308 	.tx_queue_release = mlx5_tx_queue_release,
1309 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1310 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1311 	.mac_addr_remove = mlx5_mac_addr_remove,
1312 	.mac_addr_add = mlx5_mac_addr_add,
1313 	.mac_addr_set = mlx5_mac_addr_set,
1314 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1315 	.mtu_set = mlx5_dev_set_mtu,
1316 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1317 	.vlan_offload_set = mlx5_vlan_offload_set,
1318 	.filter_ctrl = mlx5_dev_filter_ctrl,
1319 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1320 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1321 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1322 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1323 	.is_removed = mlx5_is_removed,
1324 	.get_module_info = mlx5_get_module_info,
1325 	.get_module_eeprom = mlx5_get_module_eeprom,
1326 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1327 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1328 };
1329 
1330 /**
1331  * Verify and store value for device argument.
1332  *
1333  * @param[in] key
1334  *   Key argument to verify.
1335  * @param[in] val
1336  *   Value associated with key.
1337  * @param opaque
1338  *   User data.
1339  *
1340  * @return
1341  *   0 on success, a negative errno value otherwise and rte_errno is set.
1342  */
1343 static int
1344 mlx5_args_check(const char *key, const char *val, void *opaque)
1345 {
1346 	struct mlx5_dev_config *config = opaque;
1347 	unsigned long tmp;
1348 
1349 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1350 	if (!strcmp(MLX5_REPRESENTOR, key))
1351 		return 0;
1352 	errno = 0;
1353 	tmp = strtoul(val, NULL, 0);
1354 	if (errno) {
1355 		rte_errno = errno;
1356 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1357 		return -rte_errno;
1358 	}
1359 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1360 		config->cqe_comp = !!tmp;
1361 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1362 		config->cqe_pad = !!tmp;
1363 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1364 		config->hw_padding = !!tmp;
1365 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1366 		config->mprq.enabled = !!tmp;
1367 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1368 		config->mprq.stride_num_n = tmp;
1369 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1370 		config->mprq.max_memcpy_len = tmp;
1371 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1372 		config->mprq.min_rxqs_num = tmp;
1373 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1374 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1375 				 " converted to txq_inline_max", key);
1376 		config->txq_inline_max = tmp;
1377 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1378 		config->txq_inline_max = tmp;
1379 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1380 		config->txq_inline_min = tmp;
1381 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1382 		config->txq_inline_mpw = tmp;
1383 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1384 		config->txqs_inline = tmp;
1385 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1386 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1387 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1388 		config->mps = !!tmp;
1389 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1390 		config->dbnc = !!tmp;
1391 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1392 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1393 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1394 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1395 				 " converted to txq_inline_mpw", key);
1396 		config->txq_inline_mpw = tmp;
1397 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1398 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1399 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1400 		config->rx_vec_en = !!tmp;
1401 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1402 		config->l3_vxlan_en = !!tmp;
1403 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1404 		config->vf_nl_en = !!tmp;
1405 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1406 		config->dv_esw_en = !!tmp;
1407 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1408 		config->dv_flow_en = !!tmp;
1409 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1410 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1411 		    tmp != MLX5_XMETA_MODE_META16 &&
1412 		    tmp != MLX5_XMETA_MODE_META32) {
1413 			DRV_LOG(WARNING, "invalid extensive "
1414 					 "metadata parameter");
1415 			rte_errno = EINVAL;
1416 			return -rte_errno;
1417 		}
1418 		config->dv_xmeta_en = tmp;
1419 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1420 		config->mr_ext_memseg_en = !!tmp;
1421 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1422 		config->max_dump_files_num = tmp;
1423 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1424 		config->lro.timeout = tmp;
1425 	} else {
1426 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1427 		rte_errno = EINVAL;
1428 		return -rte_errno;
1429 	}
1430 	return 0;
1431 }
1432 
1433 /**
1434  * Parse device parameters.
1435  *
1436  * @param config
1437  *   Pointer to device configuration structure.
1438  * @param devargs
1439  *   Device arguments structure.
1440  *
1441  * @return
1442  *   0 on success, a negative errno value otherwise and rte_errno is set.
1443  */
1444 static int
1445 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1446 {
1447 	const char **params = (const char *[]){
1448 		MLX5_RXQ_CQE_COMP_EN,
1449 		MLX5_RXQ_CQE_PAD_EN,
1450 		MLX5_RXQ_PKT_PAD_EN,
1451 		MLX5_RX_MPRQ_EN,
1452 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1453 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1454 		MLX5_RXQS_MIN_MPRQ,
1455 		MLX5_TXQ_INLINE,
1456 		MLX5_TXQ_INLINE_MIN,
1457 		MLX5_TXQ_INLINE_MAX,
1458 		MLX5_TXQ_INLINE_MPW,
1459 		MLX5_TXQS_MIN_INLINE,
1460 		MLX5_TXQS_MAX_VEC,
1461 		MLX5_TXQ_MPW_EN,
1462 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1463 		MLX5_TXQ_MAX_INLINE_LEN,
1464 		MLX5_TX_DB_NC,
1465 		MLX5_TX_VEC_EN,
1466 		MLX5_RX_VEC_EN,
1467 		MLX5_L3_VXLAN_EN,
1468 		MLX5_VF_NL_EN,
1469 		MLX5_DV_ESW_EN,
1470 		MLX5_DV_FLOW_EN,
1471 		MLX5_DV_XMETA_EN,
1472 		MLX5_MR_EXT_MEMSEG_EN,
1473 		MLX5_REPRESENTOR,
1474 		MLX5_MAX_DUMP_FILES_NUM,
1475 		MLX5_LRO_TIMEOUT_USEC,
1476 		NULL,
1477 	};
1478 	struct rte_kvargs *kvlist;
1479 	int ret = 0;
1480 	int i;
1481 
1482 	if (devargs == NULL)
1483 		return 0;
1484 	/* Following UGLY cast is done to pass checkpatch. */
1485 	kvlist = rte_kvargs_parse(devargs->args, params);
1486 	if (kvlist == NULL) {
1487 		rte_errno = EINVAL;
1488 		return -rte_errno;
1489 	}
1490 	/* Process parameters. */
1491 	for (i = 0; (params[i] != NULL); ++i) {
1492 		if (rte_kvargs_count(kvlist, params[i])) {
1493 			ret = rte_kvargs_process(kvlist, params[i],
1494 						 mlx5_args_check, config);
1495 			if (ret) {
1496 				rte_errno = EINVAL;
1497 				rte_kvargs_free(kvlist);
1498 				return -rte_errno;
1499 			}
1500 		}
1501 	}
1502 	rte_kvargs_free(kvlist);
1503 	return 0;
1504 }
1505 
1506 static struct rte_pci_driver mlx5_driver;
1507 
1508 /**
1509  * PMD global initialization.
1510  *
1511  * Independent from individual device, this function initializes global
1512  * per-PMD data structures distinguishing primary and secondary processes.
1513  * Hence, each initialization is called once per a process.
1514  *
1515  * @return
1516  *   0 on success, a negative errno value otherwise and rte_errno is set.
1517  */
1518 static int
1519 mlx5_init_once(void)
1520 {
1521 	struct mlx5_shared_data *sd;
1522 	struct mlx5_local_data *ld = &mlx5_local_data;
1523 	int ret = 0;
1524 
1525 	if (mlx5_init_shared_data())
1526 		return -rte_errno;
1527 	sd = mlx5_shared_data;
1528 	assert(sd);
1529 	rte_spinlock_lock(&sd->lock);
1530 	switch (rte_eal_process_type()) {
1531 	case RTE_PROC_PRIMARY:
1532 		if (sd->init_done)
1533 			break;
1534 		LIST_INIT(&sd->mem_event_cb_list);
1535 		rte_rwlock_init(&sd->mem_event_rwlock);
1536 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1537 						mlx5_mr_mem_event_cb, NULL);
1538 		ret = mlx5_mp_init_primary();
1539 		if (ret)
1540 			goto out;
1541 		sd->init_done = true;
1542 		break;
1543 	case RTE_PROC_SECONDARY:
1544 		if (ld->init_done)
1545 			break;
1546 		ret = mlx5_mp_init_secondary();
1547 		if (ret)
1548 			goto out;
1549 		++sd->secondary_cnt;
1550 		ld->init_done = true;
1551 		break;
1552 	default:
1553 		break;
1554 	}
1555 out:
1556 	rte_spinlock_unlock(&sd->lock);
1557 	return ret;
1558 }
1559 
1560 /**
1561  * Configures the minimal amount of data to inline into WQE
1562  * while sending packets.
1563  *
1564  * - the txq_inline_min has the maximal priority, if this
1565  *   key is specified in devargs
1566  * - if DevX is enabled the inline mode is queried from the
1567  *   device (HCA attributes and NIC vport context if needed).
1568  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1569  *   and none (0 bytes) for other NICs
1570  *
1571  * @param spawn
1572  *   Verbs device parameters (name, port, switch_info) to spawn.
1573  * @param config
1574  *   Device configuration parameters.
1575  */
1576 static void
1577 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1578 		    struct mlx5_dev_config *config)
1579 {
1580 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1581 		/* Application defines size of inlined data explicitly. */
1582 		switch (spawn->pci_dev->id.device_id) {
1583 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1584 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1585 			if (config->txq_inline_min <
1586 				       (int)MLX5_INLINE_HSIZE_L2) {
1587 				DRV_LOG(DEBUG,
1588 					"txq_inline_mix aligned to minimal"
1589 					" ConnectX-4 required value %d",
1590 					(int)MLX5_INLINE_HSIZE_L2);
1591 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1592 			}
1593 			break;
1594 		}
1595 		goto exit;
1596 	}
1597 	if (config->hca_attr.eth_net_offloads) {
1598 		/* We have DevX enabled, inline mode queried successfully. */
1599 		switch (config->hca_attr.wqe_inline_mode) {
1600 		case MLX5_CAP_INLINE_MODE_L2:
1601 			/* outer L2 header must be inlined. */
1602 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1603 			goto exit;
1604 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1605 			/* No inline data are required by NIC. */
1606 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1607 			config->hw_vlan_insert =
1608 				config->hca_attr.wqe_vlan_insert;
1609 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1610 			goto exit;
1611 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1612 			/* inline mode is defined by NIC vport context. */
1613 			if (!config->hca_attr.eth_virt)
1614 				break;
1615 			switch (config->hca_attr.vport_inline_mode) {
1616 			case MLX5_INLINE_MODE_NONE:
1617 				config->txq_inline_min =
1618 					MLX5_INLINE_HSIZE_NONE;
1619 				goto exit;
1620 			case MLX5_INLINE_MODE_L2:
1621 				config->txq_inline_min =
1622 					MLX5_INLINE_HSIZE_L2;
1623 				goto exit;
1624 			case MLX5_INLINE_MODE_IP:
1625 				config->txq_inline_min =
1626 					MLX5_INLINE_HSIZE_L3;
1627 				goto exit;
1628 			case MLX5_INLINE_MODE_TCP_UDP:
1629 				config->txq_inline_min =
1630 					MLX5_INLINE_HSIZE_L4;
1631 				goto exit;
1632 			case MLX5_INLINE_MODE_INNER_L2:
1633 				config->txq_inline_min =
1634 					MLX5_INLINE_HSIZE_INNER_L2;
1635 				goto exit;
1636 			case MLX5_INLINE_MODE_INNER_IP:
1637 				config->txq_inline_min =
1638 					MLX5_INLINE_HSIZE_INNER_L3;
1639 				goto exit;
1640 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1641 				config->txq_inline_min =
1642 					MLX5_INLINE_HSIZE_INNER_L4;
1643 				goto exit;
1644 			}
1645 		}
1646 	}
1647 	/*
1648 	 * We get here if we are unable to deduce
1649 	 * inline data size with DevX. Try PCI ID
1650 	 * to determine old NICs.
1651 	 */
1652 	switch (spawn->pci_dev->id.device_id) {
1653 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1654 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1655 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1656 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1657 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1658 		config->hw_vlan_insert = 0;
1659 		break;
1660 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1661 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1662 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1663 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1664 		/*
1665 		 * These NICs support VLAN insertion from WQE and
1666 		 * report the wqe_vlan_insert flag. But there is the bug
1667 		 * and PFC control may be broken, so disable feature.
1668 		 */
1669 		config->hw_vlan_insert = 0;
1670 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1671 		break;
1672 	default:
1673 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1674 		break;
1675 	}
1676 exit:
1677 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1678 }
1679 
1680 /**
1681  * Configures the metadata mask fields in the shared context.
1682  *
1683  * @param [in] dev
1684  *   Pointer to Ethernet device.
1685  */
1686 static void
1687 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1688 {
1689 	struct mlx5_priv *priv = dev->data->dev_private;
1690 	struct mlx5_ibv_shared *sh = priv->sh;
1691 	uint32_t meta, mark, reg_c0;
1692 
1693 	reg_c0 = ~priv->vport_meta_mask;
1694 	switch (priv->config.dv_xmeta_en) {
1695 	case MLX5_XMETA_MODE_LEGACY:
1696 		meta = UINT32_MAX;
1697 		mark = MLX5_FLOW_MARK_MASK;
1698 		break;
1699 	case MLX5_XMETA_MODE_META16:
1700 		meta = reg_c0 >> rte_bsf32(reg_c0);
1701 		mark = MLX5_FLOW_MARK_MASK;
1702 		break;
1703 	case MLX5_XMETA_MODE_META32:
1704 		meta = UINT32_MAX;
1705 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1706 		break;
1707 	default:
1708 		meta = 0;
1709 		mark = 0;
1710 		assert(false);
1711 		break;
1712 	}
1713 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1714 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1715 				 sh->dv_mark_mask, mark);
1716 	else
1717 		sh->dv_mark_mask = mark;
1718 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1719 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1720 				 sh->dv_meta_mask, meta);
1721 	else
1722 		sh->dv_meta_mask = meta;
1723 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1724 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1725 				 sh->dv_meta_mask, reg_c0);
1726 	else
1727 		sh->dv_regc0_mask = reg_c0;
1728 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1729 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1730 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1731 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1732 }
1733 
1734 /**
1735  * Allocate page of door-bells and register it using DevX API.
1736  *
1737  * @param [in] dev
1738  *   Pointer to Ethernet device.
1739  *
1740  * @return
1741  *   Pointer to new page on success, NULL otherwise.
1742  */
1743 static struct mlx5_devx_dbr_page *
1744 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1745 {
1746 	struct mlx5_priv *priv = dev->data->dev_private;
1747 	struct mlx5_devx_dbr_page *page;
1748 
1749 	/* Allocate space for door-bell page and management data. */
1750 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1751 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1752 	if (!page) {
1753 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1754 			dev->data->port_id);
1755 		return NULL;
1756 	}
1757 	/* Register allocated memory. */
1758 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1759 					      MLX5_DBR_PAGE_SIZE, 0);
1760 	if (!page->umem) {
1761 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1762 			dev->data->port_id);
1763 		rte_free(page);
1764 		return NULL;
1765 	}
1766 	return page;
1767 }
1768 
1769 /**
1770  * Find the next available door-bell, allocate new page if needed.
1771  *
1772  * @param [in] dev
1773  *   Pointer to Ethernet device.
1774  * @param [out] dbr_page
1775  *   Door-bell page containing the page data.
1776  *
1777  * @return
1778  *   Door-bell address offset on success, a negative error value otherwise.
1779  */
1780 int64_t
1781 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1782 {
1783 	struct mlx5_priv *priv = dev->data->dev_private;
1784 	struct mlx5_devx_dbr_page *page = NULL;
1785 	uint32_t i, j;
1786 
1787 	LIST_FOREACH(page, &priv->dbrpgs, next)
1788 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1789 			break;
1790 	if (!page) { /* No page with free door-bell exists. */
1791 		page = mlx5_alloc_dbr_page(dev);
1792 		if (!page) /* Failed to allocate new page. */
1793 			return (-1);
1794 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1795 	}
1796 	/* Loop to find bitmap part with clear bit. */
1797 	for (i = 0;
1798 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1799 	     i++)
1800 		; /* Empty. */
1801 	/* Find the first clear bit. */
1802 	j = rte_bsf64(~page->dbr_bitmap[i]);
1803 	assert(i < (MLX5_DBR_PER_PAGE / 64));
1804 	page->dbr_bitmap[i] |= (1 << j);
1805 	page->dbr_count++;
1806 	*dbr_page = page;
1807 	return (((i * 64) + j) * sizeof(uint64_t));
1808 }
1809 
1810 /**
1811  * Release a door-bell record.
1812  *
1813  * @param [in] dev
1814  *   Pointer to Ethernet device.
1815  * @param [in] umem_id
1816  *   UMEM ID of page containing the door-bell record to release.
1817  * @param [in] offset
1818  *   Offset of door-bell record in page.
1819  *
1820  * @return
1821  *   0 on success, a negative error value otherwise.
1822  */
1823 int32_t
1824 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1825 {
1826 	struct mlx5_priv *priv = dev->data->dev_private;
1827 	struct mlx5_devx_dbr_page *page = NULL;
1828 	int ret = 0;
1829 
1830 	LIST_FOREACH(page, &priv->dbrpgs, next)
1831 		/* Find the page this address belongs to. */
1832 		if (page->umem->umem_id == umem_id)
1833 			break;
1834 	if (!page)
1835 		return -EINVAL;
1836 	page->dbr_count--;
1837 	if (!page->dbr_count) {
1838 		/* Page not used, free it and remove from list. */
1839 		LIST_REMOVE(page, next);
1840 		if (page->umem)
1841 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
1842 		rte_free(page);
1843 	} else {
1844 		/* Mark in bitmap that this door-bell is not in use. */
1845 		offset /= MLX5_DBR_SIZE;
1846 		int i = offset / 64;
1847 		int j = offset % 64;
1848 
1849 		page->dbr_bitmap[i] &= ~(1 << j);
1850 	}
1851 	return ret;
1852 }
1853 
1854 /**
1855  * Check sibling device configurations.
1856  *
1857  * Sibling devices sharing the Infiniband device context
1858  * should have compatible configurations. This regards
1859  * representors and bonding slaves.
1860  *
1861  * @param priv
1862  *   Private device descriptor.
1863  * @param config
1864  *   Configuration of the device is going to be created.
1865  *
1866  * @return
1867  *   0 on success, EINVAL otherwise
1868  */
1869 static int
1870 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1871 			      struct mlx5_dev_config *config)
1872 {
1873 	struct mlx5_ibv_shared *sh = priv->sh;
1874 	struct mlx5_dev_config *sh_conf = NULL;
1875 	uint16_t port_id;
1876 
1877 	assert(sh);
1878 	/* Nothing to compare for the single/first device. */
1879 	if (sh->refcnt == 1)
1880 		return 0;
1881 	/* Find the device with shared context. */
1882 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1883 		struct mlx5_priv *opriv =
1884 			rte_eth_devices[port_id].data->dev_private;
1885 
1886 		if (opriv && opriv != priv && opriv->sh == sh) {
1887 			sh_conf = &opriv->config;
1888 			break;
1889 		}
1890 	}
1891 	if (!sh_conf)
1892 		return 0;
1893 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1894 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1895 			     " for shared %s context", sh->ibdev_name);
1896 		rte_errno = EINVAL;
1897 		return rte_errno;
1898 	}
1899 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1900 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1901 			     " for shared %s context", sh->ibdev_name);
1902 		rte_errno = EINVAL;
1903 		return rte_errno;
1904 	}
1905 	return 0;
1906 }
1907 /**
1908  * Spawn an Ethernet device from Verbs information.
1909  *
1910  * @param dpdk_dev
1911  *   Backing DPDK device.
1912  * @param spawn
1913  *   Verbs device parameters (name, port, switch_info) to spawn.
1914  * @param config
1915  *   Device configuration parameters.
1916  *
1917  * @return
1918  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1919  *   is set. The following errors are defined:
1920  *
1921  *   EBUSY: device is not supposed to be spawned.
1922  *   EEXIST: device is already spawned
1923  */
1924 static struct rte_eth_dev *
1925 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1926 	       struct mlx5_dev_spawn_data *spawn,
1927 	       struct mlx5_dev_config config)
1928 {
1929 	const struct mlx5_switch_info *switch_info = &spawn->info;
1930 	struct mlx5_ibv_shared *sh = NULL;
1931 	struct ibv_port_attr port_attr;
1932 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1933 	struct rte_eth_dev *eth_dev = NULL;
1934 	struct mlx5_priv *priv = NULL;
1935 	int err = 0;
1936 	unsigned int hw_padding = 0;
1937 	unsigned int mps;
1938 	unsigned int cqe_comp;
1939 	unsigned int cqe_pad = 0;
1940 	unsigned int tunnel_en = 0;
1941 	unsigned int mpls_en = 0;
1942 	unsigned int swp = 0;
1943 	unsigned int mprq = 0;
1944 	unsigned int mprq_min_stride_size_n = 0;
1945 	unsigned int mprq_max_stride_size_n = 0;
1946 	unsigned int mprq_min_stride_num_n = 0;
1947 	unsigned int mprq_max_stride_num_n = 0;
1948 	struct rte_ether_addr mac;
1949 	char name[RTE_ETH_NAME_MAX_LEN];
1950 	int own_domain_id = 0;
1951 	uint16_t port_id;
1952 	unsigned int i;
1953 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1954 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
1955 #endif
1956 
1957 	/* Determine if this port representor is supposed to be spawned. */
1958 	if (switch_info->representor && dpdk_dev->devargs) {
1959 		struct rte_eth_devargs eth_da;
1960 
1961 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1962 		if (err) {
1963 			rte_errno = -err;
1964 			DRV_LOG(ERR, "failed to process device arguments: %s",
1965 				strerror(rte_errno));
1966 			return NULL;
1967 		}
1968 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
1969 			if (eth_da.representor_ports[i] ==
1970 			    (uint16_t)switch_info->port_name)
1971 				break;
1972 		if (i == eth_da.nb_representor_ports) {
1973 			rte_errno = EBUSY;
1974 			return NULL;
1975 		}
1976 	}
1977 	/* Build device name. */
1978 	if (spawn->pf_bond <  0) {
1979 		/* Single device. */
1980 		if (!switch_info->representor)
1981 			strlcpy(name, dpdk_dev->name, sizeof(name));
1982 		else
1983 			snprintf(name, sizeof(name), "%s_representor_%u",
1984 				 dpdk_dev->name, switch_info->port_name);
1985 	} else {
1986 		/* Bonding device. */
1987 		if (!switch_info->representor)
1988 			snprintf(name, sizeof(name), "%s_%s",
1989 				 dpdk_dev->name, spawn->ibv_dev->name);
1990 		else
1991 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
1992 				 dpdk_dev->name, spawn->ibv_dev->name,
1993 				 switch_info->port_name);
1994 	}
1995 	/* check if the device is already spawned */
1996 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1997 		rte_errno = EEXIST;
1998 		return NULL;
1999 	}
2000 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2001 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2002 		eth_dev = rte_eth_dev_attach_secondary(name);
2003 		if (eth_dev == NULL) {
2004 			DRV_LOG(ERR, "can not attach rte ethdev");
2005 			rte_errno = ENOMEM;
2006 			return NULL;
2007 		}
2008 		eth_dev->device = dpdk_dev;
2009 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2010 		err = mlx5_proc_priv_init(eth_dev);
2011 		if (err)
2012 			return NULL;
2013 		/* Receive command fd from primary process */
2014 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2015 		if (err < 0)
2016 			return NULL;
2017 		/* Remap UAR for Tx queues. */
2018 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2019 		if (err)
2020 			return NULL;
2021 		/*
2022 		 * Ethdev pointer is still required as input since
2023 		 * the primary device is not accessible from the
2024 		 * secondary process.
2025 		 */
2026 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2027 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2028 		return eth_dev;
2029 	}
2030 	/*
2031 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2032 	 * advance to create dv/verbs device context. We proceed the
2033 	 * devargs here to get ones, and later proceed devargs again
2034 	 * to override some hardware settings.
2035 	 */
2036 	err = mlx5_args(&config, dpdk_dev->devargs);
2037 	if (err) {
2038 		err = rte_errno;
2039 		DRV_LOG(ERR, "failed to process device arguments: %s",
2040 			strerror(rte_errno));
2041 		goto error;
2042 	}
2043 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2044 	if (!sh)
2045 		return NULL;
2046 	config.devx = sh->devx;
2047 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2048 	config.dest_tir = 1;
2049 #endif
2050 #ifdef HAVE_IBV_MLX5_MOD_SWP
2051 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2052 #endif
2053 	/*
2054 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2055 	 * as all ConnectX-5 devices.
2056 	 */
2057 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2058 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2059 #endif
2060 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2061 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2062 #endif
2063 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2064 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2065 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2066 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2067 			mps = MLX5_MPW_ENHANCED;
2068 		} else {
2069 			DRV_LOG(DEBUG, "MPW is supported");
2070 			mps = MLX5_MPW;
2071 		}
2072 	} else {
2073 		DRV_LOG(DEBUG, "MPW isn't supported");
2074 		mps = MLX5_MPW_DISABLED;
2075 	}
2076 #ifdef HAVE_IBV_MLX5_MOD_SWP
2077 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2078 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2079 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2080 #endif
2081 	config.swp = !!swp;
2082 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2083 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2084 		struct mlx5dv_striding_rq_caps mprq_caps =
2085 			dv_attr.striding_rq_caps;
2086 
2087 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2088 			mprq_caps.min_single_stride_log_num_of_bytes);
2089 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2090 			mprq_caps.max_single_stride_log_num_of_bytes);
2091 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2092 			mprq_caps.min_single_wqe_log_num_of_strides);
2093 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2094 			mprq_caps.max_single_wqe_log_num_of_strides);
2095 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2096 			mprq_caps.supported_qpts);
2097 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2098 		mprq = 1;
2099 		mprq_min_stride_size_n =
2100 			mprq_caps.min_single_stride_log_num_of_bytes;
2101 		mprq_max_stride_size_n =
2102 			mprq_caps.max_single_stride_log_num_of_bytes;
2103 		mprq_min_stride_num_n =
2104 			mprq_caps.min_single_wqe_log_num_of_strides;
2105 		mprq_max_stride_num_n =
2106 			mprq_caps.max_single_wqe_log_num_of_strides;
2107 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2108 						   mprq_min_stride_num_n);
2109 	}
2110 #endif
2111 	if (RTE_CACHE_LINE_SIZE == 128 &&
2112 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2113 		cqe_comp = 0;
2114 	else
2115 		cqe_comp = 1;
2116 	config.cqe_comp = cqe_comp;
2117 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2118 	/* Whether device supports 128B Rx CQE padding. */
2119 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2120 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2121 #endif
2122 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2123 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2124 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2125 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2126 			     (dv_attr.tunnel_offloads_caps &
2127 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
2128 	}
2129 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2130 		tunnel_en ? "" : "not ");
2131 #else
2132 	DRV_LOG(WARNING,
2133 		"tunnel offloading disabled due to old OFED/rdma-core version");
2134 #endif
2135 	config.tunnel_en = tunnel_en;
2136 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2137 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2138 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2139 		   (dv_attr.tunnel_offloads_caps &
2140 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2141 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2142 		mpls_en ? "" : "not ");
2143 #else
2144 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2145 		" old OFED/rdma-core version or firmware configuration");
2146 #endif
2147 	config.mpls_en = mpls_en;
2148 	/* Check port status. */
2149 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2150 	if (err) {
2151 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2152 		goto error;
2153 	}
2154 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2155 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2156 		err = EINVAL;
2157 		goto error;
2158 	}
2159 	if (port_attr.state != IBV_PORT_ACTIVE)
2160 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2161 			mlx5_glue->port_state_str(port_attr.state),
2162 			port_attr.state);
2163 	/* Allocate private eth device data. */
2164 	priv = rte_zmalloc("ethdev private structure",
2165 			   sizeof(*priv),
2166 			   RTE_CACHE_LINE_SIZE);
2167 	if (priv == NULL) {
2168 		DRV_LOG(ERR, "priv allocation failure");
2169 		err = ENOMEM;
2170 		goto error;
2171 	}
2172 	priv->sh = sh;
2173 	priv->ibv_port = spawn->ibv_port;
2174 	priv->pci_dev = spawn->pci_dev;
2175 	priv->mtu = RTE_ETHER_MTU;
2176 #ifndef RTE_ARCH_64
2177 	/* Initialize UAR access locks for 32bit implementations. */
2178 	rte_spinlock_init(&priv->uar_lock_cq);
2179 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2180 		rte_spinlock_init(&priv->uar_lock[i]);
2181 #endif
2182 	/* Some internal functions rely on Netlink sockets, open them now. */
2183 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2184 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2185 	priv->nl_sn = 0;
2186 	priv->representor = !!switch_info->representor;
2187 	priv->master = !!switch_info->master;
2188 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2189 	priv->vport_meta_tag = 0;
2190 	priv->vport_meta_mask = 0;
2191 	priv->pf_bond = spawn->pf_bond;
2192 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2193 	/*
2194 	 * The DevX port query API is implemented. E-Switch may use
2195 	 * either vport or reg_c[0] metadata register to match on
2196 	 * vport index. The engaged part of metadata register is
2197 	 * defined by mask.
2198 	 */
2199 	if (switch_info->representor || switch_info->master) {
2200 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2201 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2202 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2203 						 &devx_port);
2204 		if (err) {
2205 			DRV_LOG(WARNING,
2206 				"can't query devx port %d on device %s",
2207 				spawn->ibv_port, spawn->ibv_dev->name);
2208 			devx_port.comp_mask = 0;
2209 		}
2210 	}
2211 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2212 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2213 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2214 		if (!priv->vport_meta_mask) {
2215 			DRV_LOG(ERR, "vport zero mask for port %d"
2216 				     " on bonding device %s",
2217 				     spawn->ibv_port, spawn->ibv_dev->name);
2218 			err = ENOTSUP;
2219 			goto error;
2220 		}
2221 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2222 			DRV_LOG(ERR, "invalid vport tag for port %d"
2223 				     " on bonding device %s",
2224 				     spawn->ibv_port, spawn->ibv_dev->name);
2225 			err = ENOTSUP;
2226 			goto error;
2227 		}
2228 	} else if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2229 		priv->vport_id = devx_port.vport_num;
2230 	} else if (spawn->pf_bond >= 0) {
2231 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2232 			     " on bonding device %s",
2233 			     spawn->ibv_port, spawn->ibv_dev->name);
2234 		err = ENOTSUP;
2235 		goto error;
2236 	} else {
2237 		/* Suppose vport index in compatible way. */
2238 		priv->vport_id = switch_info->representor ?
2239 				 switch_info->port_name + 1 : -1;
2240 	}
2241 #else
2242 	/*
2243 	 * Kernel/rdma_core support single E-Switch per PF configurations
2244 	 * only and vport_id field contains the vport index for
2245 	 * associated VF, which is deduced from representor port name.
2246 	 * For example, let's have the IB device port 10, it has
2247 	 * attached network device eth0, which has port name attribute
2248 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2249 	 * as 3 (2+1). This assigning schema should be changed if the
2250 	 * multiple E-Switch instances per PF configurations or/and PCI
2251 	 * subfunctions are added.
2252 	 */
2253 	priv->vport_id = switch_info->representor ?
2254 			 switch_info->port_name + 1 : -1;
2255 #endif
2256 	/* representor_id field keeps the unmodified VF index. */
2257 	priv->representor_id = switch_info->representor ?
2258 			       switch_info->port_name : -1;
2259 	/*
2260 	 * Look for sibling devices in order to reuse their switch domain
2261 	 * if any, otherwise allocate one.
2262 	 */
2263 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2264 		const struct mlx5_priv *opriv =
2265 			rte_eth_devices[port_id].data->dev_private;
2266 
2267 		if (!opriv ||
2268 		    opriv->sh != priv->sh ||
2269 			opriv->domain_id ==
2270 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2271 			continue;
2272 		priv->domain_id = opriv->domain_id;
2273 		break;
2274 	}
2275 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2276 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2277 		if (err) {
2278 			err = rte_errno;
2279 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2280 				strerror(rte_errno));
2281 			goto error;
2282 		}
2283 		own_domain_id = 1;
2284 	}
2285 	/* Override some values set by hardware configuration. */
2286 	mlx5_args(&config, dpdk_dev->devargs);
2287 	err = mlx5_dev_check_sibling_config(priv, &config);
2288 	if (err)
2289 		goto error;
2290 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2291 			    IBV_DEVICE_RAW_IP_CSUM);
2292 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2293 		(config.hw_csum ? "" : "not "));
2294 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2295 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2296 	DRV_LOG(DEBUG, "counters are not supported");
2297 #endif
2298 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
2299 	if (config.dv_flow_en) {
2300 		DRV_LOG(WARNING, "DV flow is not supported");
2301 		config.dv_flow_en = 0;
2302 	}
2303 #endif
2304 	config.ind_table_max_size =
2305 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2306 	/*
2307 	 * Remove this check once DPDK supports larger/variable
2308 	 * indirection tables.
2309 	 */
2310 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2311 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2312 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2313 		config.ind_table_max_size);
2314 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2315 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2316 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2317 		(config.hw_vlan_strip ? "" : "not "));
2318 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2319 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2320 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2321 		(config.hw_fcs_strip ? "" : "not "));
2322 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2323 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2324 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2325 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2326 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2327 #endif
2328 	if (config.hw_padding && !hw_padding) {
2329 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2330 		config.hw_padding = 0;
2331 	} else if (config.hw_padding) {
2332 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2333 	}
2334 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2335 		      (sh->device_attr.tso_caps.supported_qpts &
2336 		       (1 << IBV_QPT_RAW_PACKET)));
2337 	if (config.tso)
2338 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2339 	/*
2340 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2341 	 * by default.
2342 	 */
2343 	if (config.mps == MLX5_ARG_UNSET)
2344 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2345 							  MLX5_MPW_DISABLED;
2346 	else
2347 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2348 	DRV_LOG(INFO, "%sMPS is %s",
2349 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
2350 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2351 	if (config.cqe_comp && !cqe_comp) {
2352 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2353 		config.cqe_comp = 0;
2354 	}
2355 	if (config.cqe_pad && !cqe_pad) {
2356 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2357 		config.cqe_pad = 0;
2358 	} else if (config.cqe_pad) {
2359 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2360 	}
2361 	if (config.devx) {
2362 		priv->counter_fallback = 0;
2363 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2364 		if (err) {
2365 			err = -err;
2366 			goto error;
2367 		}
2368 		if (!config.hca_attr.flow_counters_dump)
2369 			priv->counter_fallback = 1;
2370 #ifndef HAVE_IBV_DEVX_ASYNC
2371 		priv->counter_fallback = 1;
2372 #endif
2373 		if (priv->counter_fallback)
2374 			DRV_LOG(INFO, "Use fall-back DV counter management");
2375 		/* Check for LRO support. */
2376 		if (config.dest_tir && config.hca_attr.lro_cap &&
2377 		    config.dv_flow_en) {
2378 			/* TBD check tunnel lro caps. */
2379 			config.lro.supported = config.hca_attr.lro_cap;
2380 			DRV_LOG(DEBUG, "Device supports LRO");
2381 			/*
2382 			 * If LRO timeout is not configured by application,
2383 			 * use the minimal supported value.
2384 			 */
2385 			if (!config.lro.timeout)
2386 				config.lro.timeout =
2387 				config.hca_attr.lro_timer_supported_periods[0];
2388 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2389 				config.lro.timeout);
2390 		}
2391 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2392 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2393 		    config.dv_flow_en) {
2394 			uint8_t reg_c_mask =
2395 				config.hca_attr.qos.flow_meter_reg_c_ids;
2396 			/*
2397 			 * Meter needs two REG_C's for color match and pre-sfx
2398 			 * flow match. Here get the REG_C for color match.
2399 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2400 			 */
2401 			reg_c_mask &= 0xfc;
2402 			if (__builtin_popcount(reg_c_mask) < 1) {
2403 				priv->mtr_en = 0;
2404 				DRV_LOG(WARNING, "No available register for"
2405 					" meter.");
2406 			} else {
2407 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2408 						      REG_C_0;
2409 				priv->mtr_en = 1;
2410 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2411 					priv->mtr_color_reg);
2412 			}
2413 		}
2414 #endif
2415 	}
2416 	if (config.mprq.enabled && mprq) {
2417 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2418 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2419 			config.mprq.stride_num_n =
2420 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2421 					mprq_min_stride_num_n);
2422 			DRV_LOG(WARNING,
2423 				"the number of strides"
2424 				" for Multi-Packet RQ is out of range,"
2425 				" setting default value (%u)",
2426 				1 << config.mprq.stride_num_n);
2427 		}
2428 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2429 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2430 	} else if (config.mprq.enabled && !mprq) {
2431 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2432 		config.mprq.enabled = 0;
2433 	}
2434 	if (config.max_dump_files_num == 0)
2435 		config.max_dump_files_num = 128;
2436 	eth_dev = rte_eth_dev_allocate(name);
2437 	if (eth_dev == NULL) {
2438 		DRV_LOG(ERR, "can not allocate rte ethdev");
2439 		err = ENOMEM;
2440 		goto error;
2441 	}
2442 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2443 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2444 	if (priv->representor) {
2445 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2446 		eth_dev->data->representor_id = priv->representor_id;
2447 	}
2448 	/*
2449 	 * Store associated network device interface index. This index
2450 	 * is permanent throughout the lifetime of device. So, we may store
2451 	 * the ifindex here and use the cached value further.
2452 	 */
2453 	assert(spawn->ifindex);
2454 	priv->if_index = spawn->ifindex;
2455 	eth_dev->data->dev_private = priv;
2456 	priv->dev_data = eth_dev->data;
2457 	eth_dev->data->mac_addrs = priv->mac;
2458 	eth_dev->device = dpdk_dev;
2459 	/* Configure the first MAC address by default. */
2460 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2461 		DRV_LOG(ERR,
2462 			"port %u cannot get MAC address, is mlx5_en"
2463 			" loaded? (errno: %s)",
2464 			eth_dev->data->port_id, strerror(rte_errno));
2465 		err = ENODEV;
2466 		goto error;
2467 	}
2468 	DRV_LOG(INFO,
2469 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2470 		eth_dev->data->port_id,
2471 		mac.addr_bytes[0], mac.addr_bytes[1],
2472 		mac.addr_bytes[2], mac.addr_bytes[3],
2473 		mac.addr_bytes[4], mac.addr_bytes[5]);
2474 #ifndef NDEBUG
2475 	{
2476 		char ifname[IF_NAMESIZE];
2477 
2478 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2479 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2480 				eth_dev->data->port_id, ifname);
2481 		else
2482 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2483 				eth_dev->data->port_id);
2484 	}
2485 #endif
2486 	/* Get actual MTU if possible. */
2487 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2488 	if (err) {
2489 		err = rte_errno;
2490 		goto error;
2491 	}
2492 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2493 		priv->mtu);
2494 	/* Initialize burst functions to prevent crashes before link-up. */
2495 	eth_dev->rx_pkt_burst = removed_rx_burst;
2496 	eth_dev->tx_pkt_burst = removed_tx_burst;
2497 	eth_dev->dev_ops = &mlx5_dev_ops;
2498 	/* Register MAC address. */
2499 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2500 	if (config.vf && config.vf_nl_en)
2501 		mlx5_nl_mac_addr_sync(eth_dev);
2502 	TAILQ_INIT(&priv->flows);
2503 	TAILQ_INIT(&priv->ctrl_flows);
2504 	TAILQ_INIT(&priv->flow_meters);
2505 	TAILQ_INIT(&priv->flow_meter_profiles);
2506 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2507 	struct mlx5dv_ctx_allocators alctr = {
2508 		.alloc = &mlx5_alloc_verbs_buf,
2509 		.free = &mlx5_free_verbs_buf,
2510 		.data = priv,
2511 	};
2512 	mlx5_glue->dv_set_context_attr(sh->ctx,
2513 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2514 				       (void *)((uintptr_t)&alctr));
2515 	/* Bring Ethernet device up. */
2516 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2517 		eth_dev->data->port_id);
2518 	mlx5_set_link_up(eth_dev);
2519 	/*
2520 	 * Even though the interrupt handler is not installed yet,
2521 	 * interrupts will still trigger on the async_fd from
2522 	 * Verbs context returned by ibv_open_device().
2523 	 */
2524 	mlx5_link_update(eth_dev, 0);
2525 #ifdef HAVE_MLX5DV_DR_ESWITCH
2526 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2527 	      (switch_info->representor || switch_info->master)))
2528 		config.dv_esw_en = 0;
2529 #else
2530 	config.dv_esw_en = 0;
2531 #endif
2532 	/* Detect minimal data bytes to inline. */
2533 	mlx5_set_min_inline(spawn, &config);
2534 	/* Store device configuration on private structure. */
2535 	priv->config = config;
2536 	/* Create context for virtual machine VLAN workaround. */
2537 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2538 	if (config.dv_flow_en) {
2539 		err = mlx5_alloc_shared_dr(priv);
2540 		if (err)
2541 			goto error;
2542 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc();
2543 		if (!priv->qrss_id_pool) {
2544 			DRV_LOG(ERR, "can't create flow id pool");
2545 			err = ENOMEM;
2546 			goto error;
2547 		}
2548 	}
2549 	/* Supported Verbs flow priority number detection. */
2550 	err = mlx5_flow_discover_priorities(eth_dev);
2551 	if (err < 0) {
2552 		err = -err;
2553 		goto error;
2554 	}
2555 	priv->config.flow_prio = err;
2556 	if (!priv->config.dv_esw_en &&
2557 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2558 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2559 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2560 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2561 	}
2562 	mlx5_set_metadata_mask(eth_dev);
2563 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2564 	    !priv->sh->dv_regc0_mask) {
2565 		DRV_LOG(ERR, "metadata mode %u is not supported "
2566 			     "(no metadata reg_c[0] is available)",
2567 			     priv->config.dv_xmeta_en);
2568 			err = ENOTSUP;
2569 			goto error;
2570 	}
2571 	/* Query availibility of metadata reg_c's. */
2572 	err = mlx5_flow_discover_mreg_c(eth_dev);
2573 	if (err < 0) {
2574 		err = -err;
2575 		goto error;
2576 	}
2577 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2578 		DRV_LOG(DEBUG,
2579 			"port %u extensive metadata register is not supported",
2580 			eth_dev->data->port_id);
2581 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2582 			DRV_LOG(ERR, "metadata mode %u is not supported "
2583 				     "(no metadata registers available)",
2584 				     priv->config.dv_xmeta_en);
2585 			err = ENOTSUP;
2586 			goto error;
2587 		}
2588 	}
2589 	if (priv->config.dv_flow_en &&
2590 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2591 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
2592 	    priv->sh->dv_regc0_mask) {
2593 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2594 						      MLX5_FLOW_MREG_HTABLE_SZ);
2595 		if (!priv->mreg_cp_tbl) {
2596 			err = ENOMEM;
2597 			goto error;
2598 		}
2599 	}
2600 	return eth_dev;
2601 error:
2602 	if (priv) {
2603 		if (priv->mreg_cp_tbl)
2604 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2605 		if (priv->sh)
2606 			mlx5_free_shared_dr(priv);
2607 		if (priv->nl_socket_route >= 0)
2608 			close(priv->nl_socket_route);
2609 		if (priv->nl_socket_rdma >= 0)
2610 			close(priv->nl_socket_rdma);
2611 		if (priv->vmwa_context)
2612 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2613 		if (priv->qrss_id_pool)
2614 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
2615 		if (own_domain_id)
2616 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2617 		rte_free(priv);
2618 		if (eth_dev != NULL)
2619 			eth_dev->data->dev_private = NULL;
2620 	}
2621 	if (eth_dev != NULL) {
2622 		/* mac_addrs must not be freed alone because part of dev_private */
2623 		eth_dev->data->mac_addrs = NULL;
2624 		rte_eth_dev_release_port(eth_dev);
2625 	}
2626 	if (sh)
2627 		mlx5_free_shared_ibctx(sh);
2628 	assert(err > 0);
2629 	rte_errno = err;
2630 	return NULL;
2631 }
2632 
2633 /**
2634  * Comparison callback to sort device data.
2635  *
2636  * This is meant to be used with qsort().
2637  *
2638  * @param a[in]
2639  *   Pointer to pointer to first data object.
2640  * @param b[in]
2641  *   Pointer to pointer to second data object.
2642  *
2643  * @return
2644  *   0 if both objects are equal, less than 0 if the first argument is less
2645  *   than the second, greater than 0 otherwise.
2646  */
2647 static int
2648 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2649 {
2650 	const struct mlx5_switch_info *si_a =
2651 		&((const struct mlx5_dev_spawn_data *)a)->info;
2652 	const struct mlx5_switch_info *si_b =
2653 		&((const struct mlx5_dev_spawn_data *)b)->info;
2654 	int ret;
2655 
2656 	/* Master device first. */
2657 	ret = si_b->master - si_a->master;
2658 	if (ret)
2659 		return ret;
2660 	/* Then representor devices. */
2661 	ret = si_b->representor - si_a->representor;
2662 	if (ret)
2663 		return ret;
2664 	/* Unidentified devices come last in no specific order. */
2665 	if (!si_a->representor)
2666 		return 0;
2667 	/* Order representors by name. */
2668 	return si_a->port_name - si_b->port_name;
2669 }
2670 
2671 /**
2672  * Match PCI information for possible slaves of bonding device.
2673  *
2674  * @param[in] ibv_dev
2675  *   Pointer to Infiniband device structure.
2676  * @param[in] pci_dev
2677  *   Pointer to PCI device structure to match PCI address.
2678  * @param[in] nl_rdma
2679  *   Netlink RDMA group socket handle.
2680  *
2681  * @return
2682  *   negative value if no bonding device found, otherwise
2683  *   positive index of slave PF in bonding.
2684  */
2685 static int
2686 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2687 			   const struct rte_pci_device *pci_dev,
2688 			   int nl_rdma)
2689 {
2690 	char ifname[IF_NAMESIZE + 1];
2691 	unsigned int ifindex;
2692 	unsigned int np, i;
2693 	FILE *file = NULL;
2694 	int pf = -1;
2695 
2696 	/*
2697 	 * Try to get master device name. If something goes
2698 	 * wrong suppose the lack of kernel support and no
2699 	 * bonding devices.
2700 	 */
2701 	if (nl_rdma < 0)
2702 		return -1;
2703 	if (!strstr(ibv_dev->name, "bond"))
2704 		return -1;
2705 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2706 	if (!np)
2707 		return -1;
2708 	/*
2709 	 * The Master device might not be on the predefined
2710 	 * port (not on port index 1, it is not garanted),
2711 	 * we have to scan all Infiniband device port and
2712 	 * find master.
2713 	 */
2714 	for (i = 1; i <= np; ++i) {
2715 		/* Check whether Infiniband port is populated. */
2716 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2717 		if (!ifindex)
2718 			continue;
2719 		if (!if_indextoname(ifindex, ifname))
2720 			continue;
2721 		/* Try to read bonding slave names from sysfs. */
2722 		MKSTR(slaves,
2723 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2724 		file = fopen(slaves, "r");
2725 		if (file)
2726 			break;
2727 	}
2728 	if (!file)
2729 		return -1;
2730 	/* Use safe format to check maximal buffer length. */
2731 	assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2732 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2733 		char tmp_str[IF_NAMESIZE + 32];
2734 		struct rte_pci_addr pci_addr;
2735 		struct mlx5_switch_info	info;
2736 
2737 		/* Process slave interface names in the loop. */
2738 		snprintf(tmp_str, sizeof(tmp_str),
2739 			 "/sys/class/net/%s", ifname);
2740 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2741 			DRV_LOG(WARNING, "can not get PCI address"
2742 					 " for netdev \"%s\"", ifname);
2743 			continue;
2744 		}
2745 		if (pci_dev->addr.domain != pci_addr.domain ||
2746 		    pci_dev->addr.bus != pci_addr.bus ||
2747 		    pci_dev->addr.devid != pci_addr.devid ||
2748 		    pci_dev->addr.function != pci_addr.function)
2749 			continue;
2750 		/* Slave interface PCI address match found. */
2751 		fclose(file);
2752 		snprintf(tmp_str, sizeof(tmp_str),
2753 			 "/sys/class/net/%s/phys_port_name", ifname);
2754 		file = fopen(tmp_str, "rb");
2755 		if (!file)
2756 			break;
2757 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2758 		if (fscanf(file, "%32s", tmp_str) == 1)
2759 			mlx5_translate_port_name(tmp_str, &info);
2760 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2761 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2762 			pf = info.port_name;
2763 		break;
2764 	}
2765 	if (file)
2766 		fclose(file);
2767 	return pf;
2768 }
2769 
2770 /**
2771  * DPDK callback to register a PCI device.
2772  *
2773  * This function spawns Ethernet devices out of a given PCI device.
2774  *
2775  * @param[in] pci_drv
2776  *   PCI driver structure (mlx5_driver).
2777  * @param[in] pci_dev
2778  *   PCI device information.
2779  *
2780  * @return
2781  *   0 on success, a negative errno value otherwise and rte_errno is set.
2782  */
2783 static int
2784 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2785 	       struct rte_pci_device *pci_dev)
2786 {
2787 	struct ibv_device **ibv_list;
2788 	/*
2789 	 * Number of found IB Devices matching with requested PCI BDF.
2790 	 * nd != 1 means there are multiple IB devices over the same
2791 	 * PCI device and we have representors and master.
2792 	 */
2793 	unsigned int nd = 0;
2794 	/*
2795 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2796 	 * we have the single multiport IB device, and there may be
2797 	 * representors attached to some of found ports.
2798 	 */
2799 	unsigned int np = 0;
2800 	/*
2801 	 * Number of DPDK ethernet devices to Spawn - either over
2802 	 * multiple IB devices or multiple ports of single IB device.
2803 	 * Actually this is the number of iterations to spawn.
2804 	 */
2805 	unsigned int ns = 0;
2806 	/*
2807 	 * Bonding device
2808 	 *   < 0 - no bonding device (single one)
2809 	 *  >= 0 - bonding device (value is slave PF index)
2810 	 */
2811 	int bd = -1;
2812 	struct mlx5_dev_spawn_data *list = NULL;
2813 	struct mlx5_dev_config dev_config;
2814 	int ret;
2815 
2816 	ret = mlx5_init_once();
2817 	if (ret) {
2818 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2819 			strerror(rte_errno));
2820 		return -rte_errno;
2821 	}
2822 	assert(pci_drv == &mlx5_driver);
2823 	errno = 0;
2824 	ibv_list = mlx5_glue->get_device_list(&ret);
2825 	if (!ibv_list) {
2826 		rte_errno = errno ? errno : ENOSYS;
2827 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2828 		return -rte_errno;
2829 	}
2830 	/*
2831 	 * First scan the list of all Infiniband devices to find
2832 	 * matching ones, gathering into the list.
2833 	 */
2834 	struct ibv_device *ibv_match[ret + 1];
2835 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2836 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2837 	unsigned int i;
2838 
2839 	while (ret-- > 0) {
2840 		struct rte_pci_addr pci_addr;
2841 
2842 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2843 		bd = mlx5_device_bond_pci_match
2844 				(ibv_list[ret], pci_dev, nl_rdma);
2845 		if (bd >= 0) {
2846 			/*
2847 			 * Bonding device detected. Only one match is allowed,
2848 			 * the bonding is supported over multi-port IB device,
2849 			 * there should be no matches on representor PCI
2850 			 * functions or non VF LAG bonding devices with
2851 			 * specified address.
2852 			 */
2853 			if (nd) {
2854 				DRV_LOG(ERR,
2855 					"multiple PCI match on bonding device"
2856 					"\"%s\" found", ibv_list[ret]->name);
2857 				rte_errno = ENOENT;
2858 				ret = -rte_errno;
2859 				goto exit;
2860 			}
2861 			DRV_LOG(INFO, "PCI information matches for"
2862 				      " slave %d bonding device \"%s\"",
2863 				      bd, ibv_list[ret]->name);
2864 			ibv_match[nd++] = ibv_list[ret];
2865 			break;
2866 		}
2867 		if (mlx5_dev_to_pci_addr
2868 			(ibv_list[ret]->ibdev_path, &pci_addr))
2869 			continue;
2870 		if (pci_dev->addr.domain != pci_addr.domain ||
2871 		    pci_dev->addr.bus != pci_addr.bus ||
2872 		    pci_dev->addr.devid != pci_addr.devid ||
2873 		    pci_dev->addr.function != pci_addr.function)
2874 			continue;
2875 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2876 			ibv_list[ret]->name);
2877 		ibv_match[nd++] = ibv_list[ret];
2878 	}
2879 	ibv_match[nd] = NULL;
2880 	if (!nd) {
2881 		/* No device matches, just complain and bail out. */
2882 		DRV_LOG(WARNING,
2883 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2884 			" are kernel drivers loaded?",
2885 			pci_dev->addr.domain, pci_dev->addr.bus,
2886 			pci_dev->addr.devid, pci_dev->addr.function);
2887 		rte_errno = ENOENT;
2888 		ret = -rte_errno;
2889 		goto exit;
2890 	}
2891 	if (nd == 1) {
2892 		/*
2893 		 * Found single matching device may have multiple ports.
2894 		 * Each port may be representor, we have to check the port
2895 		 * number and check the representors existence.
2896 		 */
2897 		if (nl_rdma >= 0)
2898 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2899 		if (!np)
2900 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2901 					 " ports number", ibv_match[0]->name);
2902 		if (bd >= 0 && !np) {
2903 			DRV_LOG(ERR, "can not get ports"
2904 				     " for bonding device");
2905 			rte_errno = ENOENT;
2906 			ret = -rte_errno;
2907 			goto exit;
2908 		}
2909 	}
2910 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2911 	if (bd >= 0) {
2912 		/*
2913 		 * This may happen if there is VF LAG kernel support and
2914 		 * application is compiled with older rdma_core library.
2915 		 */
2916 		DRV_LOG(ERR,
2917 			"No kernel/verbs support for VF LAG bonding found.");
2918 		rte_errno = ENOTSUP;
2919 		ret = -rte_errno;
2920 		goto exit;
2921 	}
2922 #endif
2923 	/*
2924 	 * Now we can determine the maximal
2925 	 * amount of devices to be spawned.
2926 	 */
2927 	list = rte_zmalloc("device spawn data",
2928 			 sizeof(struct mlx5_dev_spawn_data) *
2929 			 (np ? np : nd),
2930 			 RTE_CACHE_LINE_SIZE);
2931 	if (!list) {
2932 		DRV_LOG(ERR, "spawn data array allocation failure");
2933 		rte_errno = ENOMEM;
2934 		ret = -rte_errno;
2935 		goto exit;
2936 	}
2937 	if (bd >= 0 || np > 1) {
2938 		/*
2939 		 * Single IB device with multiple ports found,
2940 		 * it may be E-Switch master device and representors.
2941 		 * We have to perform identification trough the ports.
2942 		 */
2943 		assert(nl_rdma >= 0);
2944 		assert(ns == 0);
2945 		assert(nd == 1);
2946 		assert(np);
2947 		for (i = 1; i <= np; ++i) {
2948 			list[ns].max_port = np;
2949 			list[ns].ibv_port = i;
2950 			list[ns].ibv_dev = ibv_match[0];
2951 			list[ns].eth_dev = NULL;
2952 			list[ns].pci_dev = pci_dev;
2953 			list[ns].pf_bond = bd;
2954 			list[ns].ifindex = mlx5_nl_ifindex
2955 					(nl_rdma, list[ns].ibv_dev->name, i);
2956 			if (!list[ns].ifindex) {
2957 				/*
2958 				 * No network interface index found for the
2959 				 * specified port, it means there is no
2960 				 * representor on this port. It's OK,
2961 				 * there can be disabled ports, for example
2962 				 * if sriov_numvfs < sriov_totalvfs.
2963 				 */
2964 				continue;
2965 			}
2966 			ret = -1;
2967 			if (nl_route >= 0)
2968 				ret = mlx5_nl_switch_info
2969 					       (nl_route,
2970 						list[ns].ifindex,
2971 						&list[ns].info);
2972 			if (ret || (!list[ns].info.representor &&
2973 				    !list[ns].info.master)) {
2974 				/*
2975 				 * We failed to recognize representors with
2976 				 * Netlink, let's try to perform the task
2977 				 * with sysfs.
2978 				 */
2979 				ret =  mlx5_sysfs_switch_info
2980 						(list[ns].ifindex,
2981 						 &list[ns].info);
2982 			}
2983 			if (!ret && bd >= 0) {
2984 				switch (list[ns].info.name_type) {
2985 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2986 					if (list[ns].info.port_name == bd)
2987 						ns++;
2988 					break;
2989 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2990 					if (list[ns].info.pf_num == bd)
2991 						ns++;
2992 					break;
2993 				default:
2994 					break;
2995 				}
2996 				continue;
2997 			}
2998 			if (!ret && (list[ns].info.representor ^
2999 				     list[ns].info.master))
3000 				ns++;
3001 		}
3002 		if (!ns) {
3003 			DRV_LOG(ERR,
3004 				"unable to recognize master/representors"
3005 				" on the IB device with multiple ports");
3006 			rte_errno = ENOENT;
3007 			ret = -rte_errno;
3008 			goto exit;
3009 		}
3010 	} else {
3011 		/*
3012 		 * The existence of several matching entries (nd > 1) means
3013 		 * port representors have been instantiated. No existing Verbs
3014 		 * call nor sysfs entries can tell them apart, this can only
3015 		 * be done through Netlink calls assuming kernel drivers are
3016 		 * recent enough to support them.
3017 		 *
3018 		 * In the event of identification failure through Netlink,
3019 		 * try again through sysfs, then:
3020 		 *
3021 		 * 1. A single IB device matches (nd == 1) with single
3022 		 *    port (np=0/1) and is not a representor, assume
3023 		 *    no switch support.
3024 		 *
3025 		 * 2. Otherwise no safe assumptions can be made;
3026 		 *    complain louder and bail out.
3027 		 */
3028 		np = 1;
3029 		for (i = 0; i != nd; ++i) {
3030 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3031 			list[ns].max_port = 1;
3032 			list[ns].ibv_port = 1;
3033 			list[ns].ibv_dev = ibv_match[i];
3034 			list[ns].eth_dev = NULL;
3035 			list[ns].pci_dev = pci_dev;
3036 			list[ns].pf_bond = -1;
3037 			list[ns].ifindex = 0;
3038 			if (nl_rdma >= 0)
3039 				list[ns].ifindex = mlx5_nl_ifindex
3040 					(nl_rdma, list[ns].ibv_dev->name, 1);
3041 			if (!list[ns].ifindex) {
3042 				char ifname[IF_NAMESIZE];
3043 
3044 				/*
3045 				 * Netlink failed, it may happen with old
3046 				 * ib_core kernel driver (before 4.16).
3047 				 * We can assume there is old driver because
3048 				 * here we are processing single ports IB
3049 				 * devices. Let's try sysfs to retrieve
3050 				 * the ifindex. The method works for
3051 				 * master device only.
3052 				 */
3053 				if (nd > 1) {
3054 					/*
3055 					 * Multiple devices found, assume
3056 					 * representors, can not distinguish
3057 					 * master/representor and retrieve
3058 					 * ifindex via sysfs.
3059 					 */
3060 					continue;
3061 				}
3062 				ret = mlx5_get_master_ifname
3063 					(ibv_match[i]->ibdev_path, &ifname);
3064 				if (!ret)
3065 					list[ns].ifindex =
3066 						if_nametoindex(ifname);
3067 				if (!list[ns].ifindex) {
3068 					/*
3069 					 * No network interface index found
3070 					 * for the specified device, it means
3071 					 * there it is neither representor
3072 					 * nor master.
3073 					 */
3074 					continue;
3075 				}
3076 			}
3077 			ret = -1;
3078 			if (nl_route >= 0)
3079 				ret = mlx5_nl_switch_info
3080 					       (nl_route,
3081 						list[ns].ifindex,
3082 						&list[ns].info);
3083 			if (ret || (!list[ns].info.representor &&
3084 				    !list[ns].info.master)) {
3085 				/*
3086 				 * We failed to recognize representors with
3087 				 * Netlink, let's try to perform the task
3088 				 * with sysfs.
3089 				 */
3090 				ret =  mlx5_sysfs_switch_info
3091 						(list[ns].ifindex,
3092 						 &list[ns].info);
3093 			}
3094 			if (!ret && (list[ns].info.representor ^
3095 				     list[ns].info.master)) {
3096 				ns++;
3097 			} else if ((nd == 1) &&
3098 				   !list[ns].info.representor &&
3099 				   !list[ns].info.master) {
3100 				/*
3101 				 * Single IB device with
3102 				 * one physical port and
3103 				 * attached network device.
3104 				 * May be SRIOV is not enabled
3105 				 * or there is no representors.
3106 				 */
3107 				DRV_LOG(INFO, "no E-Switch support detected");
3108 				ns++;
3109 				break;
3110 			}
3111 		}
3112 		if (!ns) {
3113 			DRV_LOG(ERR,
3114 				"unable to recognize master/representors"
3115 				" on the multiple IB devices");
3116 			rte_errno = ENOENT;
3117 			ret = -rte_errno;
3118 			goto exit;
3119 		}
3120 	}
3121 	assert(ns);
3122 	/*
3123 	 * Sort list to probe devices in natural order for users convenience
3124 	 * (i.e. master first, then representors from lowest to highest ID).
3125 	 */
3126 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3127 	/* Default configuration. */
3128 	dev_config = (struct mlx5_dev_config){
3129 		.hw_padding = 0,
3130 		.mps = MLX5_ARG_UNSET,
3131 		.dbnc = MLX5_ARG_UNSET,
3132 		.rx_vec_en = 1,
3133 		.txq_inline_max = MLX5_ARG_UNSET,
3134 		.txq_inline_min = MLX5_ARG_UNSET,
3135 		.txq_inline_mpw = MLX5_ARG_UNSET,
3136 		.txqs_inline = MLX5_ARG_UNSET,
3137 		.vf_nl_en = 1,
3138 		.mr_ext_memseg_en = 1,
3139 		.mprq = {
3140 			.enabled = 0, /* Disabled by default. */
3141 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3142 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3143 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3144 		},
3145 		.dv_esw_en = 1,
3146 	};
3147 	/* Device specific configuration. */
3148 	switch (pci_dev->id.device_id) {
3149 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3150 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3151 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3152 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3153 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3154 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3155 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3156 		dev_config.vf = 1;
3157 		break;
3158 	default:
3159 		break;
3160 	}
3161 	for (i = 0; i != ns; ++i) {
3162 		uint32_t restore;
3163 
3164 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3165 						 &list[i],
3166 						 dev_config);
3167 		if (!list[i].eth_dev) {
3168 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3169 				break;
3170 			/* Device is disabled or already spawned. Ignore it. */
3171 			continue;
3172 		}
3173 		restore = list[i].eth_dev->data->dev_flags;
3174 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3175 		/* Restore non-PCI flags cleared by the above call. */
3176 		list[i].eth_dev->data->dev_flags |= restore;
3177 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3178 		rte_eth_dev_probing_finish(list[i].eth_dev);
3179 	}
3180 	if (i != ns) {
3181 		DRV_LOG(ERR,
3182 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3183 			" encountering an error: %s",
3184 			pci_dev->addr.domain, pci_dev->addr.bus,
3185 			pci_dev->addr.devid, pci_dev->addr.function,
3186 			strerror(rte_errno));
3187 		ret = -rte_errno;
3188 		/* Roll back. */
3189 		while (i--) {
3190 			if (!list[i].eth_dev)
3191 				continue;
3192 			mlx5_dev_close(list[i].eth_dev);
3193 			/* mac_addrs must not be freed because in dev_private */
3194 			list[i].eth_dev->data->mac_addrs = NULL;
3195 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3196 		}
3197 		/* Restore original error. */
3198 		rte_errno = -ret;
3199 	} else {
3200 		ret = 0;
3201 	}
3202 exit:
3203 	/*
3204 	 * Do the routine cleanup:
3205 	 * - close opened Netlink sockets
3206 	 * - free allocated spawn data array
3207 	 * - free the Infiniband device list
3208 	 */
3209 	if (nl_rdma >= 0)
3210 		close(nl_rdma);
3211 	if (nl_route >= 0)
3212 		close(nl_route);
3213 	if (list)
3214 		rte_free(list);
3215 	assert(ibv_list);
3216 	mlx5_glue->free_device_list(ibv_list);
3217 	return ret;
3218 }
3219 
3220 /**
3221  * Look for the ethernet device belonging to mlx5 driver.
3222  *
3223  * @param[in] port_id
3224  *   port_id to start looking for device.
3225  * @param[in] pci_dev
3226  *   Pointer to the hint PCI device. When device is being probed
3227  *   the its siblings (master and preceding representors might
3228  *   not have assigned driver yet (because the mlx5_pci_probe()
3229  *   is not completed yet, for this case match on hint PCI
3230  *   device may be used to detect sibling device.
3231  *
3232  * @return
3233  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3234  */
3235 uint16_t
3236 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3237 {
3238 	while (port_id < RTE_MAX_ETHPORTS) {
3239 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3240 
3241 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3242 		    dev->device &&
3243 		    (dev->device == &pci_dev->device ||
3244 		     (dev->device->driver &&
3245 		     dev->device->driver->name &&
3246 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3247 			break;
3248 		port_id++;
3249 	}
3250 	if (port_id >= RTE_MAX_ETHPORTS)
3251 		return RTE_MAX_ETHPORTS;
3252 	return port_id;
3253 }
3254 
3255 /**
3256  * DPDK callback to remove a PCI device.
3257  *
3258  * This function removes all Ethernet devices belong to a given PCI device.
3259  *
3260  * @param[in] pci_dev
3261  *   Pointer to the PCI device.
3262  *
3263  * @return
3264  *   0 on success, the function cannot fail.
3265  */
3266 static int
3267 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3268 {
3269 	uint16_t port_id;
3270 
3271 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3272 		rte_eth_dev_close(port_id);
3273 	return 0;
3274 }
3275 
3276 static const struct rte_pci_id mlx5_pci_id_map[] = {
3277 	{
3278 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3279 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3280 	},
3281 	{
3282 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3283 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3284 	},
3285 	{
3286 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3287 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3288 	},
3289 	{
3290 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3291 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3292 	},
3293 	{
3294 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3295 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3296 	},
3297 	{
3298 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3299 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3300 	},
3301 	{
3302 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3303 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3304 	},
3305 	{
3306 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3307 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3308 	},
3309 	{
3310 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3311 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3312 	},
3313 	{
3314 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3315 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3316 	},
3317 	{
3318 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3319 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3320 	},
3321 	{
3322 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3323 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3324 	},
3325 	{
3326 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3327 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3328 	},
3329 	{
3330 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3331 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3332 	},
3333 	{
3334 		.vendor_id = 0
3335 	}
3336 };
3337 
3338 static struct rte_pci_driver mlx5_driver = {
3339 	.driver = {
3340 		.name = MLX5_DRIVER_NAME
3341 	},
3342 	.id_table = mlx5_pci_id_map,
3343 	.probe = mlx5_pci_probe,
3344 	.remove = mlx5_pci_remove,
3345 	.dma_map = mlx5_dma_map,
3346 	.dma_unmap = mlx5_dma_unmap,
3347 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3348 		     RTE_PCI_DRV_PROBE_AGAIN,
3349 };
3350 
3351 #ifdef RTE_IBVERBS_LINK_DLOPEN
3352 
3353 /**
3354  * Suffix RTE_EAL_PMD_PATH with "-glue".
3355  *
3356  * This function performs a sanity check on RTE_EAL_PMD_PATH before
3357  * suffixing its last component.
3358  *
3359  * @param buf[out]
3360  *   Output buffer, should be large enough otherwise NULL is returned.
3361  * @param size
3362  *   Size of @p out.
3363  *
3364  * @return
3365  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
3366  */
3367 static char *
3368 mlx5_glue_path(char *buf, size_t size)
3369 {
3370 	static const char *const bad[] = { "/", ".", "..", NULL };
3371 	const char *path = RTE_EAL_PMD_PATH;
3372 	size_t len = strlen(path);
3373 	size_t off;
3374 	int i;
3375 
3376 	while (len && path[len - 1] == '/')
3377 		--len;
3378 	for (off = len; off && path[off - 1] != '/'; --off)
3379 		;
3380 	for (i = 0; bad[i]; ++i)
3381 		if (!strncmp(path + off, bad[i], (int)(len - off)))
3382 			goto error;
3383 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
3384 	if (i == -1 || (size_t)i >= size)
3385 		goto error;
3386 	return buf;
3387 error:
3388 	DRV_LOG(ERR,
3389 		"unable to append \"-glue\" to last component of"
3390 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
3391 		" please re-configure DPDK");
3392 	return NULL;
3393 }
3394 
3395 /**
3396  * Initialization routine for run-time dependency on rdma-core.
3397  */
3398 static int
3399 mlx5_glue_init(void)
3400 {
3401 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
3402 	const char *path[] = {
3403 		/*
3404 		 * A basic security check is necessary before trusting
3405 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
3406 		 */
3407 		(geteuid() == getuid() && getegid() == getgid() ?
3408 		 getenv("MLX5_GLUE_PATH") : NULL),
3409 		/*
3410 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
3411 		 * variant, otherwise let dlopen() look up libraries on its
3412 		 * own.
3413 		 */
3414 		(*RTE_EAL_PMD_PATH ?
3415 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
3416 	};
3417 	unsigned int i = 0;
3418 	void *handle = NULL;
3419 	void **sym;
3420 	const char *dlmsg;
3421 
3422 	while (!handle && i != RTE_DIM(path)) {
3423 		const char *end;
3424 		size_t len;
3425 		int ret;
3426 
3427 		if (!path[i]) {
3428 			++i;
3429 			continue;
3430 		}
3431 		end = strpbrk(path[i], ":;");
3432 		if (!end)
3433 			end = path[i] + strlen(path[i]);
3434 		len = end - path[i];
3435 		ret = 0;
3436 		do {
3437 			char name[ret + 1];
3438 
3439 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
3440 				       (int)len, path[i],
3441 				       (!len || *(end - 1) == '/') ? "" : "/");
3442 			if (ret == -1)
3443 				break;
3444 			if (sizeof(name) != (size_t)ret + 1)
3445 				continue;
3446 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
3447 				name);
3448 			handle = dlopen(name, RTLD_LAZY);
3449 			break;
3450 		} while (1);
3451 		path[i] = end + 1;
3452 		if (!*end)
3453 			++i;
3454 	}
3455 	if (!handle) {
3456 		rte_errno = EINVAL;
3457 		dlmsg = dlerror();
3458 		if (dlmsg)
3459 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
3460 		goto glue_error;
3461 	}
3462 	sym = dlsym(handle, "mlx5_glue");
3463 	if (!sym || !*sym) {
3464 		rte_errno = EINVAL;
3465 		dlmsg = dlerror();
3466 		if (dlmsg)
3467 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
3468 		goto glue_error;
3469 	}
3470 	mlx5_glue = *sym;
3471 	return 0;
3472 glue_error:
3473 	if (handle)
3474 		dlclose(handle);
3475 	DRV_LOG(WARNING,
3476 		"cannot initialize PMD due to missing run-time dependency on"
3477 		" rdma-core libraries (libibverbs, libmlx5)");
3478 	return -rte_errno;
3479 }
3480 
3481 #endif
3482 
3483 /**
3484  * Driver initialization routine.
3485  */
3486 RTE_INIT(rte_mlx5_pmd_init)
3487 {
3488 	/* Initialize driver log type. */
3489 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3490 	if (mlx5_logtype >= 0)
3491 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3492 
3493 	/* Build the static tables for Verbs conversion. */
3494 	mlx5_set_ptype_table();
3495 	mlx5_set_cksum_table();
3496 	mlx5_set_swp_types_table();
3497 	/*
3498 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3499 	 * huge pages. Calling ibv_fork_init() during init allows
3500 	 * applications to use fork() safely for purposes other than
3501 	 * using this PMD, which is not supported in forked processes.
3502 	 */
3503 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3504 	/* Match the size of Rx completion entry to the size of a cacheline. */
3505 	if (RTE_CACHE_LINE_SIZE == 128)
3506 		setenv("MLX5_CQE_SIZE", "128", 0);
3507 	/*
3508 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
3509 	 * cleanup all the Verbs resources even when the device was removed.
3510 	 */
3511 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
3512 #ifdef RTE_IBVERBS_LINK_DLOPEN
3513 	if (mlx5_glue_init())
3514 		return;
3515 	assert(mlx5_glue);
3516 #endif
3517 #ifndef NDEBUG
3518 	/* Glue structure must not contain any NULL pointers. */
3519 	{
3520 		unsigned int i;
3521 
3522 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
3523 			assert(((const void *const *)mlx5_glue)[i]);
3524 	}
3525 #endif
3526 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
3527 		DRV_LOG(ERR,
3528 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
3529 			mlx5_glue->version, MLX5_GLUE_VERSION);
3530 		return;
3531 	}
3532 	mlx5_glue->fork_init();
3533 	rte_pci_register(&mlx5_driver);
3534 }
3535 
3536 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3537 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3538 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3539