xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 7adf992fb9bf7162a7edc45b50d10fbb1d57824d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <sys/mman.h>
14 #include <linux/rtnetlink.h>
15 
16 /* Verbs header. */
17 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
18 #ifdef PEDANTIC
19 #pragma GCC diagnostic ignored "-Wpedantic"
20 #endif
21 #include <infiniband/verbs.h>
22 #ifdef PEDANTIC
23 #pragma GCC diagnostic error "-Wpedantic"
24 #endif
25 
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_ethdev_pci.h>
29 #include <rte_pci.h>
30 #include <rte_bus_pci.h>
31 #include <rte_common.h>
32 #include <rte_config.h>
33 #include <rte_kvargs.h>
34 #include <rte_rwlock.h>
35 #include <rte_spinlock.h>
36 #include <rte_string_fns.h>
37 #include <rte_alarm.h>
38 
39 #include <mlx5_glue.h>
40 #include <mlx5_devx_cmds.h>
41 #include <mlx5_common.h>
42 
43 #include "mlx5_defs.h"
44 #include "mlx5.h"
45 #include "mlx5_utils.h"
46 #include "mlx5_rxtx.h"
47 #include "mlx5_autoconf.h"
48 #include "mlx5_mr.h"
49 #include "mlx5_flow.h"
50 #include "rte_pmd_mlx5.h"
51 
52 /* Device parameter to enable RX completion queue compression. */
53 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
54 
55 /* Device parameter to enable RX completion entry padding to 128B. */
56 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
57 
58 /* Device parameter to enable padding Rx packet to cacheline size. */
59 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
60 
61 /* Device parameter to enable Multi-Packet Rx queue. */
62 #define MLX5_RX_MPRQ_EN "mprq_en"
63 
64 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
65 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
66 
67 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
68 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
69 
70 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
71 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
72 
73 /* Device parameter to configure inline send. Deprecated, ignored.*/
74 #define MLX5_TXQ_INLINE "txq_inline"
75 
76 /* Device parameter to limit packet size to inline with ordinary SEND. */
77 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
78 
79 /* Device parameter to configure minimal data size to inline. */
80 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
81 
82 /* Device parameter to limit packet size to inline with Enhanced MPW. */
83 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
84 
85 /*
86  * Device parameter to configure the number of TX queues threshold for
87  * enabling inline send.
88  */
89 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
90 
91 /*
92  * Device parameter to configure the number of TX queues threshold for
93  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
94  */
95 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
96 
97 /* Device parameter to enable multi-packet send WQEs. */
98 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
99 
100 /*
101  * Device parameter to force doorbell register mapping
102  * to non-cahed region eliminating the extra write memory barrier.
103  */
104 #define MLX5_TX_DB_NC "tx_db_nc"
105 
106 /*
107  * Device parameter to include 2 dsegs in the title WQEBB.
108  * Deprecated, ignored.
109  */
110 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
111 
112 /*
113  * Device parameter to limit the size of inlining packet.
114  * Deprecated, ignored.
115  */
116 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
117 
118 /*
119  * Device parameter to enable hardware Tx vector.
120  * Deprecated, ignored (no vectorized Tx routines anymore).
121  */
122 #define MLX5_TX_VEC_EN "tx_vec_en"
123 
124 /* Device parameter to enable hardware Rx vector. */
125 #define MLX5_RX_VEC_EN "rx_vec_en"
126 
127 /* Allow L3 VXLAN flow creation. */
128 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
129 
130 /* Activate DV E-Switch flow steering. */
131 #define MLX5_DV_ESW_EN "dv_esw_en"
132 
133 /* Activate DV flow steering. */
134 #define MLX5_DV_FLOW_EN "dv_flow_en"
135 
136 /* Enable extensive flow metadata support. */
137 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
138 
139 /* Activate Netlink support in VF mode. */
140 #define MLX5_VF_NL_EN "vf_nl_en"
141 
142 /* Enable extending memsegs when creating a MR. */
143 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
144 
145 /* Select port representors to instantiate. */
146 #define MLX5_REPRESENTOR "representor"
147 
148 /* Device parameter to configure the maximum number of dump files per queue. */
149 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
150 
151 /* Configure timeout of LRO session (in microseconds). */
152 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
153 
154 #ifndef HAVE_IBV_MLX5_MOD_MPW
155 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
156 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
157 #endif
158 
159 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
160 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
161 #endif
162 
163 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
164 
165 /* Shared memory between primary and secondary processes. */
166 struct mlx5_shared_data *mlx5_shared_data;
167 
168 /* Spinlock for mlx5_shared_data allocation. */
169 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
170 
171 /* Process local data for secondary processes. */
172 static struct mlx5_local_data mlx5_local_data;
173 
174 /** Driver-specific log messages type. */
175 int mlx5_logtype;
176 
177 /** Data associated with devices to spawn. */
178 struct mlx5_dev_spawn_data {
179 	uint32_t ifindex; /**< Network interface index. */
180 	uint32_t max_port; /**< IB device maximal port index. */
181 	uint32_t ibv_port; /**< IB device physical port index. */
182 	int pf_bond; /**< bonding device PF index. < 0 - no bonding */
183 	struct mlx5_switch_info info; /**< Switch information. */
184 	struct ibv_device *ibv_dev; /**< Associated IB device. */
185 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
186 	struct rte_pci_device *pci_dev; /**< Backend PCI device. */
187 };
188 
189 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
190 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
191 
192 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
193 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
194 
195 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
196 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
197 
198 /**
199  * Allocate ID pool structure.
200  *
201  * @param[in] max_id
202  *   The maximum id can be allocated from the pool.
203  *
204  * @return
205  *   Pointer to pool object, NULL value otherwise.
206  */
207 struct mlx5_flow_id_pool *
208 mlx5_flow_id_pool_alloc(uint32_t max_id)
209 {
210 	struct mlx5_flow_id_pool *pool;
211 	void *mem;
212 
213 	pool = rte_zmalloc("id pool allocation", sizeof(*pool),
214 			   RTE_CACHE_LINE_SIZE);
215 	if (!pool) {
216 		DRV_LOG(ERR, "can't allocate id pool");
217 		rte_errno  = ENOMEM;
218 		return NULL;
219 	}
220 	mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
221 			  RTE_CACHE_LINE_SIZE);
222 	if (!mem) {
223 		DRV_LOG(ERR, "can't allocate mem for id pool");
224 		rte_errno  = ENOMEM;
225 		goto error;
226 	}
227 	pool->free_arr = mem;
228 	pool->curr = pool->free_arr;
229 	pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
230 	pool->base_index = 0;
231 	pool->max_id = max_id;
232 	return pool;
233 error:
234 	rte_free(pool);
235 	return NULL;
236 }
237 
238 /**
239  * Release ID pool structure.
240  *
241  * @param[in] pool
242  *   Pointer to flow id pool object to free.
243  */
244 void
245 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
246 {
247 	rte_free(pool->free_arr);
248 	rte_free(pool);
249 }
250 
251 /**
252  * Generate ID.
253  *
254  * @param[in] pool
255  *   Pointer to flow id pool.
256  * @param[out] id
257  *   The generated ID.
258  *
259  * @return
260  *   0 on success, error value otherwise.
261  */
262 uint32_t
263 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
264 {
265 	if (pool->curr == pool->free_arr) {
266 		if (pool->base_index == pool->max_id) {
267 			rte_errno  = ENOMEM;
268 			DRV_LOG(ERR, "no free id");
269 			return -rte_errno;
270 		}
271 		*id = ++pool->base_index;
272 		return 0;
273 	}
274 	*id = *(--pool->curr);
275 	return 0;
276 }
277 
278 /**
279  * Release ID.
280  *
281  * @param[in] pool
282  *   Pointer to flow id pool.
283  * @param[out] id
284  *   The generated ID.
285  *
286  * @return
287  *   0 on success, error value otherwise.
288  */
289 uint32_t
290 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
291 {
292 	uint32_t size;
293 	uint32_t size2;
294 	void *mem;
295 
296 	if (pool->curr == pool->last) {
297 		size = pool->curr - pool->free_arr;
298 		size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
299 		MLX5_ASSERT(size2 > size);
300 		mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
301 		if (!mem) {
302 			DRV_LOG(ERR, "can't allocate mem for id pool");
303 			rte_errno  = ENOMEM;
304 			return -rte_errno;
305 		}
306 		memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
307 		rte_free(pool->free_arr);
308 		pool->free_arr = mem;
309 		pool->curr = pool->free_arr + size;
310 		pool->last = pool->free_arr + size2;
311 	}
312 	*pool->curr = id;
313 	pool->curr++;
314 	return 0;
315 }
316 
317 /**
318  * Initialize the counters management structure.
319  *
320  * @param[in] sh
321  *   Pointer to mlx5_ibv_shared object to free
322  */
323 static void
324 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
325 {
326 	uint8_t i;
327 
328 	TAILQ_INIT(&sh->cmng.flow_counters);
329 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
330 		TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
331 }
332 
333 /**
334  * Destroy all the resources allocated for a counter memory management.
335  *
336  * @param[in] mng
337  *   Pointer to the memory management structure.
338  */
339 static void
340 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
341 {
342 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
343 
344 	LIST_REMOVE(mng, next);
345 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
346 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
347 	rte_free(mem);
348 }
349 
350 /**
351  * Close and release all the resources of the counters management.
352  *
353  * @param[in] sh
354  *   Pointer to mlx5_ibv_shared object to free.
355  */
356 static void
357 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
358 {
359 	struct mlx5_counter_stats_mem_mng *mng;
360 	uint8_t i;
361 	int j;
362 	int retries = 1024;
363 
364 	rte_errno = 0;
365 	while (--retries) {
366 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
367 		if (rte_errno != EINPROGRESS)
368 			break;
369 		rte_pause();
370 	}
371 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
372 		struct mlx5_flow_counter_pool *pool;
373 		uint32_t batch = !!(i % 2);
374 
375 		if (!sh->cmng.ccont[i].pools)
376 			continue;
377 		pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
378 		while (pool) {
379 			if (batch) {
380 				if (pool->min_dcs)
381 					claim_zero
382 					(mlx5_devx_cmd_destroy(pool->min_dcs));
383 			}
384 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
385 				if (pool->counters_raw[j].action)
386 					claim_zero
387 					(mlx5_glue->destroy_flow_action
388 					       (pool->counters_raw[j].action));
389 				if (!batch && pool->counters_raw[j].dcs)
390 					claim_zero(mlx5_devx_cmd_destroy
391 						  (pool->counters_raw[j].dcs));
392 			}
393 			TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
394 				     next);
395 			rte_free(pool);
396 			pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
397 		}
398 		rte_free(sh->cmng.ccont[i].pools);
399 	}
400 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
401 	while (mng) {
402 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
403 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
404 	}
405 	memset(&sh->cmng, 0, sizeof(sh->cmng));
406 }
407 
408 /**
409  * Extract pdn of PD object using DV API.
410  *
411  * @param[in] pd
412  *   Pointer to the verbs PD object.
413  * @param[out] pdn
414  *   Pointer to the PD object number variable.
415  *
416  * @return
417  *   0 on success, error value otherwise.
418  */
419 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
420 static int
421 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
422 {
423 	struct mlx5dv_obj obj;
424 	struct mlx5dv_pd pd_info;
425 	int ret = 0;
426 
427 	obj.pd.in = pd;
428 	obj.pd.out = &pd_info;
429 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
430 	if (ret) {
431 		DRV_LOG(DEBUG, "Fail to get PD object info");
432 		return ret;
433 	}
434 	*pdn = pd_info.pdn;
435 	return 0;
436 }
437 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
438 
439 static int
440 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
441 {
442 	char *env;
443 	int value;
444 
445 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
446 	/* Get environment variable to store. */
447 	env = getenv(MLX5_SHUT_UP_BF);
448 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
449 	if (config->dbnc == MLX5_ARG_UNSET)
450 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
451 	else
452 		setenv(MLX5_SHUT_UP_BF,
453 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
454 	return value;
455 }
456 
457 static void
458 mlx5_restore_doorbell_mapping_env(int value)
459 {
460 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
461 	/* Restore the original environment variable state. */
462 	if (value == MLX5_ARG_UNSET)
463 		unsetenv(MLX5_SHUT_UP_BF);
464 	else
465 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
466 }
467 
468 /**
469  * Allocate shared IB device context. If there is multiport device the
470  * master and representors will share this context, if there is single
471  * port dedicated IB device, the context will be used by only given
472  * port due to unification.
473  *
474  * Routine first searches the context for the specified IB device name,
475  * if found the shared context assumed and reference counter is incremented.
476  * If no context found the new one is created and initialized with specified
477  * IB device context and parameters.
478  *
479  * @param[in] spawn
480  *   Pointer to the IB device attributes (name, port, etc).
481  * @param[in] config
482  *   Pointer to device configuration structure.
483  *
484  * @return
485  *   Pointer to mlx5_ibv_shared object on success,
486  *   otherwise NULL and rte_errno is set.
487  */
488 static struct mlx5_ibv_shared *
489 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
490 			const struct mlx5_dev_config *config)
491 {
492 	struct mlx5_ibv_shared *sh;
493 	int dbmap_env;
494 	int err = 0;
495 	uint32_t i;
496 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
497 	struct mlx5_devx_tis_attr tis_attr = { 0 };
498 #endif
499 
500 	MLX5_ASSERT(spawn);
501 	/* Secondary process should not create the shared context. */
502 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
503 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
504 	/* Search for IB context by device name. */
505 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
506 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
507 			sh->refcnt++;
508 			goto exit;
509 		}
510 	}
511 	/* No device found, we have to create new shared context. */
512 	MLX5_ASSERT(spawn->max_port);
513 	sh = rte_zmalloc("ethdev shared ib context",
514 			 sizeof(struct mlx5_ibv_shared) +
515 			 spawn->max_port *
516 			 sizeof(struct mlx5_ibv_shared_port),
517 			 RTE_CACHE_LINE_SIZE);
518 	if (!sh) {
519 		DRV_LOG(ERR, "shared context allocation failure");
520 		rte_errno  = ENOMEM;
521 		goto exit;
522 	}
523 	/*
524 	 * Configure environment variable "MLX5_BF_SHUT_UP"
525 	 * before the device creation. The rdma_core library
526 	 * checks the variable at device creation and
527 	 * stores the result internally.
528 	 */
529 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
530 	/* Try to open IB device with DV first, then usual Verbs. */
531 	errno = 0;
532 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
533 	if (sh->ctx) {
534 		sh->devx = 1;
535 		DRV_LOG(DEBUG, "DevX is supported");
536 		/* The device is created, no need for environment. */
537 		mlx5_restore_doorbell_mapping_env(dbmap_env);
538 	} else {
539 		/* The environment variable is still configured. */
540 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
541 		err = errno ? errno : ENODEV;
542 		/*
543 		 * The environment variable is not needed anymore,
544 		 * all device creation attempts are completed.
545 		 */
546 		mlx5_restore_doorbell_mapping_env(dbmap_env);
547 		if (!sh->ctx)
548 			goto error;
549 		DRV_LOG(DEBUG, "DevX is NOT supported");
550 	}
551 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
552 	if (err) {
553 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
554 		goto error;
555 	}
556 	sh->refcnt = 1;
557 	sh->max_port = spawn->max_port;
558 	strncpy(sh->ibdev_name, sh->ctx->device->name,
559 		sizeof(sh->ibdev_name));
560 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
561 		sizeof(sh->ibdev_path));
562 	pthread_mutex_init(&sh->intr_mutex, NULL);
563 	/*
564 	 * Setting port_id to max unallowed value means
565 	 * there is no interrupt subhandler installed for
566 	 * the given port index i.
567 	 */
568 	for (i = 0; i < sh->max_port; i++) {
569 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
570 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
571 	}
572 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
573 	if (sh->pd == NULL) {
574 		DRV_LOG(ERR, "PD allocation failure");
575 		err = ENOMEM;
576 		goto error;
577 	}
578 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
579 	if (sh->devx) {
580 		err = mlx5_get_pdn(sh->pd, &sh->pdn);
581 		if (err) {
582 			DRV_LOG(ERR, "Fail to extract pdn from PD");
583 			goto error;
584 		}
585 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
586 		if (!sh->td) {
587 			DRV_LOG(ERR, "TD allocation failure");
588 			err = ENOMEM;
589 			goto error;
590 		}
591 		tis_attr.transport_domain = sh->td->id;
592 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
593 		if (!sh->tis) {
594 			DRV_LOG(ERR, "TIS allocation failure");
595 			err = ENOMEM;
596 			goto error;
597 		}
598 	}
599 	sh->flow_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX);
600 	if (!sh->flow_id_pool) {
601 		DRV_LOG(ERR, "can't create flow id pool");
602 		err = ENOMEM;
603 		goto error;
604 	}
605 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
606 	/*
607 	 * Once the device is added to the list of memory event
608 	 * callback, its global MR cache table cannot be expanded
609 	 * on the fly because of deadlock. If it overflows, lookup
610 	 * should be done by searching MR list linearly, which is slow.
611 	 *
612 	 * At this point the device is not added to the memory
613 	 * event list yet, context is just being created.
614 	 */
615 	err = mlx5_mr_btree_init(&sh->mr.cache,
616 				 MLX5_MR_BTREE_CACHE_N * 2,
617 				 spawn->pci_dev->device.numa_node);
618 	if (err) {
619 		err = rte_errno;
620 		goto error;
621 	}
622 	mlx5_flow_counters_mng_init(sh);
623 	/* Add device to memory callback list. */
624 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
625 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
626 			 sh, mem_event_cb);
627 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
628 	/* Add context to the global device list. */
629 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
630 exit:
631 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
632 	return sh;
633 error:
634 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
635 	MLX5_ASSERT(sh);
636 	if (sh->tis)
637 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
638 	if (sh->td)
639 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
640 	if (sh->pd)
641 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
642 	if (sh->ctx)
643 		claim_zero(mlx5_glue->close_device(sh->ctx));
644 	if (sh->flow_id_pool)
645 		mlx5_flow_id_pool_release(sh->flow_id_pool);
646 	rte_free(sh);
647 	MLX5_ASSERT(err > 0);
648 	rte_errno = err;
649 	return NULL;
650 }
651 
652 /**
653  * Free shared IB device context. Decrement counter and if zero free
654  * all allocated resources and close handles.
655  *
656  * @param[in] sh
657  *   Pointer to mlx5_ibv_shared object to free
658  */
659 static void
660 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
661 {
662 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
663 #ifdef RTE_LIBRTE_MLX5_DEBUG
664 	/* Check the object presence in the list. */
665 	struct mlx5_ibv_shared *lctx;
666 
667 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
668 		if (lctx == sh)
669 			break;
670 	MLX5_ASSERT(lctx);
671 	if (lctx != sh) {
672 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
673 		goto exit;
674 	}
675 #endif
676 	MLX5_ASSERT(sh);
677 	MLX5_ASSERT(sh->refcnt);
678 	/* Secondary process should not free the shared context. */
679 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
680 	if (--sh->refcnt)
681 		goto exit;
682 	/* Remove from memory callback device list. */
683 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
684 	LIST_REMOVE(sh, mem_event_cb);
685 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
686 	/* Release created Memory Regions. */
687 	mlx5_mr_release(sh);
688 	/* Remove context from the global device list. */
689 	LIST_REMOVE(sh, next);
690 	/*
691 	 *  Ensure there is no async event handler installed.
692 	 *  Only primary process handles async device events.
693 	 **/
694 	mlx5_flow_counters_mng_close(sh);
695 	MLX5_ASSERT(!sh->intr_cnt);
696 	if (sh->intr_cnt)
697 		mlx5_intr_callback_unregister
698 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
699 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
700 	if (sh->devx_intr_cnt) {
701 		if (sh->intr_handle_devx.fd)
702 			rte_intr_callback_unregister(&sh->intr_handle_devx,
703 					  mlx5_dev_interrupt_handler_devx, sh);
704 		if (sh->devx_comp)
705 			mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
706 	}
707 #endif
708 	pthread_mutex_destroy(&sh->intr_mutex);
709 	if (sh->pd)
710 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
711 	if (sh->tis)
712 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
713 	if (sh->td)
714 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
715 	if (sh->ctx)
716 		claim_zero(mlx5_glue->close_device(sh->ctx));
717 	if (sh->flow_id_pool)
718 		mlx5_flow_id_pool_release(sh->flow_id_pool);
719 	rte_free(sh);
720 exit:
721 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
722 }
723 
724 /**
725  * Destroy table hash list and all the root entries per domain.
726  *
727  * @param[in] priv
728  *   Pointer to the private device data structure.
729  */
730 static void
731 mlx5_free_table_hash_list(struct mlx5_priv *priv)
732 {
733 	struct mlx5_ibv_shared *sh = priv->sh;
734 	struct mlx5_flow_tbl_data_entry *tbl_data;
735 	union mlx5_flow_tbl_key table_key = {
736 		{
737 			.table_id = 0,
738 			.reserved = 0,
739 			.domain = 0,
740 			.direction = 0,
741 		}
742 	};
743 	struct mlx5_hlist_entry *pos;
744 
745 	if (!sh->flow_tbls)
746 		return;
747 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
748 	if (pos) {
749 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
750 					entry);
751 		MLX5_ASSERT(tbl_data);
752 		mlx5_hlist_remove(sh->flow_tbls, pos);
753 		rte_free(tbl_data);
754 	}
755 	table_key.direction = 1;
756 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
757 	if (pos) {
758 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
759 					entry);
760 		MLX5_ASSERT(tbl_data);
761 		mlx5_hlist_remove(sh->flow_tbls, pos);
762 		rte_free(tbl_data);
763 	}
764 	table_key.direction = 0;
765 	table_key.domain = 1;
766 	pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
767 	if (pos) {
768 		tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
769 					entry);
770 		MLX5_ASSERT(tbl_data);
771 		mlx5_hlist_remove(sh->flow_tbls, pos);
772 		rte_free(tbl_data);
773 	}
774 	mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
775 }
776 
777 /**
778  * Initialize flow table hash list and create the root tables entry
779  * for each domain.
780  *
781  * @param[in] priv
782  *   Pointer to the private device data structure.
783  *
784  * @return
785  *   Zero on success, positive error code otherwise.
786  */
787 static int
788 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
789 {
790 	struct mlx5_ibv_shared *sh = priv->sh;
791 	char s[MLX5_HLIST_NAMESIZE];
792 	int err = 0;
793 
794 	MLX5_ASSERT(sh);
795 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
796 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
797 	if (!sh->flow_tbls) {
798 		DRV_LOG(ERR, "flow tables with hash creation failed.\n");
799 		err = ENOMEM;
800 		return err;
801 	}
802 #ifndef HAVE_MLX5DV_DR
803 	/*
804 	 * In case we have not DR support, the zero tables should be created
805 	 * because DV expect to see them even if they cannot be created by
806 	 * RDMA-CORE.
807 	 */
808 	union mlx5_flow_tbl_key table_key = {
809 		{
810 			.table_id = 0,
811 			.reserved = 0,
812 			.domain = 0,
813 			.direction = 0,
814 		}
815 	};
816 	struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
817 							  sizeof(*tbl_data), 0);
818 
819 	if (!tbl_data) {
820 		err = ENOMEM;
821 		goto error;
822 	}
823 	tbl_data->entry.key = table_key.v64;
824 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
825 	if (err)
826 		goto error;
827 	rte_atomic32_init(&tbl_data->tbl.refcnt);
828 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
829 	table_key.direction = 1;
830 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
831 	if (!tbl_data) {
832 		err = ENOMEM;
833 		goto error;
834 	}
835 	tbl_data->entry.key = table_key.v64;
836 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
837 	if (err)
838 		goto error;
839 	rte_atomic32_init(&tbl_data->tbl.refcnt);
840 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
841 	table_key.direction = 0;
842 	table_key.domain = 1;
843 	tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
844 	if (!tbl_data) {
845 		err = ENOMEM;
846 		goto error;
847 	}
848 	tbl_data->entry.key = table_key.v64;
849 	err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
850 	if (err)
851 		goto error;
852 	rte_atomic32_init(&tbl_data->tbl.refcnt);
853 	rte_atomic32_inc(&tbl_data->tbl.refcnt);
854 	return err;
855 error:
856 	mlx5_free_table_hash_list(priv);
857 #endif /* HAVE_MLX5DV_DR */
858 	return err;
859 }
860 
861 /**
862  * Initialize DR related data within private structure.
863  * Routine checks the reference counter and does actual
864  * resources creation/initialization only if counter is zero.
865  *
866  * @param[in] priv
867  *   Pointer to the private device data structure.
868  *
869  * @return
870  *   Zero on success, positive error code otherwise.
871  */
872 static int
873 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
874 {
875 	struct mlx5_ibv_shared *sh = priv->sh;
876 	char s[MLX5_HLIST_NAMESIZE];
877 	int err = 0;
878 
879 	if (!sh->flow_tbls)
880 		err = mlx5_alloc_table_hash_list(priv);
881 	else
882 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
883 			(void *)sh->flow_tbls);
884 	if (err)
885 		return err;
886 	/* Create tags hash list table. */
887 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
888 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
889 	if (!sh->tag_table) {
890 		DRV_LOG(ERR, "tags with hash creation failed.\n");
891 		err = ENOMEM;
892 		goto error;
893 	}
894 #ifdef HAVE_MLX5DV_DR
895 	void *domain;
896 
897 	if (sh->dv_refcnt) {
898 		/* Shared DV/DR structures is already initialized. */
899 		sh->dv_refcnt++;
900 		priv->dr_shared = 1;
901 		return 0;
902 	}
903 	/* Reference counter is zero, we should initialize structures. */
904 	domain = mlx5_glue->dr_create_domain(sh->ctx,
905 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
906 	if (!domain) {
907 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
908 		err = errno;
909 		goto error;
910 	}
911 	sh->rx_domain = domain;
912 	domain = mlx5_glue->dr_create_domain(sh->ctx,
913 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
914 	if (!domain) {
915 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
916 		err = errno;
917 		goto error;
918 	}
919 	pthread_mutex_init(&sh->dv_mutex, NULL);
920 	sh->tx_domain = domain;
921 #ifdef HAVE_MLX5DV_DR_ESWITCH
922 	if (priv->config.dv_esw_en) {
923 		domain  = mlx5_glue->dr_create_domain
924 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
925 		if (!domain) {
926 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
927 			err = errno;
928 			goto error;
929 		}
930 		sh->fdb_domain = domain;
931 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
932 	}
933 #endif
934 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
935 #endif /* HAVE_MLX5DV_DR */
936 	sh->dv_refcnt++;
937 	priv->dr_shared = 1;
938 	return 0;
939 error:
940 	/* Rollback the created objects. */
941 	if (sh->rx_domain) {
942 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
943 		sh->rx_domain = NULL;
944 	}
945 	if (sh->tx_domain) {
946 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
947 		sh->tx_domain = NULL;
948 	}
949 	if (sh->fdb_domain) {
950 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
951 		sh->fdb_domain = NULL;
952 	}
953 	if (sh->esw_drop_action) {
954 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
955 		sh->esw_drop_action = NULL;
956 	}
957 	if (sh->pop_vlan_action) {
958 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
959 		sh->pop_vlan_action = NULL;
960 	}
961 	if (sh->tag_table) {
962 		/* tags should be destroyed with flow before. */
963 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
964 		sh->tag_table = NULL;
965 	}
966 	mlx5_free_table_hash_list(priv);
967 	return err;
968 }
969 
970 /**
971  * Destroy DR related data within private structure.
972  *
973  * @param[in] priv
974  *   Pointer to the private device data structure.
975  */
976 static void
977 mlx5_free_shared_dr(struct mlx5_priv *priv)
978 {
979 	struct mlx5_ibv_shared *sh;
980 
981 	if (!priv->dr_shared)
982 		return;
983 	priv->dr_shared = 0;
984 	sh = priv->sh;
985 	MLX5_ASSERT(sh);
986 #ifdef HAVE_MLX5DV_DR
987 	MLX5_ASSERT(sh->dv_refcnt);
988 	if (sh->dv_refcnt && --sh->dv_refcnt)
989 		return;
990 	if (sh->rx_domain) {
991 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
992 		sh->rx_domain = NULL;
993 	}
994 	if (sh->tx_domain) {
995 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
996 		sh->tx_domain = NULL;
997 	}
998 #ifdef HAVE_MLX5DV_DR_ESWITCH
999 	if (sh->fdb_domain) {
1000 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1001 		sh->fdb_domain = NULL;
1002 	}
1003 	if (sh->esw_drop_action) {
1004 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1005 		sh->esw_drop_action = NULL;
1006 	}
1007 #endif
1008 	if (sh->pop_vlan_action) {
1009 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1010 		sh->pop_vlan_action = NULL;
1011 	}
1012 	pthread_mutex_destroy(&sh->dv_mutex);
1013 #endif /* HAVE_MLX5DV_DR */
1014 	if (sh->tag_table) {
1015 		/* tags should be destroyed with flow before. */
1016 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1017 		sh->tag_table = NULL;
1018 	}
1019 	mlx5_free_table_hash_list(priv);
1020 }
1021 
1022 /**
1023  * Initialize shared data between primary and secondary process.
1024  *
1025  * A memzone is reserved by primary process and secondary processes attach to
1026  * the memzone.
1027  *
1028  * @return
1029  *   0 on success, a negative errno value otherwise and rte_errno is set.
1030  */
1031 static int
1032 mlx5_init_shared_data(void)
1033 {
1034 	const struct rte_memzone *mz;
1035 	int ret = 0;
1036 
1037 	rte_spinlock_lock(&mlx5_shared_data_lock);
1038 	if (mlx5_shared_data == NULL) {
1039 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1040 			/* Allocate shared memory. */
1041 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1042 						 sizeof(*mlx5_shared_data),
1043 						 SOCKET_ID_ANY, 0);
1044 			if (mz == NULL) {
1045 				DRV_LOG(ERR,
1046 					"Cannot allocate mlx5 shared data");
1047 				ret = -rte_errno;
1048 				goto error;
1049 			}
1050 			mlx5_shared_data = mz->addr;
1051 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1052 			rte_spinlock_init(&mlx5_shared_data->lock);
1053 		} else {
1054 			/* Lookup allocated shared memory. */
1055 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1056 			if (mz == NULL) {
1057 				DRV_LOG(ERR,
1058 					"Cannot attach mlx5 shared data");
1059 				ret = -rte_errno;
1060 				goto error;
1061 			}
1062 			mlx5_shared_data = mz->addr;
1063 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1064 		}
1065 	}
1066 error:
1067 	rte_spinlock_unlock(&mlx5_shared_data_lock);
1068 	return ret;
1069 }
1070 
1071 /**
1072  * Retrieve integer value from environment variable.
1073  *
1074  * @param[in] name
1075  *   Environment variable name.
1076  *
1077  * @return
1078  *   Integer value, 0 if the variable is not set.
1079  */
1080 int
1081 mlx5_getenv_int(const char *name)
1082 {
1083 	const char *val = getenv(name);
1084 
1085 	if (val == NULL)
1086 		return 0;
1087 	return atoi(val);
1088 }
1089 
1090 /**
1091  * Verbs callback to allocate a memory. This function should allocate the space
1092  * according to the size provided residing inside a huge page.
1093  * Please note that all allocation must respect the alignment from libmlx5
1094  * (i.e. currently sysconf(_SC_PAGESIZE)).
1095  *
1096  * @param[in] size
1097  *   The size in bytes of the memory to allocate.
1098  * @param[in] data
1099  *   A pointer to the callback data.
1100  *
1101  * @return
1102  *   Allocated buffer, NULL otherwise and rte_errno is set.
1103  */
1104 static void *
1105 mlx5_alloc_verbs_buf(size_t size, void *data)
1106 {
1107 	struct mlx5_priv *priv = data;
1108 	void *ret;
1109 	size_t alignment = sysconf(_SC_PAGESIZE);
1110 	unsigned int socket = SOCKET_ID_ANY;
1111 
1112 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1113 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1114 
1115 		socket = ctrl->socket;
1116 	} else if (priv->verbs_alloc_ctx.type ==
1117 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1118 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1119 
1120 		socket = ctrl->socket;
1121 	}
1122 	MLX5_ASSERT(data != NULL);
1123 	ret = rte_malloc_socket(__func__, size, alignment, socket);
1124 	if (!ret && size)
1125 		rte_errno = ENOMEM;
1126 	return ret;
1127 }
1128 
1129 /**
1130  * Verbs callback to free a memory.
1131  *
1132  * @param[in] ptr
1133  *   A pointer to the memory to free.
1134  * @param[in] data
1135  *   A pointer to the callback data.
1136  */
1137 static void
1138 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1139 {
1140 	MLX5_ASSERT(data != NULL);
1141 	rte_free(ptr);
1142 }
1143 
1144 /**
1145  * DPDK callback to add udp tunnel port
1146  *
1147  * @param[in] dev
1148  *   A pointer to eth_dev
1149  * @param[in] udp_tunnel
1150  *   A pointer to udp tunnel
1151  *
1152  * @return
1153  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1154  */
1155 int
1156 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1157 			 struct rte_eth_udp_tunnel *udp_tunnel)
1158 {
1159 	MLX5_ASSERT(udp_tunnel != NULL);
1160 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1161 	    udp_tunnel->udp_port == 4789)
1162 		return 0;
1163 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1164 	    udp_tunnel->udp_port == 4790)
1165 		return 0;
1166 	return -ENOTSUP;
1167 }
1168 
1169 /**
1170  * Initialize process private data structure.
1171  *
1172  * @param dev
1173  *   Pointer to Ethernet device structure.
1174  *
1175  * @return
1176  *   0 on success, a negative errno value otherwise and rte_errno is set.
1177  */
1178 int
1179 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1180 {
1181 	struct mlx5_priv *priv = dev->data->dev_private;
1182 	struct mlx5_proc_priv *ppriv;
1183 	size_t ppriv_size;
1184 
1185 	/*
1186 	 * UAR register table follows the process private structure. BlueFlame
1187 	 * registers for Tx queues are stored in the table.
1188 	 */
1189 	ppriv_size =
1190 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1191 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1192 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1193 	if (!ppriv) {
1194 		rte_errno = ENOMEM;
1195 		return -rte_errno;
1196 	}
1197 	ppriv->uar_table_sz = ppriv_size;
1198 	dev->process_private = ppriv;
1199 	return 0;
1200 }
1201 
1202 /**
1203  * Un-initialize process private data structure.
1204  *
1205  * @param dev
1206  *   Pointer to Ethernet device structure.
1207  */
1208 static void
1209 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1210 {
1211 	if (!dev->process_private)
1212 		return;
1213 	rte_free(dev->process_private);
1214 	dev->process_private = NULL;
1215 }
1216 
1217 /**
1218  * DPDK callback to close the device.
1219  *
1220  * Destroy all queues and objects, free memory.
1221  *
1222  * @param dev
1223  *   Pointer to Ethernet device structure.
1224  */
1225 static void
1226 mlx5_dev_close(struct rte_eth_dev *dev)
1227 {
1228 	struct mlx5_priv *priv = dev->data->dev_private;
1229 	unsigned int i;
1230 	int ret;
1231 
1232 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1233 		dev->data->port_id,
1234 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1235 	/* In case mlx5_dev_stop() has not been called. */
1236 	mlx5_dev_interrupt_handler_uninstall(dev);
1237 	mlx5_dev_interrupt_handler_devx_uninstall(dev);
1238 	mlx5_traffic_disable(dev);
1239 	mlx5_flow_flush(dev, NULL);
1240 	mlx5_flow_meter_flush(dev, NULL);
1241 	/* Prevent crashes when queues are still in use. */
1242 	dev->rx_pkt_burst = removed_rx_burst;
1243 	dev->tx_pkt_burst = removed_tx_burst;
1244 	rte_wmb();
1245 	/* Disable datapath on secondary process. */
1246 	mlx5_mp_req_stop_rxtx(dev);
1247 	if (priv->rxqs != NULL) {
1248 		/* XXX race condition if mlx5_rx_burst() is still running. */
1249 		usleep(1000);
1250 		for (i = 0; (i != priv->rxqs_n); ++i)
1251 			mlx5_rxq_release(dev, i);
1252 		priv->rxqs_n = 0;
1253 		priv->rxqs = NULL;
1254 	}
1255 	if (priv->txqs != NULL) {
1256 		/* XXX race condition if mlx5_tx_burst() is still running. */
1257 		usleep(1000);
1258 		for (i = 0; (i != priv->txqs_n); ++i)
1259 			mlx5_txq_release(dev, i);
1260 		priv->txqs_n = 0;
1261 		priv->txqs = NULL;
1262 	}
1263 	mlx5_proc_priv_uninit(dev);
1264 	if (priv->mreg_cp_tbl)
1265 		mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1266 	mlx5_mprq_free_mp(dev);
1267 	mlx5_free_shared_dr(priv);
1268 	if (priv->rss_conf.rss_key != NULL)
1269 		rte_free(priv->rss_conf.rss_key);
1270 	if (priv->reta_idx != NULL)
1271 		rte_free(priv->reta_idx);
1272 	if (priv->config.vf)
1273 		mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1274 				       dev->data->mac_addrs,
1275 				       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1276 	if (priv->nl_socket_route >= 0)
1277 		close(priv->nl_socket_route);
1278 	if (priv->nl_socket_rdma >= 0)
1279 		close(priv->nl_socket_rdma);
1280 	if (priv->vmwa_context)
1281 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1282 	if (priv->sh) {
1283 		/*
1284 		 * Free the shared context in last turn, because the cleanup
1285 		 * routines above may use some shared fields, like
1286 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1287 		 * ifindex if Netlink fails.
1288 		 */
1289 		mlx5_free_shared_ibctx(priv->sh);
1290 		priv->sh = NULL;
1291 	}
1292 	ret = mlx5_hrxq_verify(dev);
1293 	if (ret)
1294 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1295 			dev->data->port_id);
1296 	ret = mlx5_ind_table_obj_verify(dev);
1297 	if (ret)
1298 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1299 			dev->data->port_id);
1300 	ret = mlx5_rxq_obj_verify(dev);
1301 	if (ret)
1302 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1303 			dev->data->port_id);
1304 	ret = mlx5_rxq_verify(dev);
1305 	if (ret)
1306 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1307 			dev->data->port_id);
1308 	ret = mlx5_txq_obj_verify(dev);
1309 	if (ret)
1310 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1311 			dev->data->port_id);
1312 	ret = mlx5_txq_verify(dev);
1313 	if (ret)
1314 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1315 			dev->data->port_id);
1316 	ret = mlx5_flow_verify(dev);
1317 	if (ret)
1318 		DRV_LOG(WARNING, "port %u some flows still remain",
1319 			dev->data->port_id);
1320 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1321 		unsigned int c = 0;
1322 		uint16_t port_id;
1323 
1324 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1325 			struct mlx5_priv *opriv =
1326 				rte_eth_devices[port_id].data->dev_private;
1327 
1328 			if (!opriv ||
1329 			    opriv->domain_id != priv->domain_id ||
1330 			    &rte_eth_devices[port_id] == dev)
1331 				continue;
1332 			++c;
1333 			break;
1334 		}
1335 		if (!c)
1336 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1337 	}
1338 	memset(priv, 0, sizeof(*priv));
1339 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1340 	/*
1341 	 * Reset mac_addrs to NULL such that it is not freed as part of
1342 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1343 	 * it is freed when dev_private is freed.
1344 	 */
1345 	dev->data->mac_addrs = NULL;
1346 }
1347 
1348 const struct eth_dev_ops mlx5_dev_ops = {
1349 	.dev_configure = mlx5_dev_configure,
1350 	.dev_start = mlx5_dev_start,
1351 	.dev_stop = mlx5_dev_stop,
1352 	.dev_set_link_down = mlx5_set_link_down,
1353 	.dev_set_link_up = mlx5_set_link_up,
1354 	.dev_close = mlx5_dev_close,
1355 	.promiscuous_enable = mlx5_promiscuous_enable,
1356 	.promiscuous_disable = mlx5_promiscuous_disable,
1357 	.allmulticast_enable = mlx5_allmulticast_enable,
1358 	.allmulticast_disable = mlx5_allmulticast_disable,
1359 	.link_update = mlx5_link_update,
1360 	.stats_get = mlx5_stats_get,
1361 	.stats_reset = mlx5_stats_reset,
1362 	.xstats_get = mlx5_xstats_get,
1363 	.xstats_reset = mlx5_xstats_reset,
1364 	.xstats_get_names = mlx5_xstats_get_names,
1365 	.fw_version_get = mlx5_fw_version_get,
1366 	.dev_infos_get = mlx5_dev_infos_get,
1367 	.read_clock = mlx5_read_clock,
1368 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1369 	.vlan_filter_set = mlx5_vlan_filter_set,
1370 	.rx_queue_setup = mlx5_rx_queue_setup,
1371 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1372 	.tx_queue_setup = mlx5_tx_queue_setup,
1373 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1374 	.rx_queue_release = mlx5_rx_queue_release,
1375 	.tx_queue_release = mlx5_tx_queue_release,
1376 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1377 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1378 	.mac_addr_remove = mlx5_mac_addr_remove,
1379 	.mac_addr_add = mlx5_mac_addr_add,
1380 	.mac_addr_set = mlx5_mac_addr_set,
1381 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1382 	.mtu_set = mlx5_dev_set_mtu,
1383 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1384 	.vlan_offload_set = mlx5_vlan_offload_set,
1385 	.reta_update = mlx5_dev_rss_reta_update,
1386 	.reta_query = mlx5_dev_rss_reta_query,
1387 	.rss_hash_update = mlx5_rss_hash_update,
1388 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1389 	.filter_ctrl = mlx5_dev_filter_ctrl,
1390 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1391 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1392 	.rxq_info_get = mlx5_rxq_info_get,
1393 	.txq_info_get = mlx5_txq_info_get,
1394 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1395 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1396 	.rx_queue_count = mlx5_rx_queue_count,
1397 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1398 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1399 	.is_removed = mlx5_is_removed,
1400 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1401 	.get_module_info = mlx5_get_module_info,
1402 	.get_module_eeprom = mlx5_get_module_eeprom,
1403 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1404 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1405 };
1406 
1407 /* Available operations from secondary process. */
1408 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1409 	.stats_get = mlx5_stats_get,
1410 	.stats_reset = mlx5_stats_reset,
1411 	.xstats_get = mlx5_xstats_get,
1412 	.xstats_reset = mlx5_xstats_reset,
1413 	.xstats_get_names = mlx5_xstats_get_names,
1414 	.fw_version_get = mlx5_fw_version_get,
1415 	.dev_infos_get = mlx5_dev_infos_get,
1416 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1417 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1418 	.rxq_info_get = mlx5_rxq_info_get,
1419 	.txq_info_get = mlx5_txq_info_get,
1420 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1421 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1422 	.get_module_info = mlx5_get_module_info,
1423 	.get_module_eeprom = mlx5_get_module_eeprom,
1424 };
1425 
1426 /* Available operations in flow isolated mode. */
1427 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1428 	.dev_configure = mlx5_dev_configure,
1429 	.dev_start = mlx5_dev_start,
1430 	.dev_stop = mlx5_dev_stop,
1431 	.dev_set_link_down = mlx5_set_link_down,
1432 	.dev_set_link_up = mlx5_set_link_up,
1433 	.dev_close = mlx5_dev_close,
1434 	.promiscuous_enable = mlx5_promiscuous_enable,
1435 	.promiscuous_disable = mlx5_promiscuous_disable,
1436 	.allmulticast_enable = mlx5_allmulticast_enable,
1437 	.allmulticast_disable = mlx5_allmulticast_disable,
1438 	.link_update = mlx5_link_update,
1439 	.stats_get = mlx5_stats_get,
1440 	.stats_reset = mlx5_stats_reset,
1441 	.xstats_get = mlx5_xstats_get,
1442 	.xstats_reset = mlx5_xstats_reset,
1443 	.xstats_get_names = mlx5_xstats_get_names,
1444 	.fw_version_get = mlx5_fw_version_get,
1445 	.dev_infos_get = mlx5_dev_infos_get,
1446 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1447 	.vlan_filter_set = mlx5_vlan_filter_set,
1448 	.rx_queue_setup = mlx5_rx_queue_setup,
1449 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1450 	.tx_queue_setup = mlx5_tx_queue_setup,
1451 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1452 	.rx_queue_release = mlx5_rx_queue_release,
1453 	.tx_queue_release = mlx5_tx_queue_release,
1454 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1455 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1456 	.mac_addr_remove = mlx5_mac_addr_remove,
1457 	.mac_addr_add = mlx5_mac_addr_add,
1458 	.mac_addr_set = mlx5_mac_addr_set,
1459 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1460 	.mtu_set = mlx5_dev_set_mtu,
1461 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1462 	.vlan_offload_set = mlx5_vlan_offload_set,
1463 	.filter_ctrl = mlx5_dev_filter_ctrl,
1464 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1465 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1466 	.rxq_info_get = mlx5_rxq_info_get,
1467 	.txq_info_get = mlx5_txq_info_get,
1468 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1469 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1470 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1471 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1472 	.is_removed = mlx5_is_removed,
1473 	.get_module_info = mlx5_get_module_info,
1474 	.get_module_eeprom = mlx5_get_module_eeprom,
1475 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1476 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1477 };
1478 
1479 /**
1480  * Verify and store value for device argument.
1481  *
1482  * @param[in] key
1483  *   Key argument to verify.
1484  * @param[in] val
1485  *   Value associated with key.
1486  * @param opaque
1487  *   User data.
1488  *
1489  * @return
1490  *   0 on success, a negative errno value otherwise and rte_errno is set.
1491  */
1492 static int
1493 mlx5_args_check(const char *key, const char *val, void *opaque)
1494 {
1495 	struct mlx5_dev_config *config = opaque;
1496 	unsigned long tmp;
1497 
1498 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1499 	if (!strcmp(MLX5_REPRESENTOR, key))
1500 		return 0;
1501 	errno = 0;
1502 	tmp = strtoul(val, NULL, 0);
1503 	if (errno) {
1504 		rte_errno = errno;
1505 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1506 		return -rte_errno;
1507 	}
1508 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1509 		config->cqe_comp = !!tmp;
1510 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1511 		config->cqe_pad = !!tmp;
1512 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1513 		config->hw_padding = !!tmp;
1514 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1515 		config->mprq.enabled = !!tmp;
1516 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1517 		config->mprq.stride_num_n = tmp;
1518 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1519 		config->mprq.max_memcpy_len = tmp;
1520 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1521 		config->mprq.min_rxqs_num = tmp;
1522 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1523 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1524 				 " converted to txq_inline_max", key);
1525 		config->txq_inline_max = tmp;
1526 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1527 		config->txq_inline_max = tmp;
1528 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1529 		config->txq_inline_min = tmp;
1530 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1531 		config->txq_inline_mpw = tmp;
1532 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1533 		config->txqs_inline = tmp;
1534 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1535 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1536 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1537 		config->mps = !!tmp;
1538 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1539 		if (tmp != MLX5_TXDB_CACHED &&
1540 		    tmp != MLX5_TXDB_NCACHED &&
1541 		    tmp != MLX5_TXDB_HEURISTIC) {
1542 			DRV_LOG(ERR, "invalid Tx doorbell "
1543 				     "mapping parameter");
1544 			rte_errno = EINVAL;
1545 			return -rte_errno;
1546 		}
1547 		config->dbnc = tmp;
1548 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1549 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1550 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1551 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1552 				 " converted to txq_inline_mpw", key);
1553 		config->txq_inline_mpw = tmp;
1554 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1555 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1556 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1557 		config->rx_vec_en = !!tmp;
1558 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1559 		config->l3_vxlan_en = !!tmp;
1560 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1561 		config->vf_nl_en = !!tmp;
1562 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1563 		config->dv_esw_en = !!tmp;
1564 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1565 		config->dv_flow_en = !!tmp;
1566 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1567 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1568 		    tmp != MLX5_XMETA_MODE_META16 &&
1569 		    tmp != MLX5_XMETA_MODE_META32) {
1570 			DRV_LOG(ERR, "invalid extensive "
1571 				     "metadata parameter");
1572 			rte_errno = EINVAL;
1573 			return -rte_errno;
1574 		}
1575 		config->dv_xmeta_en = tmp;
1576 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1577 		config->mr_ext_memseg_en = !!tmp;
1578 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1579 		config->max_dump_files_num = tmp;
1580 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1581 		config->lro.timeout = tmp;
1582 	} else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1583 		DRV_LOG(DEBUG, "class argument is %s.", val);
1584 	} else {
1585 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1586 		rte_errno = EINVAL;
1587 		return -rte_errno;
1588 	}
1589 	return 0;
1590 }
1591 
1592 /**
1593  * Parse device parameters.
1594  *
1595  * @param config
1596  *   Pointer to device configuration structure.
1597  * @param devargs
1598  *   Device arguments structure.
1599  *
1600  * @return
1601  *   0 on success, a negative errno value otherwise and rte_errno is set.
1602  */
1603 static int
1604 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1605 {
1606 	const char **params = (const char *[]){
1607 		MLX5_RXQ_CQE_COMP_EN,
1608 		MLX5_RXQ_CQE_PAD_EN,
1609 		MLX5_RXQ_PKT_PAD_EN,
1610 		MLX5_RX_MPRQ_EN,
1611 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1612 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1613 		MLX5_RXQS_MIN_MPRQ,
1614 		MLX5_TXQ_INLINE,
1615 		MLX5_TXQ_INLINE_MIN,
1616 		MLX5_TXQ_INLINE_MAX,
1617 		MLX5_TXQ_INLINE_MPW,
1618 		MLX5_TXQS_MIN_INLINE,
1619 		MLX5_TXQS_MAX_VEC,
1620 		MLX5_TXQ_MPW_EN,
1621 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1622 		MLX5_TXQ_MAX_INLINE_LEN,
1623 		MLX5_TX_DB_NC,
1624 		MLX5_TX_VEC_EN,
1625 		MLX5_RX_VEC_EN,
1626 		MLX5_L3_VXLAN_EN,
1627 		MLX5_VF_NL_EN,
1628 		MLX5_DV_ESW_EN,
1629 		MLX5_DV_FLOW_EN,
1630 		MLX5_DV_XMETA_EN,
1631 		MLX5_MR_EXT_MEMSEG_EN,
1632 		MLX5_REPRESENTOR,
1633 		MLX5_MAX_DUMP_FILES_NUM,
1634 		MLX5_LRO_TIMEOUT_USEC,
1635 		MLX5_CLASS_ARG_NAME,
1636 		NULL,
1637 	};
1638 	struct rte_kvargs *kvlist;
1639 	int ret = 0;
1640 	int i;
1641 
1642 	if (devargs == NULL)
1643 		return 0;
1644 	/* Following UGLY cast is done to pass checkpatch. */
1645 	kvlist = rte_kvargs_parse(devargs->args, params);
1646 	if (kvlist == NULL) {
1647 		rte_errno = EINVAL;
1648 		return -rte_errno;
1649 	}
1650 	/* Process parameters. */
1651 	for (i = 0; (params[i] != NULL); ++i) {
1652 		if (rte_kvargs_count(kvlist, params[i])) {
1653 			ret = rte_kvargs_process(kvlist, params[i],
1654 						 mlx5_args_check, config);
1655 			if (ret) {
1656 				rte_errno = EINVAL;
1657 				rte_kvargs_free(kvlist);
1658 				return -rte_errno;
1659 			}
1660 		}
1661 	}
1662 	rte_kvargs_free(kvlist);
1663 	return 0;
1664 }
1665 
1666 static struct rte_pci_driver mlx5_driver;
1667 
1668 /**
1669  * PMD global initialization.
1670  *
1671  * Independent from individual device, this function initializes global
1672  * per-PMD data structures distinguishing primary and secondary processes.
1673  * Hence, each initialization is called once per a process.
1674  *
1675  * @return
1676  *   0 on success, a negative errno value otherwise and rte_errno is set.
1677  */
1678 static int
1679 mlx5_init_once(void)
1680 {
1681 	struct mlx5_shared_data *sd;
1682 	struct mlx5_local_data *ld = &mlx5_local_data;
1683 	int ret = 0;
1684 
1685 	if (mlx5_init_shared_data())
1686 		return -rte_errno;
1687 	sd = mlx5_shared_data;
1688 	MLX5_ASSERT(sd);
1689 	rte_spinlock_lock(&sd->lock);
1690 	switch (rte_eal_process_type()) {
1691 	case RTE_PROC_PRIMARY:
1692 		if (sd->init_done)
1693 			break;
1694 		LIST_INIT(&sd->mem_event_cb_list);
1695 		rte_rwlock_init(&sd->mem_event_rwlock);
1696 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1697 						mlx5_mr_mem_event_cb, NULL);
1698 		ret = mlx5_mp_init_primary();
1699 		if (ret)
1700 			goto out;
1701 		sd->init_done = true;
1702 		break;
1703 	case RTE_PROC_SECONDARY:
1704 		if (ld->init_done)
1705 			break;
1706 		ret = mlx5_mp_init_secondary();
1707 		if (ret)
1708 			goto out;
1709 		++sd->secondary_cnt;
1710 		ld->init_done = true;
1711 		break;
1712 	default:
1713 		break;
1714 	}
1715 out:
1716 	rte_spinlock_unlock(&sd->lock);
1717 	return ret;
1718 }
1719 
1720 /**
1721  * Configures the minimal amount of data to inline into WQE
1722  * while sending packets.
1723  *
1724  * - the txq_inline_min has the maximal priority, if this
1725  *   key is specified in devargs
1726  * - if DevX is enabled the inline mode is queried from the
1727  *   device (HCA attributes and NIC vport context if needed).
1728  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1729  *   and none (0 bytes) for other NICs
1730  *
1731  * @param spawn
1732  *   Verbs device parameters (name, port, switch_info) to spawn.
1733  * @param config
1734  *   Device configuration parameters.
1735  */
1736 static void
1737 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1738 		    struct mlx5_dev_config *config)
1739 {
1740 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1741 		/* Application defines size of inlined data explicitly. */
1742 		switch (spawn->pci_dev->id.device_id) {
1743 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1744 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1745 			if (config->txq_inline_min <
1746 				       (int)MLX5_INLINE_HSIZE_L2) {
1747 				DRV_LOG(DEBUG,
1748 					"txq_inline_mix aligned to minimal"
1749 					" ConnectX-4 required value %d",
1750 					(int)MLX5_INLINE_HSIZE_L2);
1751 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1752 			}
1753 			break;
1754 		}
1755 		goto exit;
1756 	}
1757 	if (config->hca_attr.eth_net_offloads) {
1758 		/* We have DevX enabled, inline mode queried successfully. */
1759 		switch (config->hca_attr.wqe_inline_mode) {
1760 		case MLX5_CAP_INLINE_MODE_L2:
1761 			/* outer L2 header must be inlined. */
1762 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1763 			goto exit;
1764 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1765 			/* No inline data are required by NIC. */
1766 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1767 			config->hw_vlan_insert =
1768 				config->hca_attr.wqe_vlan_insert;
1769 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1770 			goto exit;
1771 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1772 			/* inline mode is defined by NIC vport context. */
1773 			if (!config->hca_attr.eth_virt)
1774 				break;
1775 			switch (config->hca_attr.vport_inline_mode) {
1776 			case MLX5_INLINE_MODE_NONE:
1777 				config->txq_inline_min =
1778 					MLX5_INLINE_HSIZE_NONE;
1779 				goto exit;
1780 			case MLX5_INLINE_MODE_L2:
1781 				config->txq_inline_min =
1782 					MLX5_INLINE_HSIZE_L2;
1783 				goto exit;
1784 			case MLX5_INLINE_MODE_IP:
1785 				config->txq_inline_min =
1786 					MLX5_INLINE_HSIZE_L3;
1787 				goto exit;
1788 			case MLX5_INLINE_MODE_TCP_UDP:
1789 				config->txq_inline_min =
1790 					MLX5_INLINE_HSIZE_L4;
1791 				goto exit;
1792 			case MLX5_INLINE_MODE_INNER_L2:
1793 				config->txq_inline_min =
1794 					MLX5_INLINE_HSIZE_INNER_L2;
1795 				goto exit;
1796 			case MLX5_INLINE_MODE_INNER_IP:
1797 				config->txq_inline_min =
1798 					MLX5_INLINE_HSIZE_INNER_L3;
1799 				goto exit;
1800 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1801 				config->txq_inline_min =
1802 					MLX5_INLINE_HSIZE_INNER_L4;
1803 				goto exit;
1804 			}
1805 		}
1806 	}
1807 	/*
1808 	 * We get here if we are unable to deduce
1809 	 * inline data size with DevX. Try PCI ID
1810 	 * to determine old NICs.
1811 	 */
1812 	switch (spawn->pci_dev->id.device_id) {
1813 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1814 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1815 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1816 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1817 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1818 		config->hw_vlan_insert = 0;
1819 		break;
1820 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1821 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1822 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1823 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1824 		/*
1825 		 * These NICs support VLAN insertion from WQE and
1826 		 * report the wqe_vlan_insert flag. But there is the bug
1827 		 * and PFC control may be broken, so disable feature.
1828 		 */
1829 		config->hw_vlan_insert = 0;
1830 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1831 		break;
1832 	default:
1833 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1834 		break;
1835 	}
1836 exit:
1837 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1838 }
1839 
1840 /**
1841  * Configures the metadata mask fields in the shared context.
1842  *
1843  * @param [in] dev
1844  *   Pointer to Ethernet device.
1845  */
1846 static void
1847 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1848 {
1849 	struct mlx5_priv *priv = dev->data->dev_private;
1850 	struct mlx5_ibv_shared *sh = priv->sh;
1851 	uint32_t meta, mark, reg_c0;
1852 
1853 	reg_c0 = ~priv->vport_meta_mask;
1854 	switch (priv->config.dv_xmeta_en) {
1855 	case MLX5_XMETA_MODE_LEGACY:
1856 		meta = UINT32_MAX;
1857 		mark = MLX5_FLOW_MARK_MASK;
1858 		break;
1859 	case MLX5_XMETA_MODE_META16:
1860 		meta = reg_c0 >> rte_bsf32(reg_c0);
1861 		mark = MLX5_FLOW_MARK_MASK;
1862 		break;
1863 	case MLX5_XMETA_MODE_META32:
1864 		meta = UINT32_MAX;
1865 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1866 		break;
1867 	default:
1868 		meta = 0;
1869 		mark = 0;
1870 		MLX5_ASSERT(false);
1871 		break;
1872 	}
1873 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1874 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1875 				 sh->dv_mark_mask, mark);
1876 	else
1877 		sh->dv_mark_mask = mark;
1878 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1879 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1880 				 sh->dv_meta_mask, meta);
1881 	else
1882 		sh->dv_meta_mask = meta;
1883 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1884 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1885 				 sh->dv_meta_mask, reg_c0);
1886 	else
1887 		sh->dv_regc0_mask = reg_c0;
1888 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1889 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1890 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1891 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1892 }
1893 
1894 /**
1895  * Allocate page of door-bells and register it using DevX API.
1896  *
1897  * @param [in] dev
1898  *   Pointer to Ethernet device.
1899  *
1900  * @return
1901  *   Pointer to new page on success, NULL otherwise.
1902  */
1903 static struct mlx5_devx_dbr_page *
1904 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1905 {
1906 	struct mlx5_priv *priv = dev->data->dev_private;
1907 	struct mlx5_devx_dbr_page *page;
1908 
1909 	/* Allocate space for door-bell page and management data. */
1910 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1911 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1912 	if (!page) {
1913 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1914 			dev->data->port_id);
1915 		return NULL;
1916 	}
1917 	/* Register allocated memory. */
1918 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1919 					      MLX5_DBR_PAGE_SIZE, 0);
1920 	if (!page->umem) {
1921 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1922 			dev->data->port_id);
1923 		rte_free(page);
1924 		return NULL;
1925 	}
1926 	return page;
1927 }
1928 
1929 /**
1930  * Find the next available door-bell, allocate new page if needed.
1931  *
1932  * @param [in] dev
1933  *   Pointer to Ethernet device.
1934  * @param [out] dbr_page
1935  *   Door-bell page containing the page data.
1936  *
1937  * @return
1938  *   Door-bell address offset on success, a negative error value otherwise.
1939  */
1940 int64_t
1941 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1942 {
1943 	struct mlx5_priv *priv = dev->data->dev_private;
1944 	struct mlx5_devx_dbr_page *page = NULL;
1945 	uint32_t i, j;
1946 
1947 	LIST_FOREACH(page, &priv->dbrpgs, next)
1948 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1949 			break;
1950 	if (!page) { /* No page with free door-bell exists. */
1951 		page = mlx5_alloc_dbr_page(dev);
1952 		if (!page) /* Failed to allocate new page. */
1953 			return (-1);
1954 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1955 	}
1956 	/* Loop to find bitmap part with clear bit. */
1957 	for (i = 0;
1958 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1959 	     i++)
1960 		; /* Empty. */
1961 	/* Find the first clear bit. */
1962 	j = rte_bsf64(~page->dbr_bitmap[i]);
1963 	MLX5_ASSERT(i < (MLX5_DBR_PER_PAGE / 64));
1964 	page->dbr_bitmap[i] |= (1 << j);
1965 	page->dbr_count++;
1966 	*dbr_page = page;
1967 	return (((i * 64) + j) * sizeof(uint64_t));
1968 }
1969 
1970 /**
1971  * Release a door-bell record.
1972  *
1973  * @param [in] dev
1974  *   Pointer to Ethernet device.
1975  * @param [in] umem_id
1976  *   UMEM ID of page containing the door-bell record to release.
1977  * @param [in] offset
1978  *   Offset of door-bell record in page.
1979  *
1980  * @return
1981  *   0 on success, a negative error value otherwise.
1982  */
1983 int32_t
1984 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1985 {
1986 	struct mlx5_priv *priv = dev->data->dev_private;
1987 	struct mlx5_devx_dbr_page *page = NULL;
1988 	int ret = 0;
1989 
1990 	LIST_FOREACH(page, &priv->dbrpgs, next)
1991 		/* Find the page this address belongs to. */
1992 		if (page->umem->umem_id == umem_id)
1993 			break;
1994 	if (!page)
1995 		return -EINVAL;
1996 	page->dbr_count--;
1997 	if (!page->dbr_count) {
1998 		/* Page not used, free it and remove from list. */
1999 		LIST_REMOVE(page, next);
2000 		if (page->umem)
2001 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
2002 		rte_free(page);
2003 	} else {
2004 		/* Mark in bitmap that this door-bell is not in use. */
2005 		offset /= MLX5_DBR_SIZE;
2006 		int i = offset / 64;
2007 		int j = offset % 64;
2008 
2009 		page->dbr_bitmap[i] &= ~(1 << j);
2010 	}
2011 	return ret;
2012 }
2013 
2014 int
2015 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2016 {
2017 	static const char *const dynf_names[] = {
2018 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2019 		RTE_MBUF_DYNFLAG_METADATA_NAME
2020 	};
2021 	unsigned int i;
2022 
2023 	if (n < RTE_DIM(dynf_names))
2024 		return -ENOMEM;
2025 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2026 		if (names[i] == NULL)
2027 			return -EINVAL;
2028 		strcpy(names[i], dynf_names[i]);
2029 	}
2030 	return RTE_DIM(dynf_names);
2031 }
2032 
2033 /**
2034  * Check sibling device configurations.
2035  *
2036  * Sibling devices sharing the Infiniband device context
2037  * should have compatible configurations. This regards
2038  * representors and bonding slaves.
2039  *
2040  * @param priv
2041  *   Private device descriptor.
2042  * @param config
2043  *   Configuration of the device is going to be created.
2044  *
2045  * @return
2046  *   0 on success, EINVAL otherwise
2047  */
2048 static int
2049 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2050 			      struct mlx5_dev_config *config)
2051 {
2052 	struct mlx5_ibv_shared *sh = priv->sh;
2053 	struct mlx5_dev_config *sh_conf = NULL;
2054 	uint16_t port_id;
2055 
2056 	MLX5_ASSERT(sh);
2057 	/* Nothing to compare for the single/first device. */
2058 	if (sh->refcnt == 1)
2059 		return 0;
2060 	/* Find the device with shared context. */
2061 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2062 		struct mlx5_priv *opriv =
2063 			rte_eth_devices[port_id].data->dev_private;
2064 
2065 		if (opriv && opriv != priv && opriv->sh == sh) {
2066 			sh_conf = &opriv->config;
2067 			break;
2068 		}
2069 	}
2070 	if (!sh_conf)
2071 		return 0;
2072 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2073 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2074 			     " for shared %s context", sh->ibdev_name);
2075 		rte_errno = EINVAL;
2076 		return rte_errno;
2077 	}
2078 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2079 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2080 			     " for shared %s context", sh->ibdev_name);
2081 		rte_errno = EINVAL;
2082 		return rte_errno;
2083 	}
2084 	return 0;
2085 }
2086 /**
2087  * Spawn an Ethernet device from Verbs information.
2088  *
2089  * @param dpdk_dev
2090  *   Backing DPDK device.
2091  * @param spawn
2092  *   Verbs device parameters (name, port, switch_info) to spawn.
2093  * @param config
2094  *   Device configuration parameters.
2095  *
2096  * @return
2097  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2098  *   is set. The following errors are defined:
2099  *
2100  *   EBUSY: device is not supposed to be spawned.
2101  *   EEXIST: device is already spawned
2102  */
2103 static struct rte_eth_dev *
2104 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2105 	       struct mlx5_dev_spawn_data *spawn,
2106 	       struct mlx5_dev_config config)
2107 {
2108 	const struct mlx5_switch_info *switch_info = &spawn->info;
2109 	struct mlx5_ibv_shared *sh = NULL;
2110 	struct ibv_port_attr port_attr;
2111 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2112 	struct rte_eth_dev *eth_dev = NULL;
2113 	struct mlx5_priv *priv = NULL;
2114 	int err = 0;
2115 	unsigned int hw_padding = 0;
2116 	unsigned int mps;
2117 	unsigned int cqe_comp;
2118 	unsigned int cqe_pad = 0;
2119 	unsigned int tunnel_en = 0;
2120 	unsigned int mpls_en = 0;
2121 	unsigned int swp = 0;
2122 	unsigned int mprq = 0;
2123 	unsigned int mprq_min_stride_size_n = 0;
2124 	unsigned int mprq_max_stride_size_n = 0;
2125 	unsigned int mprq_min_stride_num_n = 0;
2126 	unsigned int mprq_max_stride_num_n = 0;
2127 	struct rte_ether_addr mac;
2128 	char name[RTE_ETH_NAME_MAX_LEN];
2129 	int own_domain_id = 0;
2130 	uint16_t port_id;
2131 	unsigned int i;
2132 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2133 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2134 #endif
2135 
2136 	/* Determine if this port representor is supposed to be spawned. */
2137 	if (switch_info->representor && dpdk_dev->devargs) {
2138 		struct rte_eth_devargs eth_da;
2139 
2140 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2141 		if (err) {
2142 			rte_errno = -err;
2143 			DRV_LOG(ERR, "failed to process device arguments: %s",
2144 				strerror(rte_errno));
2145 			return NULL;
2146 		}
2147 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
2148 			if (eth_da.representor_ports[i] ==
2149 			    (uint16_t)switch_info->port_name)
2150 				break;
2151 		if (i == eth_da.nb_representor_ports) {
2152 			rte_errno = EBUSY;
2153 			return NULL;
2154 		}
2155 	}
2156 	/* Build device name. */
2157 	if (spawn->pf_bond <  0) {
2158 		/* Single device. */
2159 		if (!switch_info->representor)
2160 			strlcpy(name, dpdk_dev->name, sizeof(name));
2161 		else
2162 			snprintf(name, sizeof(name), "%s_representor_%u",
2163 				 dpdk_dev->name, switch_info->port_name);
2164 	} else {
2165 		/* Bonding device. */
2166 		if (!switch_info->representor)
2167 			snprintf(name, sizeof(name), "%s_%s",
2168 				 dpdk_dev->name, spawn->ibv_dev->name);
2169 		else
2170 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
2171 				 dpdk_dev->name, spawn->ibv_dev->name,
2172 				 switch_info->port_name);
2173 	}
2174 	/* check if the device is already spawned */
2175 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2176 		rte_errno = EEXIST;
2177 		return NULL;
2178 	}
2179 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2180 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2181 		eth_dev = rte_eth_dev_attach_secondary(name);
2182 		if (eth_dev == NULL) {
2183 			DRV_LOG(ERR, "can not attach rte ethdev");
2184 			rte_errno = ENOMEM;
2185 			return NULL;
2186 		}
2187 		eth_dev->device = dpdk_dev;
2188 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
2189 		err = mlx5_proc_priv_init(eth_dev);
2190 		if (err)
2191 			return NULL;
2192 		/* Receive command fd from primary process */
2193 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2194 		if (err < 0)
2195 			return NULL;
2196 		/* Remap UAR for Tx queues. */
2197 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
2198 		if (err)
2199 			return NULL;
2200 		/*
2201 		 * Ethdev pointer is still required as input since
2202 		 * the primary device is not accessible from the
2203 		 * secondary process.
2204 		 */
2205 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2206 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2207 		return eth_dev;
2208 	}
2209 	/*
2210 	 * Some parameters ("tx_db_nc" in particularly) are needed in
2211 	 * advance to create dv/verbs device context. We proceed the
2212 	 * devargs here to get ones, and later proceed devargs again
2213 	 * to override some hardware settings.
2214 	 */
2215 	err = mlx5_args(&config, dpdk_dev->devargs);
2216 	if (err) {
2217 		err = rte_errno;
2218 		DRV_LOG(ERR, "failed to process device arguments: %s",
2219 			strerror(rte_errno));
2220 		goto error;
2221 	}
2222 	sh = mlx5_alloc_shared_ibctx(spawn, &config);
2223 	if (!sh)
2224 		return NULL;
2225 	config.devx = sh->devx;
2226 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2227 	config.dest_tir = 1;
2228 #endif
2229 #ifdef HAVE_IBV_MLX5_MOD_SWP
2230 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2231 #endif
2232 	/*
2233 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2234 	 * as all ConnectX-5 devices.
2235 	 */
2236 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2237 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2238 #endif
2239 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2240 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2241 #endif
2242 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2243 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2244 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2245 			DRV_LOG(DEBUG, "enhanced MPW is supported");
2246 			mps = MLX5_MPW_ENHANCED;
2247 		} else {
2248 			DRV_LOG(DEBUG, "MPW is supported");
2249 			mps = MLX5_MPW;
2250 		}
2251 	} else {
2252 		DRV_LOG(DEBUG, "MPW isn't supported");
2253 		mps = MLX5_MPW_DISABLED;
2254 	}
2255 #ifdef HAVE_IBV_MLX5_MOD_SWP
2256 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2257 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2258 	DRV_LOG(DEBUG, "SWP support: %u", swp);
2259 #endif
2260 	config.swp = !!swp;
2261 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2262 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2263 		struct mlx5dv_striding_rq_caps mprq_caps =
2264 			dv_attr.striding_rq_caps;
2265 
2266 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2267 			mprq_caps.min_single_stride_log_num_of_bytes);
2268 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2269 			mprq_caps.max_single_stride_log_num_of_bytes);
2270 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2271 			mprq_caps.min_single_wqe_log_num_of_strides);
2272 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2273 			mprq_caps.max_single_wqe_log_num_of_strides);
2274 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2275 			mprq_caps.supported_qpts);
2276 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2277 		mprq = 1;
2278 		mprq_min_stride_size_n =
2279 			mprq_caps.min_single_stride_log_num_of_bytes;
2280 		mprq_max_stride_size_n =
2281 			mprq_caps.max_single_stride_log_num_of_bytes;
2282 		mprq_min_stride_num_n =
2283 			mprq_caps.min_single_wqe_log_num_of_strides;
2284 		mprq_max_stride_num_n =
2285 			mprq_caps.max_single_wqe_log_num_of_strides;
2286 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2287 						   mprq_min_stride_num_n);
2288 	}
2289 #endif
2290 	if (RTE_CACHE_LINE_SIZE == 128 &&
2291 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2292 		cqe_comp = 0;
2293 	else
2294 		cqe_comp = 1;
2295 	config.cqe_comp = cqe_comp;
2296 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2297 	/* Whether device supports 128B Rx CQE padding. */
2298 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2299 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2300 #endif
2301 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2302 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2303 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
2304 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2305 			     (dv_attr.tunnel_offloads_caps &
2306 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2307 			     (dv_attr.tunnel_offloads_caps &
2308 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2309 	}
2310 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2311 		tunnel_en ? "" : "not ");
2312 #else
2313 	DRV_LOG(WARNING,
2314 		"tunnel offloading disabled due to old OFED/rdma-core version");
2315 #endif
2316 	config.tunnel_en = tunnel_en;
2317 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2318 	mpls_en = ((dv_attr.tunnel_offloads_caps &
2319 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2320 		   (dv_attr.tunnel_offloads_caps &
2321 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2322 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2323 		mpls_en ? "" : "not ");
2324 #else
2325 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2326 		" old OFED/rdma-core version or firmware configuration");
2327 #endif
2328 	config.mpls_en = mpls_en;
2329 	/* Check port status. */
2330 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2331 	if (err) {
2332 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
2333 		goto error;
2334 	}
2335 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2336 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
2337 		err = EINVAL;
2338 		goto error;
2339 	}
2340 	if (port_attr.state != IBV_PORT_ACTIVE)
2341 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2342 			mlx5_glue->port_state_str(port_attr.state),
2343 			port_attr.state);
2344 	/* Allocate private eth device data. */
2345 	priv = rte_zmalloc("ethdev private structure",
2346 			   sizeof(*priv),
2347 			   RTE_CACHE_LINE_SIZE);
2348 	if (priv == NULL) {
2349 		DRV_LOG(ERR, "priv allocation failure");
2350 		err = ENOMEM;
2351 		goto error;
2352 	}
2353 	priv->sh = sh;
2354 	priv->ibv_port = spawn->ibv_port;
2355 	priv->pci_dev = spawn->pci_dev;
2356 	priv->mtu = RTE_ETHER_MTU;
2357 #ifndef RTE_ARCH_64
2358 	/* Initialize UAR access locks for 32bit implementations. */
2359 	rte_spinlock_init(&priv->uar_lock_cq);
2360 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2361 		rte_spinlock_init(&priv->uar_lock[i]);
2362 #endif
2363 	/* Some internal functions rely on Netlink sockets, open them now. */
2364 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2365 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
2366 	priv->representor = !!switch_info->representor;
2367 	priv->master = !!switch_info->master;
2368 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2369 	priv->vport_meta_tag = 0;
2370 	priv->vport_meta_mask = 0;
2371 	priv->pf_bond = spawn->pf_bond;
2372 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2373 	/*
2374 	 * The DevX port query API is implemented. E-Switch may use
2375 	 * either vport or reg_c[0] metadata register to match on
2376 	 * vport index. The engaged part of metadata register is
2377 	 * defined by mask.
2378 	 */
2379 	if (switch_info->representor || switch_info->master) {
2380 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2381 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2382 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2383 						 &devx_port);
2384 		if (err) {
2385 			DRV_LOG(WARNING,
2386 				"can't query devx port %d on device %s",
2387 				spawn->ibv_port, spawn->ibv_dev->name);
2388 			devx_port.comp_mask = 0;
2389 		}
2390 	}
2391 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2392 		priv->vport_meta_tag = devx_port.reg_c_0.value;
2393 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
2394 		if (!priv->vport_meta_mask) {
2395 			DRV_LOG(ERR, "vport zero mask for port %d"
2396 				     " on bonding device %s",
2397 				     spawn->ibv_port, spawn->ibv_dev->name);
2398 			err = ENOTSUP;
2399 			goto error;
2400 		}
2401 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2402 			DRV_LOG(ERR, "invalid vport tag for port %d"
2403 				     " on bonding device %s",
2404 				     spawn->ibv_port, spawn->ibv_dev->name);
2405 			err = ENOTSUP;
2406 			goto error;
2407 		}
2408 	}
2409 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2410 		priv->vport_id = devx_port.vport_num;
2411 	} else if (spawn->pf_bond >= 0) {
2412 		DRV_LOG(ERR, "can't deduce vport index for port %d"
2413 			     " on bonding device %s",
2414 			     spawn->ibv_port, spawn->ibv_dev->name);
2415 		err = ENOTSUP;
2416 		goto error;
2417 	} else {
2418 		/* Suppose vport index in compatible way. */
2419 		priv->vport_id = switch_info->representor ?
2420 				 switch_info->port_name + 1 : -1;
2421 	}
2422 #else
2423 	/*
2424 	 * Kernel/rdma_core support single E-Switch per PF configurations
2425 	 * only and vport_id field contains the vport index for
2426 	 * associated VF, which is deduced from representor port name.
2427 	 * For example, let's have the IB device port 10, it has
2428 	 * attached network device eth0, which has port name attribute
2429 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
2430 	 * as 3 (2+1). This assigning schema should be changed if the
2431 	 * multiple E-Switch instances per PF configurations or/and PCI
2432 	 * subfunctions are added.
2433 	 */
2434 	priv->vport_id = switch_info->representor ?
2435 			 switch_info->port_name + 1 : -1;
2436 #endif
2437 	/* representor_id field keeps the unmodified VF index. */
2438 	priv->representor_id = switch_info->representor ?
2439 			       switch_info->port_name : -1;
2440 	/*
2441 	 * Look for sibling devices in order to reuse their switch domain
2442 	 * if any, otherwise allocate one.
2443 	 */
2444 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2445 		const struct mlx5_priv *opriv =
2446 			rte_eth_devices[port_id].data->dev_private;
2447 
2448 		if (!opriv ||
2449 		    opriv->sh != priv->sh ||
2450 			opriv->domain_id ==
2451 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2452 			continue;
2453 		priv->domain_id = opriv->domain_id;
2454 		break;
2455 	}
2456 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2457 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
2458 		if (err) {
2459 			err = rte_errno;
2460 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
2461 				strerror(rte_errno));
2462 			goto error;
2463 		}
2464 		own_domain_id = 1;
2465 	}
2466 	/* Override some values set by hardware configuration. */
2467 	mlx5_args(&config, dpdk_dev->devargs);
2468 	err = mlx5_dev_check_sibling_config(priv, &config);
2469 	if (err)
2470 		goto error;
2471 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2472 			    IBV_DEVICE_RAW_IP_CSUM);
2473 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2474 		(config.hw_csum ? "" : "not "));
2475 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2476 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2477 	DRV_LOG(DEBUG, "counters are not supported");
2478 #endif
2479 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2480 	if (config.dv_flow_en) {
2481 		DRV_LOG(WARNING, "DV flow is not supported");
2482 		config.dv_flow_en = 0;
2483 	}
2484 #endif
2485 	config.ind_table_max_size =
2486 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2487 	/*
2488 	 * Remove this check once DPDK supports larger/variable
2489 	 * indirection tables.
2490 	 */
2491 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2492 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2493 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2494 		config.ind_table_max_size);
2495 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2496 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2497 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2498 		(config.hw_vlan_strip ? "" : "not "));
2499 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2500 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2501 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2502 		(config.hw_fcs_strip ? "" : "not "));
2503 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2504 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2505 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2506 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2507 			IBV_DEVICE_PCI_WRITE_END_PADDING);
2508 #endif
2509 	if (config.hw_padding && !hw_padding) {
2510 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2511 		config.hw_padding = 0;
2512 	} else if (config.hw_padding) {
2513 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2514 	}
2515 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2516 		      (sh->device_attr.tso_caps.supported_qpts &
2517 		       (1 << IBV_QPT_RAW_PACKET)));
2518 	if (config.tso)
2519 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2520 	/*
2521 	 * MPW is disabled by default, while the Enhanced MPW is enabled
2522 	 * by default.
2523 	 */
2524 	if (config.mps == MLX5_ARG_UNSET)
2525 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2526 							  MLX5_MPW_DISABLED;
2527 	else
2528 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2529 	DRV_LOG(INFO, "%sMPS is %s",
2530 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2531 		config.mps == MLX5_MPW ? "legacy " : "",
2532 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2533 	if (config.cqe_comp && !cqe_comp) {
2534 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2535 		config.cqe_comp = 0;
2536 	}
2537 	if (config.cqe_pad && !cqe_pad) {
2538 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2539 		config.cqe_pad = 0;
2540 	} else if (config.cqe_pad) {
2541 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2542 	}
2543 	if (config.devx) {
2544 		priv->counter_fallback = 0;
2545 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2546 		if (err) {
2547 			err = -err;
2548 			goto error;
2549 		}
2550 		if (!config.hca_attr.flow_counters_dump)
2551 			priv->counter_fallback = 1;
2552 #ifndef HAVE_IBV_DEVX_ASYNC
2553 		priv->counter_fallback = 1;
2554 #endif
2555 		if (priv->counter_fallback)
2556 			DRV_LOG(INFO, "Use fall-back DV counter management");
2557 		/* Check for LRO support. */
2558 		if (config.dest_tir && config.hca_attr.lro_cap &&
2559 		    config.dv_flow_en) {
2560 			/* TBD check tunnel lro caps. */
2561 			config.lro.supported = config.hca_attr.lro_cap;
2562 			DRV_LOG(DEBUG, "Device supports LRO");
2563 			/*
2564 			 * If LRO timeout is not configured by application,
2565 			 * use the minimal supported value.
2566 			 */
2567 			if (!config.lro.timeout)
2568 				config.lro.timeout =
2569 				config.hca_attr.lro_timer_supported_periods[0];
2570 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2571 				config.lro.timeout);
2572 		}
2573 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2574 		if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2575 		    config.dv_flow_en) {
2576 			uint8_t reg_c_mask =
2577 				config.hca_attr.qos.flow_meter_reg_c_ids;
2578 			/*
2579 			 * Meter needs two REG_C's for color match and pre-sfx
2580 			 * flow match. Here get the REG_C for color match.
2581 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2582 			 */
2583 			reg_c_mask &= 0xfc;
2584 			if (__builtin_popcount(reg_c_mask) < 1) {
2585 				priv->mtr_en = 0;
2586 				DRV_LOG(WARNING, "No available register for"
2587 					" meter.");
2588 			} else {
2589 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2590 						      REG_C_0;
2591 				priv->mtr_en = 1;
2592 				priv->mtr_reg_share =
2593 				      config.hca_attr.qos.flow_meter_reg_share;
2594 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2595 					priv->mtr_color_reg);
2596 			}
2597 		}
2598 #endif
2599 	}
2600 	if (config.mprq.enabled && mprq) {
2601 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2602 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2603 			config.mprq.stride_num_n =
2604 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2605 					mprq_min_stride_num_n);
2606 			DRV_LOG(WARNING,
2607 				"the number of strides"
2608 				" for Multi-Packet RQ is out of range,"
2609 				" setting default value (%u)",
2610 				1 << config.mprq.stride_num_n);
2611 		}
2612 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2613 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2614 	} else if (config.mprq.enabled && !mprq) {
2615 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2616 		config.mprq.enabled = 0;
2617 	}
2618 	if (config.max_dump_files_num == 0)
2619 		config.max_dump_files_num = 128;
2620 	eth_dev = rte_eth_dev_allocate(name);
2621 	if (eth_dev == NULL) {
2622 		DRV_LOG(ERR, "can not allocate rte ethdev");
2623 		err = ENOMEM;
2624 		goto error;
2625 	}
2626 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2627 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2628 	if (priv->representor) {
2629 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2630 		eth_dev->data->representor_id = priv->representor_id;
2631 	}
2632 	/*
2633 	 * Store associated network device interface index. This index
2634 	 * is permanent throughout the lifetime of device. So, we may store
2635 	 * the ifindex here and use the cached value further.
2636 	 */
2637 	MLX5_ASSERT(spawn->ifindex);
2638 	priv->if_index = spawn->ifindex;
2639 	eth_dev->data->dev_private = priv;
2640 	priv->dev_data = eth_dev->data;
2641 	eth_dev->data->mac_addrs = priv->mac;
2642 	eth_dev->device = dpdk_dev;
2643 	/* Configure the first MAC address by default. */
2644 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2645 		DRV_LOG(ERR,
2646 			"port %u cannot get MAC address, is mlx5_en"
2647 			" loaded? (errno: %s)",
2648 			eth_dev->data->port_id, strerror(rte_errno));
2649 		err = ENODEV;
2650 		goto error;
2651 	}
2652 	DRV_LOG(INFO,
2653 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2654 		eth_dev->data->port_id,
2655 		mac.addr_bytes[0], mac.addr_bytes[1],
2656 		mac.addr_bytes[2], mac.addr_bytes[3],
2657 		mac.addr_bytes[4], mac.addr_bytes[5]);
2658 #ifdef RTE_LIBRTE_MLX5_DEBUG
2659 	{
2660 		char ifname[IF_NAMESIZE];
2661 
2662 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2663 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2664 				eth_dev->data->port_id, ifname);
2665 		else
2666 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2667 				eth_dev->data->port_id);
2668 	}
2669 #endif
2670 	/* Get actual MTU if possible. */
2671 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2672 	if (err) {
2673 		err = rte_errno;
2674 		goto error;
2675 	}
2676 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2677 		priv->mtu);
2678 	/* Initialize burst functions to prevent crashes before link-up. */
2679 	eth_dev->rx_pkt_burst = removed_rx_burst;
2680 	eth_dev->tx_pkt_burst = removed_tx_burst;
2681 	eth_dev->dev_ops = &mlx5_dev_ops;
2682 	/* Register MAC address. */
2683 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2684 	if (config.vf && config.vf_nl_en)
2685 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2686 				      mlx5_ifindex(eth_dev),
2687 				      eth_dev->data->mac_addrs,
2688 				      MLX5_MAX_MAC_ADDRESSES);
2689 	TAILQ_INIT(&priv->flows);
2690 	TAILQ_INIT(&priv->ctrl_flows);
2691 	TAILQ_INIT(&priv->flow_meters);
2692 	TAILQ_INIT(&priv->flow_meter_profiles);
2693 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2694 	struct mlx5dv_ctx_allocators alctr = {
2695 		.alloc = &mlx5_alloc_verbs_buf,
2696 		.free = &mlx5_free_verbs_buf,
2697 		.data = priv,
2698 	};
2699 	mlx5_glue->dv_set_context_attr(sh->ctx,
2700 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2701 				       (void *)((uintptr_t)&alctr));
2702 	/* Bring Ethernet device up. */
2703 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2704 		eth_dev->data->port_id);
2705 	mlx5_set_link_up(eth_dev);
2706 	/*
2707 	 * Even though the interrupt handler is not installed yet,
2708 	 * interrupts will still trigger on the async_fd from
2709 	 * Verbs context returned by ibv_open_device().
2710 	 */
2711 	mlx5_link_update(eth_dev, 0);
2712 #ifdef HAVE_MLX5DV_DR_ESWITCH
2713 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2714 	      (switch_info->representor || switch_info->master)))
2715 		config.dv_esw_en = 0;
2716 #else
2717 	config.dv_esw_en = 0;
2718 #endif
2719 	/* Detect minimal data bytes to inline. */
2720 	mlx5_set_min_inline(spawn, &config);
2721 	/* Store device configuration on private structure. */
2722 	priv->config = config;
2723 	/* Create context for virtual machine VLAN workaround. */
2724 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2725 	if (config.dv_flow_en) {
2726 		err = mlx5_alloc_shared_dr(priv);
2727 		if (err)
2728 			goto error;
2729 		/*
2730 		 * RSS id is shared with meter flow id. Meter flow id can only
2731 		 * use the 24 MSB of the register.
2732 		 */
2733 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2734 				     MLX5_MTR_COLOR_BITS);
2735 		if (!priv->qrss_id_pool) {
2736 			DRV_LOG(ERR, "can't create flow id pool");
2737 			err = ENOMEM;
2738 			goto error;
2739 		}
2740 	}
2741 	/* Supported Verbs flow priority number detection. */
2742 	err = mlx5_flow_discover_priorities(eth_dev);
2743 	if (err < 0) {
2744 		err = -err;
2745 		goto error;
2746 	}
2747 	priv->config.flow_prio = err;
2748 	if (!priv->config.dv_esw_en &&
2749 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2750 		DRV_LOG(WARNING, "metadata mode %u is not supported "
2751 				 "(no E-Switch)", priv->config.dv_xmeta_en);
2752 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2753 	}
2754 	mlx5_set_metadata_mask(eth_dev);
2755 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2756 	    !priv->sh->dv_regc0_mask) {
2757 		DRV_LOG(ERR, "metadata mode %u is not supported "
2758 			     "(no metadata reg_c[0] is available)",
2759 			     priv->config.dv_xmeta_en);
2760 			err = ENOTSUP;
2761 			goto error;
2762 	}
2763 	/* Query availibility of metadata reg_c's. */
2764 	err = mlx5_flow_discover_mreg_c(eth_dev);
2765 	if (err < 0) {
2766 		err = -err;
2767 		goto error;
2768 	}
2769 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2770 		DRV_LOG(DEBUG,
2771 			"port %u extensive metadata register is not supported",
2772 			eth_dev->data->port_id);
2773 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2774 			DRV_LOG(ERR, "metadata mode %u is not supported "
2775 				     "(no metadata registers available)",
2776 				     priv->config.dv_xmeta_en);
2777 			err = ENOTSUP;
2778 			goto error;
2779 		}
2780 	}
2781 	if (priv->config.dv_flow_en &&
2782 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2783 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
2784 	    priv->sh->dv_regc0_mask) {
2785 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2786 						      MLX5_FLOW_MREG_HTABLE_SZ);
2787 		if (!priv->mreg_cp_tbl) {
2788 			err = ENOMEM;
2789 			goto error;
2790 		}
2791 	}
2792 	return eth_dev;
2793 error:
2794 	if (priv) {
2795 		if (priv->mreg_cp_tbl)
2796 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2797 		if (priv->sh)
2798 			mlx5_free_shared_dr(priv);
2799 		if (priv->nl_socket_route >= 0)
2800 			close(priv->nl_socket_route);
2801 		if (priv->nl_socket_rdma >= 0)
2802 			close(priv->nl_socket_rdma);
2803 		if (priv->vmwa_context)
2804 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2805 		if (priv->qrss_id_pool)
2806 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
2807 		if (own_domain_id)
2808 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2809 		rte_free(priv);
2810 		if (eth_dev != NULL)
2811 			eth_dev->data->dev_private = NULL;
2812 	}
2813 	if (eth_dev != NULL) {
2814 		/* mac_addrs must not be freed alone because part of dev_private */
2815 		eth_dev->data->mac_addrs = NULL;
2816 		rte_eth_dev_release_port(eth_dev);
2817 	}
2818 	if (sh)
2819 		mlx5_free_shared_ibctx(sh);
2820 	MLX5_ASSERT(err > 0);
2821 	rte_errno = err;
2822 	return NULL;
2823 }
2824 
2825 /**
2826  * Comparison callback to sort device data.
2827  *
2828  * This is meant to be used with qsort().
2829  *
2830  * @param a[in]
2831  *   Pointer to pointer to first data object.
2832  * @param b[in]
2833  *   Pointer to pointer to second data object.
2834  *
2835  * @return
2836  *   0 if both objects are equal, less than 0 if the first argument is less
2837  *   than the second, greater than 0 otherwise.
2838  */
2839 static int
2840 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2841 {
2842 	const struct mlx5_switch_info *si_a =
2843 		&((const struct mlx5_dev_spawn_data *)a)->info;
2844 	const struct mlx5_switch_info *si_b =
2845 		&((const struct mlx5_dev_spawn_data *)b)->info;
2846 	int ret;
2847 
2848 	/* Master device first. */
2849 	ret = si_b->master - si_a->master;
2850 	if (ret)
2851 		return ret;
2852 	/* Then representor devices. */
2853 	ret = si_b->representor - si_a->representor;
2854 	if (ret)
2855 		return ret;
2856 	/* Unidentified devices come last in no specific order. */
2857 	if (!si_a->representor)
2858 		return 0;
2859 	/* Order representors by name. */
2860 	return si_a->port_name - si_b->port_name;
2861 }
2862 
2863 /**
2864  * Match PCI information for possible slaves of bonding device.
2865  *
2866  * @param[in] ibv_dev
2867  *   Pointer to Infiniband device structure.
2868  * @param[in] pci_dev
2869  *   Pointer to PCI device structure to match PCI address.
2870  * @param[in] nl_rdma
2871  *   Netlink RDMA group socket handle.
2872  *
2873  * @return
2874  *   negative value if no bonding device found, otherwise
2875  *   positive index of slave PF in bonding.
2876  */
2877 static int
2878 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2879 			   const struct rte_pci_device *pci_dev,
2880 			   int nl_rdma)
2881 {
2882 	char ifname[IF_NAMESIZE + 1];
2883 	unsigned int ifindex;
2884 	unsigned int np, i;
2885 	FILE *file = NULL;
2886 	int pf = -1;
2887 
2888 	/*
2889 	 * Try to get master device name. If something goes
2890 	 * wrong suppose the lack of kernel support and no
2891 	 * bonding devices.
2892 	 */
2893 	if (nl_rdma < 0)
2894 		return -1;
2895 	if (!strstr(ibv_dev->name, "bond"))
2896 		return -1;
2897 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2898 	if (!np)
2899 		return -1;
2900 	/*
2901 	 * The Master device might not be on the predefined
2902 	 * port (not on port index 1, it is not garanted),
2903 	 * we have to scan all Infiniband device port and
2904 	 * find master.
2905 	 */
2906 	for (i = 1; i <= np; ++i) {
2907 		/* Check whether Infiniband port is populated. */
2908 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2909 		if (!ifindex)
2910 			continue;
2911 		if (!if_indextoname(ifindex, ifname))
2912 			continue;
2913 		/* Try to read bonding slave names from sysfs. */
2914 		MKSTR(slaves,
2915 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2916 		file = fopen(slaves, "r");
2917 		if (file)
2918 			break;
2919 	}
2920 	if (!file)
2921 		return -1;
2922 	/* Use safe format to check maximal buffer length. */
2923 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2924 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2925 		char tmp_str[IF_NAMESIZE + 32];
2926 		struct rte_pci_addr pci_addr;
2927 		struct mlx5_switch_info	info;
2928 
2929 		/* Process slave interface names in the loop. */
2930 		snprintf(tmp_str, sizeof(tmp_str),
2931 			 "/sys/class/net/%s", ifname);
2932 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2933 			DRV_LOG(WARNING, "can not get PCI address"
2934 					 " for netdev \"%s\"", ifname);
2935 			continue;
2936 		}
2937 		if (pci_dev->addr.domain != pci_addr.domain ||
2938 		    pci_dev->addr.bus != pci_addr.bus ||
2939 		    pci_dev->addr.devid != pci_addr.devid ||
2940 		    pci_dev->addr.function != pci_addr.function)
2941 			continue;
2942 		/* Slave interface PCI address match found. */
2943 		fclose(file);
2944 		snprintf(tmp_str, sizeof(tmp_str),
2945 			 "/sys/class/net/%s/phys_port_name", ifname);
2946 		file = fopen(tmp_str, "rb");
2947 		if (!file)
2948 			break;
2949 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2950 		if (fscanf(file, "%32s", tmp_str) == 1)
2951 			mlx5_translate_port_name(tmp_str, &info);
2952 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2953 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2954 			pf = info.port_name;
2955 		break;
2956 	}
2957 	if (file)
2958 		fclose(file);
2959 	return pf;
2960 }
2961 
2962 /**
2963  * DPDK callback to register a PCI device.
2964  *
2965  * This function spawns Ethernet devices out of a given PCI device.
2966  *
2967  * @param[in] pci_drv
2968  *   PCI driver structure (mlx5_driver).
2969  * @param[in] pci_dev
2970  *   PCI device information.
2971  *
2972  * @return
2973  *   0 on success, a negative errno value otherwise and rte_errno is set.
2974  */
2975 static int
2976 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2977 	       struct rte_pci_device *pci_dev)
2978 {
2979 	struct ibv_device **ibv_list;
2980 	/*
2981 	 * Number of found IB Devices matching with requested PCI BDF.
2982 	 * nd != 1 means there are multiple IB devices over the same
2983 	 * PCI device and we have representors and master.
2984 	 */
2985 	unsigned int nd = 0;
2986 	/*
2987 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2988 	 * we have the single multiport IB device, and there may be
2989 	 * representors attached to some of found ports.
2990 	 */
2991 	unsigned int np = 0;
2992 	/*
2993 	 * Number of DPDK ethernet devices to Spawn - either over
2994 	 * multiple IB devices or multiple ports of single IB device.
2995 	 * Actually this is the number of iterations to spawn.
2996 	 */
2997 	unsigned int ns = 0;
2998 	/*
2999 	 * Bonding device
3000 	 *   < 0 - no bonding device (single one)
3001 	 *  >= 0 - bonding device (value is slave PF index)
3002 	 */
3003 	int bd = -1;
3004 	struct mlx5_dev_spawn_data *list = NULL;
3005 	struct mlx5_dev_config dev_config;
3006 	int ret;
3007 
3008 	if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
3009 		DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
3010 			" driver.");
3011 		return 1;
3012 	}
3013 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
3014 		mlx5_pmd_socket_init();
3015 	ret = mlx5_init_once();
3016 	if (ret) {
3017 		DRV_LOG(ERR, "unable to init PMD global data: %s",
3018 			strerror(rte_errno));
3019 		return -rte_errno;
3020 	}
3021 	MLX5_ASSERT(pci_drv == &mlx5_driver);
3022 	errno = 0;
3023 	ibv_list = mlx5_glue->get_device_list(&ret);
3024 	if (!ibv_list) {
3025 		rte_errno = errno ? errno : ENOSYS;
3026 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
3027 		return -rte_errno;
3028 	}
3029 	/*
3030 	 * First scan the list of all Infiniband devices to find
3031 	 * matching ones, gathering into the list.
3032 	 */
3033 	struct ibv_device *ibv_match[ret + 1];
3034 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3035 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3036 	unsigned int i;
3037 
3038 	while (ret-- > 0) {
3039 		struct rte_pci_addr pci_addr;
3040 
3041 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3042 		bd = mlx5_device_bond_pci_match
3043 				(ibv_list[ret], pci_dev, nl_rdma);
3044 		if (bd >= 0) {
3045 			/*
3046 			 * Bonding device detected. Only one match is allowed,
3047 			 * the bonding is supported over multi-port IB device,
3048 			 * there should be no matches on representor PCI
3049 			 * functions or non VF LAG bonding devices with
3050 			 * specified address.
3051 			 */
3052 			if (nd) {
3053 				DRV_LOG(ERR,
3054 					"multiple PCI match on bonding device"
3055 					"\"%s\" found", ibv_list[ret]->name);
3056 				rte_errno = ENOENT;
3057 				ret = -rte_errno;
3058 				goto exit;
3059 			}
3060 			DRV_LOG(INFO, "PCI information matches for"
3061 				      " slave %d bonding device \"%s\"",
3062 				      bd, ibv_list[ret]->name);
3063 			ibv_match[nd++] = ibv_list[ret];
3064 			break;
3065 		}
3066 		if (mlx5_dev_to_pci_addr
3067 			(ibv_list[ret]->ibdev_path, &pci_addr))
3068 			continue;
3069 		if (pci_dev->addr.domain != pci_addr.domain ||
3070 		    pci_dev->addr.bus != pci_addr.bus ||
3071 		    pci_dev->addr.devid != pci_addr.devid ||
3072 		    pci_dev->addr.function != pci_addr.function)
3073 			continue;
3074 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3075 			ibv_list[ret]->name);
3076 		ibv_match[nd++] = ibv_list[ret];
3077 	}
3078 	ibv_match[nd] = NULL;
3079 	if (!nd) {
3080 		/* No device matches, just complain and bail out. */
3081 		DRV_LOG(WARNING,
3082 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
3083 			" are kernel drivers loaded?",
3084 			pci_dev->addr.domain, pci_dev->addr.bus,
3085 			pci_dev->addr.devid, pci_dev->addr.function);
3086 		rte_errno = ENOENT;
3087 		ret = -rte_errno;
3088 		goto exit;
3089 	}
3090 	if (nd == 1) {
3091 		/*
3092 		 * Found single matching device may have multiple ports.
3093 		 * Each port may be representor, we have to check the port
3094 		 * number and check the representors existence.
3095 		 */
3096 		if (nl_rdma >= 0)
3097 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3098 		if (!np)
3099 			DRV_LOG(WARNING, "can not get IB device \"%s\""
3100 					 " ports number", ibv_match[0]->name);
3101 		if (bd >= 0 && !np) {
3102 			DRV_LOG(ERR, "can not get ports"
3103 				     " for bonding device");
3104 			rte_errno = ENOENT;
3105 			ret = -rte_errno;
3106 			goto exit;
3107 		}
3108 	}
3109 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3110 	if (bd >= 0) {
3111 		/*
3112 		 * This may happen if there is VF LAG kernel support and
3113 		 * application is compiled with older rdma_core library.
3114 		 */
3115 		DRV_LOG(ERR,
3116 			"No kernel/verbs support for VF LAG bonding found.");
3117 		rte_errno = ENOTSUP;
3118 		ret = -rte_errno;
3119 		goto exit;
3120 	}
3121 #endif
3122 	/*
3123 	 * Now we can determine the maximal
3124 	 * amount of devices to be spawned.
3125 	 */
3126 	list = rte_zmalloc("device spawn data",
3127 			 sizeof(struct mlx5_dev_spawn_data) *
3128 			 (np ? np : nd),
3129 			 RTE_CACHE_LINE_SIZE);
3130 	if (!list) {
3131 		DRV_LOG(ERR, "spawn data array allocation failure");
3132 		rte_errno = ENOMEM;
3133 		ret = -rte_errno;
3134 		goto exit;
3135 	}
3136 	if (bd >= 0 || np > 1) {
3137 		/*
3138 		 * Single IB device with multiple ports found,
3139 		 * it may be E-Switch master device and representors.
3140 		 * We have to perform identification trough the ports.
3141 		 */
3142 		MLX5_ASSERT(nl_rdma >= 0);
3143 		MLX5_ASSERT(ns == 0);
3144 		MLX5_ASSERT(nd == 1);
3145 		MLX5_ASSERT(np);
3146 		for (i = 1; i <= np; ++i) {
3147 			list[ns].max_port = np;
3148 			list[ns].ibv_port = i;
3149 			list[ns].ibv_dev = ibv_match[0];
3150 			list[ns].eth_dev = NULL;
3151 			list[ns].pci_dev = pci_dev;
3152 			list[ns].pf_bond = bd;
3153 			list[ns].ifindex = mlx5_nl_ifindex
3154 					(nl_rdma, list[ns].ibv_dev->name, i);
3155 			if (!list[ns].ifindex) {
3156 				/*
3157 				 * No network interface index found for the
3158 				 * specified port, it means there is no
3159 				 * representor on this port. It's OK,
3160 				 * there can be disabled ports, for example
3161 				 * if sriov_numvfs < sriov_totalvfs.
3162 				 */
3163 				continue;
3164 			}
3165 			ret = -1;
3166 			if (nl_route >= 0)
3167 				ret = mlx5_nl_switch_info
3168 					       (nl_route,
3169 						list[ns].ifindex,
3170 						&list[ns].info);
3171 			if (ret || (!list[ns].info.representor &&
3172 				    !list[ns].info.master)) {
3173 				/*
3174 				 * We failed to recognize representors with
3175 				 * Netlink, let's try to perform the task
3176 				 * with sysfs.
3177 				 */
3178 				ret =  mlx5_sysfs_switch_info
3179 						(list[ns].ifindex,
3180 						 &list[ns].info);
3181 			}
3182 			if (!ret && bd >= 0) {
3183 				switch (list[ns].info.name_type) {
3184 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3185 					if (list[ns].info.port_name == bd)
3186 						ns++;
3187 					break;
3188 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3189 					if (list[ns].info.pf_num == bd)
3190 						ns++;
3191 					break;
3192 				default:
3193 					break;
3194 				}
3195 				continue;
3196 			}
3197 			if (!ret && (list[ns].info.representor ^
3198 				     list[ns].info.master))
3199 				ns++;
3200 		}
3201 		if (!ns) {
3202 			DRV_LOG(ERR,
3203 				"unable to recognize master/representors"
3204 				" on the IB device with multiple ports");
3205 			rte_errno = ENOENT;
3206 			ret = -rte_errno;
3207 			goto exit;
3208 		}
3209 	} else {
3210 		/*
3211 		 * The existence of several matching entries (nd > 1) means
3212 		 * port representors have been instantiated. No existing Verbs
3213 		 * call nor sysfs entries can tell them apart, this can only
3214 		 * be done through Netlink calls assuming kernel drivers are
3215 		 * recent enough to support them.
3216 		 *
3217 		 * In the event of identification failure through Netlink,
3218 		 * try again through sysfs, then:
3219 		 *
3220 		 * 1. A single IB device matches (nd == 1) with single
3221 		 *    port (np=0/1) and is not a representor, assume
3222 		 *    no switch support.
3223 		 *
3224 		 * 2. Otherwise no safe assumptions can be made;
3225 		 *    complain louder and bail out.
3226 		 */
3227 		np = 1;
3228 		for (i = 0; i != nd; ++i) {
3229 			memset(&list[ns].info, 0, sizeof(list[ns].info));
3230 			list[ns].max_port = 1;
3231 			list[ns].ibv_port = 1;
3232 			list[ns].ibv_dev = ibv_match[i];
3233 			list[ns].eth_dev = NULL;
3234 			list[ns].pci_dev = pci_dev;
3235 			list[ns].pf_bond = -1;
3236 			list[ns].ifindex = 0;
3237 			if (nl_rdma >= 0)
3238 				list[ns].ifindex = mlx5_nl_ifindex
3239 					(nl_rdma, list[ns].ibv_dev->name, 1);
3240 			if (!list[ns].ifindex) {
3241 				char ifname[IF_NAMESIZE];
3242 
3243 				/*
3244 				 * Netlink failed, it may happen with old
3245 				 * ib_core kernel driver (before 4.16).
3246 				 * We can assume there is old driver because
3247 				 * here we are processing single ports IB
3248 				 * devices. Let's try sysfs to retrieve
3249 				 * the ifindex. The method works for
3250 				 * master device only.
3251 				 */
3252 				if (nd > 1) {
3253 					/*
3254 					 * Multiple devices found, assume
3255 					 * representors, can not distinguish
3256 					 * master/representor and retrieve
3257 					 * ifindex via sysfs.
3258 					 */
3259 					continue;
3260 				}
3261 				ret = mlx5_get_master_ifname
3262 					(ibv_match[i]->ibdev_path, &ifname);
3263 				if (!ret)
3264 					list[ns].ifindex =
3265 						if_nametoindex(ifname);
3266 				if (!list[ns].ifindex) {
3267 					/*
3268 					 * No network interface index found
3269 					 * for the specified device, it means
3270 					 * there it is neither representor
3271 					 * nor master.
3272 					 */
3273 					continue;
3274 				}
3275 			}
3276 			ret = -1;
3277 			if (nl_route >= 0)
3278 				ret = mlx5_nl_switch_info
3279 					       (nl_route,
3280 						list[ns].ifindex,
3281 						&list[ns].info);
3282 			if (ret || (!list[ns].info.representor &&
3283 				    !list[ns].info.master)) {
3284 				/*
3285 				 * We failed to recognize representors with
3286 				 * Netlink, let's try to perform the task
3287 				 * with sysfs.
3288 				 */
3289 				ret =  mlx5_sysfs_switch_info
3290 						(list[ns].ifindex,
3291 						 &list[ns].info);
3292 			}
3293 			if (!ret && (list[ns].info.representor ^
3294 				     list[ns].info.master)) {
3295 				ns++;
3296 			} else if ((nd == 1) &&
3297 				   !list[ns].info.representor &&
3298 				   !list[ns].info.master) {
3299 				/*
3300 				 * Single IB device with
3301 				 * one physical port and
3302 				 * attached network device.
3303 				 * May be SRIOV is not enabled
3304 				 * or there is no representors.
3305 				 */
3306 				DRV_LOG(INFO, "no E-Switch support detected");
3307 				ns++;
3308 				break;
3309 			}
3310 		}
3311 		if (!ns) {
3312 			DRV_LOG(ERR,
3313 				"unable to recognize master/representors"
3314 				" on the multiple IB devices");
3315 			rte_errno = ENOENT;
3316 			ret = -rte_errno;
3317 			goto exit;
3318 		}
3319 	}
3320 	MLX5_ASSERT(ns);
3321 	/*
3322 	 * Sort list to probe devices in natural order for users convenience
3323 	 * (i.e. master first, then representors from lowest to highest ID).
3324 	 */
3325 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3326 	/* Default configuration. */
3327 	dev_config = (struct mlx5_dev_config){
3328 		.hw_padding = 0,
3329 		.mps = MLX5_ARG_UNSET,
3330 		.dbnc = MLX5_ARG_UNSET,
3331 		.rx_vec_en = 1,
3332 		.txq_inline_max = MLX5_ARG_UNSET,
3333 		.txq_inline_min = MLX5_ARG_UNSET,
3334 		.txq_inline_mpw = MLX5_ARG_UNSET,
3335 		.txqs_inline = MLX5_ARG_UNSET,
3336 		.vf_nl_en = 1,
3337 		.mr_ext_memseg_en = 1,
3338 		.mprq = {
3339 			.enabled = 0, /* Disabled by default. */
3340 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3341 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3342 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3343 		},
3344 		.dv_esw_en = 1,
3345 		.dv_flow_en = 1,
3346 	};
3347 	/* Device specific configuration. */
3348 	switch (pci_dev->id.device_id) {
3349 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3350 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3351 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3352 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3353 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3354 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3355 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3356 		dev_config.vf = 1;
3357 		break;
3358 	default:
3359 		break;
3360 	}
3361 	for (i = 0; i != ns; ++i) {
3362 		uint32_t restore;
3363 
3364 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3365 						 &list[i],
3366 						 dev_config);
3367 		if (!list[i].eth_dev) {
3368 			if (rte_errno != EBUSY && rte_errno != EEXIST)
3369 				break;
3370 			/* Device is disabled or already spawned. Ignore it. */
3371 			continue;
3372 		}
3373 		restore = list[i].eth_dev->data->dev_flags;
3374 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3375 		/* Restore non-PCI flags cleared by the above call. */
3376 		list[i].eth_dev->data->dev_flags |= restore;
3377 		mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3378 		rte_eth_dev_probing_finish(list[i].eth_dev);
3379 	}
3380 	if (i != ns) {
3381 		DRV_LOG(ERR,
3382 			"probe of PCI device " PCI_PRI_FMT " aborted after"
3383 			" encountering an error: %s",
3384 			pci_dev->addr.domain, pci_dev->addr.bus,
3385 			pci_dev->addr.devid, pci_dev->addr.function,
3386 			strerror(rte_errno));
3387 		ret = -rte_errno;
3388 		/* Roll back. */
3389 		while (i--) {
3390 			if (!list[i].eth_dev)
3391 				continue;
3392 			mlx5_dev_close(list[i].eth_dev);
3393 			/* mac_addrs must not be freed because in dev_private */
3394 			list[i].eth_dev->data->mac_addrs = NULL;
3395 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3396 		}
3397 		/* Restore original error. */
3398 		rte_errno = -ret;
3399 	} else {
3400 		ret = 0;
3401 	}
3402 exit:
3403 	/*
3404 	 * Do the routine cleanup:
3405 	 * - close opened Netlink sockets
3406 	 * - free allocated spawn data array
3407 	 * - free the Infiniband device list
3408 	 */
3409 	if (nl_rdma >= 0)
3410 		close(nl_rdma);
3411 	if (nl_route >= 0)
3412 		close(nl_route);
3413 	if (list)
3414 		rte_free(list);
3415 	MLX5_ASSERT(ibv_list);
3416 	mlx5_glue->free_device_list(ibv_list);
3417 	return ret;
3418 }
3419 
3420 /**
3421  * Look for the ethernet device belonging to mlx5 driver.
3422  *
3423  * @param[in] port_id
3424  *   port_id to start looking for device.
3425  * @param[in] pci_dev
3426  *   Pointer to the hint PCI device. When device is being probed
3427  *   the its siblings (master and preceding representors might
3428  *   not have assigned driver yet (because the mlx5_pci_probe()
3429  *   is not completed yet, for this case match on hint PCI
3430  *   device may be used to detect sibling device.
3431  *
3432  * @return
3433  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3434  */
3435 uint16_t
3436 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3437 {
3438 	while (port_id < RTE_MAX_ETHPORTS) {
3439 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3440 
3441 		if (dev->state != RTE_ETH_DEV_UNUSED &&
3442 		    dev->device &&
3443 		    (dev->device == &pci_dev->device ||
3444 		     (dev->device->driver &&
3445 		     dev->device->driver->name &&
3446 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3447 			break;
3448 		port_id++;
3449 	}
3450 	if (port_id >= RTE_MAX_ETHPORTS)
3451 		return RTE_MAX_ETHPORTS;
3452 	return port_id;
3453 }
3454 
3455 /**
3456  * DPDK callback to remove a PCI device.
3457  *
3458  * This function removes all Ethernet devices belong to a given PCI device.
3459  *
3460  * @param[in] pci_dev
3461  *   Pointer to the PCI device.
3462  *
3463  * @return
3464  *   0 on success, the function cannot fail.
3465  */
3466 static int
3467 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3468 {
3469 	uint16_t port_id;
3470 
3471 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3472 		rte_eth_dev_close(port_id);
3473 	return 0;
3474 }
3475 
3476 static const struct rte_pci_id mlx5_pci_id_map[] = {
3477 	{
3478 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3479 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3480 	},
3481 	{
3482 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3483 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3484 	},
3485 	{
3486 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3487 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3488 	},
3489 	{
3490 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3491 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3492 	},
3493 	{
3494 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3495 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3496 	},
3497 	{
3498 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3499 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3500 	},
3501 	{
3502 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3503 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3504 	},
3505 	{
3506 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3507 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3508 	},
3509 	{
3510 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3511 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3512 	},
3513 	{
3514 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3515 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3516 	},
3517 	{
3518 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3519 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3520 	},
3521 	{
3522 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3523 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3524 	},
3525 	{
3526 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3527 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3528 	},
3529 	{
3530 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3531 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3532 	},
3533 	{
3534 		.vendor_id = 0
3535 	}
3536 };
3537 
3538 static struct rte_pci_driver mlx5_driver = {
3539 	.driver = {
3540 		.name = MLX5_DRIVER_NAME
3541 	},
3542 	.id_table = mlx5_pci_id_map,
3543 	.probe = mlx5_pci_probe,
3544 	.remove = mlx5_pci_remove,
3545 	.dma_map = mlx5_dma_map,
3546 	.dma_unmap = mlx5_dma_unmap,
3547 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3548 		     RTE_PCI_DRV_PROBE_AGAIN,
3549 };
3550 
3551 /**
3552  * Driver initialization routine.
3553  */
3554 RTE_INIT(rte_mlx5_pmd_init)
3555 {
3556 	/* Initialize driver log type. */
3557 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3558 	if (mlx5_logtype >= 0)
3559 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3560 
3561 	/* Build the static tables for Verbs conversion. */
3562 	mlx5_set_ptype_table();
3563 	mlx5_set_cksum_table();
3564 	mlx5_set_swp_types_table();
3565 	if (mlx5_glue)
3566 		rte_pci_register(&mlx5_driver);
3567 }
3568 
3569 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3570 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3571 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3572