xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 0964a95120fa024888fbc0ea5e34d1abef1b93dc)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <assert.h>
10 #include <dlfcn.h>
11 #include <stdint.h>
12 #include <stdlib.h>
13 #include <errno.h>
14 #include <net/if.h>
15 #include <sys/mman.h>
16 #include <linux/rtnetlink.h>
17 
18 /* Verbs header. */
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27 
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
31 #include <rte_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_kvargs.h>
36 #include <rte_rwlock.h>
37 #include <rte_spinlock.h>
38 #include <rte_string_fns.h>
39 #include <rte_alarm.h>
40 
41 #include "mlx5.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_autoconf.h"
45 #include "mlx5_defs.h"
46 #include "mlx5_glue.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49 
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52 
53 /* Device parameter to enable RX completion entry padding to 128B. */
54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
55 
56 /* Device parameter to enable padding Rx packet to cacheline size. */
57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
58 
59 /* Device parameter to enable Multi-Packet Rx queue. */
60 #define MLX5_RX_MPRQ_EN "mprq_en"
61 
62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
64 
65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
67 
68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
70 
71 /* Device parameter to configure inline send. Deprecated, ignored.*/
72 #define MLX5_TXQ_INLINE "txq_inline"
73 
74 /* Device parameter to limit packet size to inline with ordinary SEND. */
75 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
76 
77 /* Device parameter to configure minimal data size to inline. */
78 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
79 
80 /* Device parameter to limit packet size to inline with Enhanced MPW. */
81 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
82 
83 /*
84  * Device parameter to configure the number of TX queues threshold for
85  * enabling inline send.
86  */
87 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
88 
89 /*
90  * Device parameter to configure the number of TX queues threshold for
91  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
92  */
93 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
94 
95 /* Device parameter to enable multi-packet send WQEs. */
96 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
97 
98 /*
99  * Device parameter to include 2 dsegs in the title WQEBB.
100  * Deprecated, ignored.
101  */
102 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
103 
104 /*
105  * Device parameter to limit the size of inlining packet.
106  * Deprecated, ignored.
107  */
108 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
109 
110 /*
111  * Device parameter to enable hardware Tx vector.
112  * Deprecated, ignored (no vectorized Tx routines anymore).
113  */
114 #define MLX5_TX_VEC_EN "tx_vec_en"
115 
116 /* Device parameter to enable hardware Rx vector. */
117 #define MLX5_RX_VEC_EN "rx_vec_en"
118 
119 /* Allow L3 VXLAN flow creation. */
120 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
121 
122 /* Activate DV E-Switch flow steering. */
123 #define MLX5_DV_ESW_EN "dv_esw_en"
124 
125 /* Activate DV flow steering. */
126 #define MLX5_DV_FLOW_EN "dv_flow_en"
127 
128 /* Activate Netlink support in VF mode. */
129 #define MLX5_VF_NL_EN "vf_nl_en"
130 
131 /* Enable extending memsegs when creating a MR. */
132 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
133 
134 /* Select port representors to instantiate. */
135 #define MLX5_REPRESENTOR "representor"
136 
137 /* Device parameter to configure the maximum number of dump files per queue. */
138 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
139 
140 /* Configure timeout of LRO session (in microseconds). */
141 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
142 
143 #ifndef HAVE_IBV_MLX5_MOD_MPW
144 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
145 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
146 #endif
147 
148 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
149 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
150 #endif
151 
152 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
153 
154 /* Shared memory between primary and secondary processes. */
155 struct mlx5_shared_data *mlx5_shared_data;
156 
157 /* Spinlock for mlx5_shared_data allocation. */
158 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
159 
160 /* Process local data for secondary processes. */
161 static struct mlx5_local_data mlx5_local_data;
162 
163 /** Driver-specific log messages type. */
164 int mlx5_logtype;
165 
166 /** Data associated with devices to spawn. */
167 struct mlx5_dev_spawn_data {
168 	uint32_t ifindex; /**< Network interface index. */
169 	uint32_t max_port; /**< IB device maximal port index. */
170 	uint32_t ibv_port; /**< IB device physical port index. */
171 	int pf_bond; /**< bonding device PF index. < 0 - no bonding */
172 	struct mlx5_switch_info info; /**< Switch information. */
173 	struct ibv_device *ibv_dev; /**< Associated IB device. */
174 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
175 	struct rte_pci_device *pci_dev; /**< Backend PCI device. */
176 };
177 
178 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
179 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
180 
181 /**
182  * Initialize the counters management structure.
183  *
184  * @param[in] sh
185  *   Pointer to mlx5_ibv_shared object to free
186  */
187 static void
188 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
189 {
190 	uint8_t i;
191 
192 	TAILQ_INIT(&sh->cmng.flow_counters);
193 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
194 		TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
195 }
196 
197 /**
198  * Destroy all the resources allocated for a counter memory management.
199  *
200  * @param[in] mng
201  *   Pointer to the memory management structure.
202  */
203 static void
204 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
205 {
206 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
207 
208 	LIST_REMOVE(mng, next);
209 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
210 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
211 	rte_free(mem);
212 }
213 
214 /**
215  * Close and release all the resources of the counters management.
216  *
217  * @param[in] sh
218  *   Pointer to mlx5_ibv_shared object to free.
219  */
220 static void
221 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
222 {
223 	struct mlx5_counter_stats_mem_mng *mng;
224 	uint8_t i;
225 	int j;
226 	int retries = 1024;
227 
228 	rte_errno = 0;
229 	while (--retries) {
230 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
231 		if (rte_errno != EINPROGRESS)
232 			break;
233 		rte_pause();
234 	}
235 	for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
236 		struct mlx5_flow_counter_pool *pool;
237 		uint32_t batch = !!(i % 2);
238 
239 		if (!sh->cmng.ccont[i].pools)
240 			continue;
241 		pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
242 		while (pool) {
243 			if (batch) {
244 				if (pool->min_dcs)
245 					claim_zero
246 					(mlx5_devx_cmd_destroy(pool->min_dcs));
247 			}
248 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
249 				if (pool->counters_raw[j].action)
250 					claim_zero
251 					(mlx5_glue->destroy_flow_action
252 					       (pool->counters_raw[j].action));
253 				if (!batch && pool->counters_raw[j].dcs)
254 					claim_zero(mlx5_devx_cmd_destroy
255 						  (pool->counters_raw[j].dcs));
256 			}
257 			TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
258 				     next);
259 			rte_free(pool);
260 			pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
261 		}
262 		rte_free(sh->cmng.ccont[i].pools);
263 	}
264 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
265 	while (mng) {
266 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
267 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
268 	}
269 	memset(&sh->cmng, 0, sizeof(sh->cmng));
270 }
271 
272 /**
273  * Extract pdn of PD object using DV API.
274  *
275  * @param[in] pd
276  *   Pointer to the verbs PD object.
277  * @param[out] pdn
278  *   Pointer to the PD object number variable.
279  *
280  * @return
281  *   0 on success, error value otherwise.
282  */
283 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
284 static int
285 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
286 {
287 	struct mlx5dv_obj obj;
288 	struct mlx5dv_pd pd_info;
289 	int ret = 0;
290 
291 	obj.pd.in = pd;
292 	obj.pd.out = &pd_info;
293 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
294 	if (ret) {
295 		DRV_LOG(DEBUG, "Fail to get PD object info");
296 		return ret;
297 	}
298 	*pdn = pd_info.pdn;
299 	return 0;
300 }
301 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
302 
303 /**
304  * Allocate shared IB device context. If there is multiport device the
305  * master and representors will share this context, if there is single
306  * port dedicated IB device, the context will be used by only given
307  * port due to unification.
308  *
309  * Routine first searches the context for the specified IB device name,
310  * if found the shared context assumed and reference counter is incremented.
311  * If no context found the new one is created and initialized with specified
312  * IB device context and parameters.
313  *
314  * @param[in] spawn
315  *   Pointer to the IB device attributes (name, port, etc).
316  *
317  * @return
318  *   Pointer to mlx5_ibv_shared object on success,
319  *   otherwise NULL and rte_errno is set.
320  */
321 static struct mlx5_ibv_shared *
322 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
323 {
324 	struct mlx5_ibv_shared *sh;
325 	int err = 0;
326 	uint32_t i;
327 
328 	assert(spawn);
329 	/* Secondary process should not create the shared context. */
330 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
331 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
332 	/* Search for IB context by device name. */
333 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
334 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
335 			sh->refcnt++;
336 			goto exit;
337 		}
338 	}
339 	/* No device found, we have to create new shared context. */
340 	assert(spawn->max_port);
341 	sh = rte_zmalloc("ethdev shared ib context",
342 			 sizeof(struct mlx5_ibv_shared) +
343 			 spawn->max_port *
344 			 sizeof(struct mlx5_ibv_shared_port),
345 			 RTE_CACHE_LINE_SIZE);
346 	if (!sh) {
347 		DRV_LOG(ERR, "shared context allocation failure");
348 		rte_errno  = ENOMEM;
349 		goto exit;
350 	}
351 	/* Try to open IB device with DV first, then usual Verbs. */
352 	errno = 0;
353 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
354 	if (sh->ctx) {
355 		sh->devx = 1;
356 		DRV_LOG(DEBUG, "DevX is supported");
357 	} else {
358 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
359 		if (!sh->ctx) {
360 			err = errno ? errno : ENODEV;
361 			goto error;
362 		}
363 		DRV_LOG(DEBUG, "DevX is NOT supported");
364 	}
365 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
366 	if (err) {
367 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
368 		goto error;
369 	}
370 	sh->refcnt = 1;
371 	sh->max_port = spawn->max_port;
372 	strncpy(sh->ibdev_name, sh->ctx->device->name,
373 		sizeof(sh->ibdev_name));
374 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
375 		sizeof(sh->ibdev_path));
376 	pthread_mutex_init(&sh->intr_mutex, NULL);
377 	/*
378 	 * Setting port_id to max unallowed value means
379 	 * there is no interrupt subhandler installed for
380 	 * the given port index i.
381 	 */
382 	for (i = 0; i < sh->max_port; i++)
383 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
384 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
385 	if (sh->pd == NULL) {
386 		DRV_LOG(ERR, "PD allocation failure");
387 		err = ENOMEM;
388 		goto error;
389 	}
390 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
391 	err = mlx5_get_pdn(sh->pd, &sh->pdn);
392 	if (err) {
393 		DRV_LOG(ERR, "Fail to extract pdn from PD");
394 		goto error;
395 	}
396 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
397 	/*
398 	 * Once the device is added to the list of memory event
399 	 * callback, its global MR cache table cannot be expanded
400 	 * on the fly because of deadlock. If it overflows, lookup
401 	 * should be done by searching MR list linearly, which is slow.
402 	 *
403 	 * At this point the device is not added to the memory
404 	 * event list yet, context is just being created.
405 	 */
406 	err = mlx5_mr_btree_init(&sh->mr.cache,
407 				 MLX5_MR_BTREE_CACHE_N * 2,
408 				 spawn->pci_dev->device.numa_node);
409 	if (err) {
410 		err = rte_errno;
411 		goto error;
412 	}
413 	mlx5_flow_counters_mng_init(sh);
414 	/* Add device to memory callback list. */
415 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
416 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
417 			 sh, mem_event_cb);
418 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
419 	/* Add context to the global device list. */
420 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
421 exit:
422 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
423 	return sh;
424 error:
425 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
426 	assert(sh);
427 	if (sh->pd)
428 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
429 	if (sh->ctx)
430 		claim_zero(mlx5_glue->close_device(sh->ctx));
431 	rte_free(sh);
432 	assert(err > 0);
433 	rte_errno = err;
434 	return NULL;
435 }
436 
437 /**
438  * Free shared IB device context. Decrement counter and if zero free
439  * all allocated resources and close handles.
440  *
441  * @param[in] sh
442  *   Pointer to mlx5_ibv_shared object to free
443  */
444 static void
445 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
446 {
447 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
448 #ifndef NDEBUG
449 	/* Check the object presence in the list. */
450 	struct mlx5_ibv_shared *lctx;
451 
452 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
453 		if (lctx == sh)
454 			break;
455 	assert(lctx);
456 	if (lctx != sh) {
457 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
458 		goto exit;
459 	}
460 #endif
461 	assert(sh);
462 	assert(sh->refcnt);
463 	/* Secondary process should not free the shared context. */
464 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
465 	if (--sh->refcnt)
466 		goto exit;
467 	/* Release created Memory Regions. */
468 	mlx5_mr_release(sh);
469 	/* Remove from memory callback device list. */
470 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
471 	LIST_REMOVE(sh, mem_event_cb);
472 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
473 	/* Remove context from the global device list. */
474 	LIST_REMOVE(sh, next);
475 	/*
476 	 *  Ensure there is no async event handler installed.
477 	 *  Only primary process handles async device events.
478 	 **/
479 	mlx5_flow_counters_mng_close(sh);
480 	assert(!sh->intr_cnt);
481 	if (sh->intr_cnt)
482 		mlx5_intr_callback_unregister
483 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
484 	pthread_mutex_destroy(&sh->intr_mutex);
485 	if (sh->pd)
486 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
487 	if (sh->ctx)
488 		claim_zero(mlx5_glue->close_device(sh->ctx));
489 	rte_free(sh);
490 exit:
491 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
492 }
493 
494 /**
495  * Initialize DR related data within private structure.
496  * Routine checks the reference counter and does actual
497  * resources creation/initialization only if counter is zero.
498  *
499  * @param[in] priv
500  *   Pointer to the private device data structure.
501  *
502  * @return
503  *   Zero on success, positive error code otherwise.
504  */
505 static int
506 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
507 {
508 #ifdef HAVE_MLX5DV_DR
509 	struct mlx5_ibv_shared *sh = priv->sh;
510 	int err = 0;
511 	void *domain;
512 
513 	assert(sh);
514 	if (sh->dv_refcnt) {
515 		/* Shared DV/DR structures is already initialized. */
516 		sh->dv_refcnt++;
517 		priv->dr_shared = 1;
518 		return 0;
519 	}
520 	/* Reference counter is zero, we should initialize structures. */
521 	domain = mlx5_glue->dr_create_domain(sh->ctx,
522 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
523 	if (!domain) {
524 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
525 		err = errno;
526 		goto error;
527 	}
528 	sh->rx_domain = domain;
529 	domain = mlx5_glue->dr_create_domain(sh->ctx,
530 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
531 	if (!domain) {
532 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
533 		err = errno;
534 		goto error;
535 	}
536 	pthread_mutex_init(&sh->dv_mutex, NULL);
537 	sh->tx_domain = domain;
538 #ifdef HAVE_MLX5DV_DR_ESWITCH
539 	if (priv->config.dv_esw_en) {
540 		domain  = mlx5_glue->dr_create_domain
541 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
542 		if (!domain) {
543 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
544 			err = errno;
545 			goto error;
546 		}
547 		sh->fdb_domain = domain;
548 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
549 	}
550 #endif
551 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
552 	sh->dv_refcnt++;
553 	priv->dr_shared = 1;
554 	return 0;
555 
556 error:
557        /* Rollback the created objects. */
558 	if (sh->rx_domain) {
559 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
560 		sh->rx_domain = NULL;
561 	}
562 	if (sh->tx_domain) {
563 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
564 		sh->tx_domain = NULL;
565 	}
566 	if (sh->fdb_domain) {
567 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
568 		sh->fdb_domain = NULL;
569 	}
570 	if (sh->esw_drop_action) {
571 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
572 		sh->esw_drop_action = NULL;
573 	}
574 	if (sh->pop_vlan_action) {
575 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
576 		sh->pop_vlan_action = NULL;
577 	}
578 	return err;
579 #else
580 	(void)priv;
581 	return 0;
582 #endif
583 }
584 
585 /**
586  * Destroy DR related data within private structure.
587  *
588  * @param[in] priv
589  *   Pointer to the private device data structure.
590  */
591 static void
592 mlx5_free_shared_dr(struct mlx5_priv *priv)
593 {
594 #ifdef HAVE_MLX5DV_DR
595 	struct mlx5_ibv_shared *sh;
596 
597 	if (!priv->dr_shared)
598 		return;
599 	priv->dr_shared = 0;
600 	sh = priv->sh;
601 	assert(sh);
602 	assert(sh->dv_refcnt);
603 	if (sh->dv_refcnt && --sh->dv_refcnt)
604 		return;
605 	if (sh->rx_domain) {
606 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
607 		sh->rx_domain = NULL;
608 	}
609 	if (sh->tx_domain) {
610 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
611 		sh->tx_domain = NULL;
612 	}
613 #ifdef HAVE_MLX5DV_DR_ESWITCH
614 	if (sh->fdb_domain) {
615 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
616 		sh->fdb_domain = NULL;
617 	}
618 	if (sh->esw_drop_action) {
619 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
620 		sh->esw_drop_action = NULL;
621 	}
622 #endif
623 	if (sh->pop_vlan_action) {
624 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
625 		sh->pop_vlan_action = NULL;
626 	}
627 	pthread_mutex_destroy(&sh->dv_mutex);
628 #else
629 	(void)priv;
630 #endif
631 }
632 
633 /**
634  * Initialize shared data between primary and secondary process.
635  *
636  * A memzone is reserved by primary process and secondary processes attach to
637  * the memzone.
638  *
639  * @return
640  *   0 on success, a negative errno value otherwise and rte_errno is set.
641  */
642 static int
643 mlx5_init_shared_data(void)
644 {
645 	const struct rte_memzone *mz;
646 	int ret = 0;
647 
648 	rte_spinlock_lock(&mlx5_shared_data_lock);
649 	if (mlx5_shared_data == NULL) {
650 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
651 			/* Allocate shared memory. */
652 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
653 						 sizeof(*mlx5_shared_data),
654 						 SOCKET_ID_ANY, 0);
655 			if (mz == NULL) {
656 				DRV_LOG(ERR,
657 					"Cannot allocate mlx5 shared data\n");
658 				ret = -rte_errno;
659 				goto error;
660 			}
661 			mlx5_shared_data = mz->addr;
662 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
663 			rte_spinlock_init(&mlx5_shared_data->lock);
664 		} else {
665 			/* Lookup allocated shared memory. */
666 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
667 			if (mz == NULL) {
668 				DRV_LOG(ERR,
669 					"Cannot attach mlx5 shared data\n");
670 				ret = -rte_errno;
671 				goto error;
672 			}
673 			mlx5_shared_data = mz->addr;
674 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
675 		}
676 	}
677 error:
678 	rte_spinlock_unlock(&mlx5_shared_data_lock);
679 	return ret;
680 }
681 
682 /**
683  * Retrieve integer value from environment variable.
684  *
685  * @param[in] name
686  *   Environment variable name.
687  *
688  * @return
689  *   Integer value, 0 if the variable is not set.
690  */
691 int
692 mlx5_getenv_int(const char *name)
693 {
694 	const char *val = getenv(name);
695 
696 	if (val == NULL)
697 		return 0;
698 	return atoi(val);
699 }
700 
701 /**
702  * Verbs callback to allocate a memory. This function should allocate the space
703  * according to the size provided residing inside a huge page.
704  * Please note that all allocation must respect the alignment from libmlx5
705  * (i.e. currently sysconf(_SC_PAGESIZE)).
706  *
707  * @param[in] size
708  *   The size in bytes of the memory to allocate.
709  * @param[in] data
710  *   A pointer to the callback data.
711  *
712  * @return
713  *   Allocated buffer, NULL otherwise and rte_errno is set.
714  */
715 static void *
716 mlx5_alloc_verbs_buf(size_t size, void *data)
717 {
718 	struct mlx5_priv *priv = data;
719 	void *ret;
720 	size_t alignment = sysconf(_SC_PAGESIZE);
721 	unsigned int socket = SOCKET_ID_ANY;
722 
723 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
724 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
725 
726 		socket = ctrl->socket;
727 	} else if (priv->verbs_alloc_ctx.type ==
728 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
729 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
730 
731 		socket = ctrl->socket;
732 	}
733 	assert(data != NULL);
734 	ret = rte_malloc_socket(__func__, size, alignment, socket);
735 	if (!ret && size)
736 		rte_errno = ENOMEM;
737 	return ret;
738 }
739 
740 /**
741  * Verbs callback to free a memory.
742  *
743  * @param[in] ptr
744  *   A pointer to the memory to free.
745  * @param[in] data
746  *   A pointer to the callback data.
747  */
748 static void
749 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
750 {
751 	assert(data != NULL);
752 	rte_free(ptr);
753 }
754 
755 /**
756  * DPDK callback to add udp tunnel port
757  *
758  * @param[in] dev
759  *   A pointer to eth_dev
760  * @param[in] udp_tunnel
761  *   A pointer to udp tunnel
762  *
763  * @return
764  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
765  */
766 int
767 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
768 			 struct rte_eth_udp_tunnel *udp_tunnel)
769 {
770 	assert(udp_tunnel != NULL);
771 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
772 	    udp_tunnel->udp_port == 4789)
773 		return 0;
774 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
775 	    udp_tunnel->udp_port == 4790)
776 		return 0;
777 	return -ENOTSUP;
778 }
779 
780 /**
781  * Initialize process private data structure.
782  *
783  * @param dev
784  *   Pointer to Ethernet device structure.
785  *
786  * @return
787  *   0 on success, a negative errno value otherwise and rte_errno is set.
788  */
789 int
790 mlx5_proc_priv_init(struct rte_eth_dev *dev)
791 {
792 	struct mlx5_priv *priv = dev->data->dev_private;
793 	struct mlx5_proc_priv *ppriv;
794 	size_t ppriv_size;
795 
796 	/*
797 	 * UAR register table follows the process private structure. BlueFlame
798 	 * registers for Tx queues are stored in the table.
799 	 */
800 	ppriv_size =
801 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
802 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
803 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
804 	if (!ppriv) {
805 		rte_errno = ENOMEM;
806 		return -rte_errno;
807 	}
808 	ppriv->uar_table_sz = ppriv_size;
809 	dev->process_private = ppriv;
810 	return 0;
811 }
812 
813 /**
814  * Un-initialize process private data structure.
815  *
816  * @param dev
817  *   Pointer to Ethernet device structure.
818  */
819 static void
820 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
821 {
822 	if (!dev->process_private)
823 		return;
824 	rte_free(dev->process_private);
825 	dev->process_private = NULL;
826 }
827 
828 /**
829  * DPDK callback to close the device.
830  *
831  * Destroy all queues and objects, free memory.
832  *
833  * @param dev
834  *   Pointer to Ethernet device structure.
835  */
836 static void
837 mlx5_dev_close(struct rte_eth_dev *dev)
838 {
839 	struct mlx5_priv *priv = dev->data->dev_private;
840 	unsigned int i;
841 	int ret;
842 
843 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
844 		dev->data->port_id,
845 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
846 	/* In case mlx5_dev_stop() has not been called. */
847 	mlx5_dev_interrupt_handler_uninstall(dev);
848 	mlx5_traffic_disable(dev);
849 	mlx5_flow_flush(dev, NULL);
850 	/* Prevent crashes when queues are still in use. */
851 	dev->rx_pkt_burst = removed_rx_burst;
852 	dev->tx_pkt_burst = removed_tx_burst;
853 	rte_wmb();
854 	/* Disable datapath on secondary process. */
855 	mlx5_mp_req_stop_rxtx(dev);
856 	if (priv->rxqs != NULL) {
857 		/* XXX race condition if mlx5_rx_burst() is still running. */
858 		usleep(1000);
859 		for (i = 0; (i != priv->rxqs_n); ++i)
860 			mlx5_rxq_release(dev, i);
861 		priv->rxqs_n = 0;
862 		priv->rxqs = NULL;
863 	}
864 	if (priv->txqs != NULL) {
865 		/* XXX race condition if mlx5_tx_burst() is still running. */
866 		usleep(1000);
867 		for (i = 0; (i != priv->txqs_n); ++i)
868 			mlx5_txq_release(dev, i);
869 		priv->txqs_n = 0;
870 		priv->txqs = NULL;
871 	}
872 	mlx5_proc_priv_uninit(dev);
873 	mlx5_mprq_free_mp(dev);
874 	mlx5_free_shared_dr(priv);
875 	if (priv->rss_conf.rss_key != NULL)
876 		rte_free(priv->rss_conf.rss_key);
877 	if (priv->reta_idx != NULL)
878 		rte_free(priv->reta_idx);
879 	if (priv->config.vf)
880 		mlx5_nl_mac_addr_flush(dev);
881 	if (priv->nl_socket_route >= 0)
882 		close(priv->nl_socket_route);
883 	if (priv->nl_socket_rdma >= 0)
884 		close(priv->nl_socket_rdma);
885 	if (priv->vmwa_context)
886 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
887 	if (priv->sh) {
888 		/*
889 		 * Free the shared context in last turn, because the cleanup
890 		 * routines above may use some shared fields, like
891 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
892 		 * ifindex if Netlink fails.
893 		 */
894 		mlx5_free_shared_ibctx(priv->sh);
895 		priv->sh = NULL;
896 	}
897 	ret = mlx5_hrxq_verify(dev);
898 	if (ret)
899 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
900 			dev->data->port_id);
901 	ret = mlx5_ind_table_obj_verify(dev);
902 	if (ret)
903 		DRV_LOG(WARNING, "port %u some indirection table still remain",
904 			dev->data->port_id);
905 	ret = mlx5_rxq_obj_verify(dev);
906 	if (ret)
907 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
908 			dev->data->port_id);
909 	ret = mlx5_rxq_verify(dev);
910 	if (ret)
911 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
912 			dev->data->port_id);
913 	ret = mlx5_txq_ibv_verify(dev);
914 	if (ret)
915 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
916 			dev->data->port_id);
917 	ret = mlx5_txq_verify(dev);
918 	if (ret)
919 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
920 			dev->data->port_id);
921 	ret = mlx5_flow_verify(dev);
922 	if (ret)
923 		DRV_LOG(WARNING, "port %u some flows still remain",
924 			dev->data->port_id);
925 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
926 		unsigned int c = 0;
927 		uint16_t port_id;
928 
929 		MLX5_ETH_FOREACH_DEV(port_id) {
930 			struct mlx5_priv *opriv =
931 				rte_eth_devices[port_id].data->dev_private;
932 
933 			if (!opriv ||
934 			    opriv->domain_id != priv->domain_id ||
935 			    &rte_eth_devices[port_id] == dev)
936 				continue;
937 			++c;
938 			break;
939 		}
940 		if (!c)
941 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
942 	}
943 	memset(priv, 0, sizeof(*priv));
944 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
945 	/*
946 	 * Reset mac_addrs to NULL such that it is not freed as part of
947 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
948 	 * it is freed when dev_private is freed.
949 	 */
950 	dev->data->mac_addrs = NULL;
951 }
952 
953 const struct eth_dev_ops mlx5_dev_ops = {
954 	.dev_configure = mlx5_dev_configure,
955 	.dev_start = mlx5_dev_start,
956 	.dev_stop = mlx5_dev_stop,
957 	.dev_set_link_down = mlx5_set_link_down,
958 	.dev_set_link_up = mlx5_set_link_up,
959 	.dev_close = mlx5_dev_close,
960 	.promiscuous_enable = mlx5_promiscuous_enable,
961 	.promiscuous_disable = mlx5_promiscuous_disable,
962 	.allmulticast_enable = mlx5_allmulticast_enable,
963 	.allmulticast_disable = mlx5_allmulticast_disable,
964 	.link_update = mlx5_link_update,
965 	.stats_get = mlx5_stats_get,
966 	.stats_reset = mlx5_stats_reset,
967 	.xstats_get = mlx5_xstats_get,
968 	.xstats_reset = mlx5_xstats_reset,
969 	.xstats_get_names = mlx5_xstats_get_names,
970 	.fw_version_get = mlx5_fw_version_get,
971 	.dev_infos_get = mlx5_dev_infos_get,
972 	.read_clock = mlx5_read_clock,
973 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
974 	.vlan_filter_set = mlx5_vlan_filter_set,
975 	.rx_queue_setup = mlx5_rx_queue_setup,
976 	.tx_queue_setup = mlx5_tx_queue_setup,
977 	.rx_queue_release = mlx5_rx_queue_release,
978 	.tx_queue_release = mlx5_tx_queue_release,
979 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
980 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
981 	.mac_addr_remove = mlx5_mac_addr_remove,
982 	.mac_addr_add = mlx5_mac_addr_add,
983 	.mac_addr_set = mlx5_mac_addr_set,
984 	.set_mc_addr_list = mlx5_set_mc_addr_list,
985 	.mtu_set = mlx5_dev_set_mtu,
986 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
987 	.vlan_offload_set = mlx5_vlan_offload_set,
988 	.reta_update = mlx5_dev_rss_reta_update,
989 	.reta_query = mlx5_dev_rss_reta_query,
990 	.rss_hash_update = mlx5_rss_hash_update,
991 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
992 	.filter_ctrl = mlx5_dev_filter_ctrl,
993 	.rx_descriptor_status = mlx5_rx_descriptor_status,
994 	.tx_descriptor_status = mlx5_tx_descriptor_status,
995 	.rx_queue_count = mlx5_rx_queue_count,
996 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
997 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
998 	.is_removed = mlx5_is_removed,
999 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1000 	.get_module_info = mlx5_get_module_info,
1001 	.get_module_eeprom = mlx5_get_module_eeprom,
1002 };
1003 
1004 /* Available operations from secondary process. */
1005 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1006 	.stats_get = mlx5_stats_get,
1007 	.stats_reset = mlx5_stats_reset,
1008 	.xstats_get = mlx5_xstats_get,
1009 	.xstats_reset = mlx5_xstats_reset,
1010 	.xstats_get_names = mlx5_xstats_get_names,
1011 	.fw_version_get = mlx5_fw_version_get,
1012 	.dev_infos_get = mlx5_dev_infos_get,
1013 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1014 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1015 	.get_module_info = mlx5_get_module_info,
1016 	.get_module_eeprom = mlx5_get_module_eeprom,
1017 };
1018 
1019 /* Available operations in flow isolated mode. */
1020 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1021 	.dev_configure = mlx5_dev_configure,
1022 	.dev_start = mlx5_dev_start,
1023 	.dev_stop = mlx5_dev_stop,
1024 	.dev_set_link_down = mlx5_set_link_down,
1025 	.dev_set_link_up = mlx5_set_link_up,
1026 	.dev_close = mlx5_dev_close,
1027 	.promiscuous_enable = mlx5_promiscuous_enable,
1028 	.promiscuous_disable = mlx5_promiscuous_disable,
1029 	.allmulticast_enable = mlx5_allmulticast_enable,
1030 	.allmulticast_disable = mlx5_allmulticast_disable,
1031 	.link_update = mlx5_link_update,
1032 	.stats_get = mlx5_stats_get,
1033 	.stats_reset = mlx5_stats_reset,
1034 	.xstats_get = mlx5_xstats_get,
1035 	.xstats_reset = mlx5_xstats_reset,
1036 	.xstats_get_names = mlx5_xstats_get_names,
1037 	.fw_version_get = mlx5_fw_version_get,
1038 	.dev_infos_get = mlx5_dev_infos_get,
1039 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1040 	.vlan_filter_set = mlx5_vlan_filter_set,
1041 	.rx_queue_setup = mlx5_rx_queue_setup,
1042 	.tx_queue_setup = mlx5_tx_queue_setup,
1043 	.rx_queue_release = mlx5_rx_queue_release,
1044 	.tx_queue_release = mlx5_tx_queue_release,
1045 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1046 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1047 	.mac_addr_remove = mlx5_mac_addr_remove,
1048 	.mac_addr_add = mlx5_mac_addr_add,
1049 	.mac_addr_set = mlx5_mac_addr_set,
1050 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1051 	.mtu_set = mlx5_dev_set_mtu,
1052 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1053 	.vlan_offload_set = mlx5_vlan_offload_set,
1054 	.filter_ctrl = mlx5_dev_filter_ctrl,
1055 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1056 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1057 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1058 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1059 	.is_removed = mlx5_is_removed,
1060 	.get_module_info = mlx5_get_module_info,
1061 	.get_module_eeprom = mlx5_get_module_eeprom,
1062 };
1063 
1064 /**
1065  * Verify and store value for device argument.
1066  *
1067  * @param[in] key
1068  *   Key argument to verify.
1069  * @param[in] val
1070  *   Value associated with key.
1071  * @param opaque
1072  *   User data.
1073  *
1074  * @return
1075  *   0 on success, a negative errno value otherwise and rte_errno is set.
1076  */
1077 static int
1078 mlx5_args_check(const char *key, const char *val, void *opaque)
1079 {
1080 	struct mlx5_dev_config *config = opaque;
1081 	unsigned long tmp;
1082 
1083 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1084 	if (!strcmp(MLX5_REPRESENTOR, key))
1085 		return 0;
1086 	errno = 0;
1087 	tmp = strtoul(val, NULL, 0);
1088 	if (errno) {
1089 		rte_errno = errno;
1090 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1091 		return -rte_errno;
1092 	}
1093 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1094 		config->cqe_comp = !!tmp;
1095 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1096 		config->cqe_pad = !!tmp;
1097 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1098 		config->hw_padding = !!tmp;
1099 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1100 		config->mprq.enabled = !!tmp;
1101 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1102 		config->mprq.stride_num_n = tmp;
1103 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1104 		config->mprq.max_memcpy_len = tmp;
1105 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1106 		config->mprq.min_rxqs_num = tmp;
1107 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1108 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1109 				 " converted to txq_inline_max", key);
1110 		config->txq_inline_max = tmp;
1111 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1112 		config->txq_inline_max = tmp;
1113 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1114 		config->txq_inline_min = tmp;
1115 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1116 		config->txq_inline_mpw = tmp;
1117 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1118 		config->txqs_inline = tmp;
1119 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1120 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1121 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1122 		config->mps = !!tmp;
1123 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1124 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1125 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1126 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1127 				 " converted to txq_inline_mpw", key);
1128 		config->txq_inline_mpw = tmp;
1129 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1130 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1131 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1132 		config->rx_vec_en = !!tmp;
1133 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1134 		config->l3_vxlan_en = !!tmp;
1135 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1136 		config->vf_nl_en = !!tmp;
1137 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1138 		config->dv_esw_en = !!tmp;
1139 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1140 		config->dv_flow_en = !!tmp;
1141 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1142 		config->mr_ext_memseg_en = !!tmp;
1143 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1144 		config->max_dump_files_num = tmp;
1145 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1146 		config->lro.timeout = tmp;
1147 	} else {
1148 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1149 		rte_errno = EINVAL;
1150 		return -rte_errno;
1151 	}
1152 	return 0;
1153 }
1154 
1155 /**
1156  * Parse device parameters.
1157  *
1158  * @param config
1159  *   Pointer to device configuration structure.
1160  * @param devargs
1161  *   Device arguments structure.
1162  *
1163  * @return
1164  *   0 on success, a negative errno value otherwise and rte_errno is set.
1165  */
1166 static int
1167 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1168 {
1169 	const char **params = (const char *[]){
1170 		MLX5_RXQ_CQE_COMP_EN,
1171 		MLX5_RXQ_CQE_PAD_EN,
1172 		MLX5_RXQ_PKT_PAD_EN,
1173 		MLX5_RX_MPRQ_EN,
1174 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1175 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1176 		MLX5_RXQS_MIN_MPRQ,
1177 		MLX5_TXQ_INLINE,
1178 		MLX5_TXQ_INLINE_MIN,
1179 		MLX5_TXQ_INLINE_MAX,
1180 		MLX5_TXQ_INLINE_MPW,
1181 		MLX5_TXQS_MIN_INLINE,
1182 		MLX5_TXQS_MAX_VEC,
1183 		MLX5_TXQ_MPW_EN,
1184 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1185 		MLX5_TXQ_MAX_INLINE_LEN,
1186 		MLX5_TX_VEC_EN,
1187 		MLX5_RX_VEC_EN,
1188 		MLX5_L3_VXLAN_EN,
1189 		MLX5_VF_NL_EN,
1190 		MLX5_DV_ESW_EN,
1191 		MLX5_DV_FLOW_EN,
1192 		MLX5_MR_EXT_MEMSEG_EN,
1193 		MLX5_REPRESENTOR,
1194 		MLX5_MAX_DUMP_FILES_NUM,
1195 		MLX5_LRO_TIMEOUT_USEC,
1196 		NULL,
1197 	};
1198 	struct rte_kvargs *kvlist;
1199 	int ret = 0;
1200 	int i;
1201 
1202 	if (devargs == NULL)
1203 		return 0;
1204 	/* Following UGLY cast is done to pass checkpatch. */
1205 	kvlist = rte_kvargs_parse(devargs->args, params);
1206 	if (kvlist == NULL) {
1207 		rte_errno = EINVAL;
1208 		return -rte_errno;
1209 	}
1210 	/* Process parameters. */
1211 	for (i = 0; (params[i] != NULL); ++i) {
1212 		if (rte_kvargs_count(kvlist, params[i])) {
1213 			ret = rte_kvargs_process(kvlist, params[i],
1214 						 mlx5_args_check, config);
1215 			if (ret) {
1216 				rte_errno = EINVAL;
1217 				rte_kvargs_free(kvlist);
1218 				return -rte_errno;
1219 			}
1220 		}
1221 	}
1222 	rte_kvargs_free(kvlist);
1223 	return 0;
1224 }
1225 
1226 static struct rte_pci_driver mlx5_driver;
1227 
1228 /**
1229  * PMD global initialization.
1230  *
1231  * Independent from individual device, this function initializes global
1232  * per-PMD data structures distinguishing primary and secondary processes.
1233  * Hence, each initialization is called once per a process.
1234  *
1235  * @return
1236  *   0 on success, a negative errno value otherwise and rte_errno is set.
1237  */
1238 static int
1239 mlx5_init_once(void)
1240 {
1241 	struct mlx5_shared_data *sd;
1242 	struct mlx5_local_data *ld = &mlx5_local_data;
1243 	int ret = 0;
1244 
1245 	if (mlx5_init_shared_data())
1246 		return -rte_errno;
1247 	sd = mlx5_shared_data;
1248 	assert(sd);
1249 	rte_spinlock_lock(&sd->lock);
1250 	switch (rte_eal_process_type()) {
1251 	case RTE_PROC_PRIMARY:
1252 		if (sd->init_done)
1253 			break;
1254 		LIST_INIT(&sd->mem_event_cb_list);
1255 		rte_rwlock_init(&sd->mem_event_rwlock);
1256 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1257 						mlx5_mr_mem_event_cb, NULL);
1258 		ret = mlx5_mp_init_primary();
1259 		if (ret)
1260 			goto out;
1261 		sd->init_done = true;
1262 		break;
1263 	case RTE_PROC_SECONDARY:
1264 		if (ld->init_done)
1265 			break;
1266 		ret = mlx5_mp_init_secondary();
1267 		if (ret)
1268 			goto out;
1269 		++sd->secondary_cnt;
1270 		ld->init_done = true;
1271 		break;
1272 	default:
1273 		break;
1274 	}
1275 out:
1276 	rte_spinlock_unlock(&sd->lock);
1277 	return ret;
1278 }
1279 
1280 /**
1281  * Configures the minimal amount of data to inline into WQE
1282  * while sending packets.
1283  *
1284  * - the txq_inline_min has the maximal priority, if this
1285  *   key is specified in devargs
1286  * - if DevX is enabled the inline mode is queried from the
1287  *   device (HCA attributes and NIC vport context if needed).
1288  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1289  *   and none (0 bytes) for other NICs
1290  *
1291  * @param spawn
1292  *   Verbs device parameters (name, port, switch_info) to spawn.
1293  * @param config
1294  *   Device configuration parameters.
1295  */
1296 static void
1297 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1298 		    struct mlx5_dev_config *config)
1299 {
1300 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1301 		/* Application defines size of inlined data explicitly. */
1302 		switch (spawn->pci_dev->id.device_id) {
1303 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1304 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1305 			if (config->txq_inline_min <
1306 				       (int)MLX5_INLINE_HSIZE_L2) {
1307 				DRV_LOG(DEBUG,
1308 					"txq_inline_mix aligned to minimal"
1309 					" ConnectX-4 required value %d",
1310 					(int)MLX5_INLINE_HSIZE_L2);
1311 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1312 			}
1313 			break;
1314 		}
1315 		goto exit;
1316 	}
1317 	if (config->hca_attr.eth_net_offloads) {
1318 		/* We have DevX enabled, inline mode queried successfully. */
1319 		switch (config->hca_attr.wqe_inline_mode) {
1320 		case MLX5_CAP_INLINE_MODE_L2:
1321 			/* outer L2 header must be inlined. */
1322 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1323 			goto exit;
1324 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1325 			/* No inline data are required by NIC. */
1326 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1327 			config->hw_vlan_insert =
1328 				config->hca_attr.wqe_vlan_insert;
1329 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1330 			goto exit;
1331 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1332 			/* inline mode is defined by NIC vport context. */
1333 			if (!config->hca_attr.eth_virt)
1334 				break;
1335 			switch (config->hca_attr.vport_inline_mode) {
1336 			case MLX5_INLINE_MODE_NONE:
1337 				config->txq_inline_min =
1338 					MLX5_INLINE_HSIZE_NONE;
1339 				goto exit;
1340 			case MLX5_INLINE_MODE_L2:
1341 				config->txq_inline_min =
1342 					MLX5_INLINE_HSIZE_L2;
1343 				goto exit;
1344 			case MLX5_INLINE_MODE_IP:
1345 				config->txq_inline_min =
1346 					MLX5_INLINE_HSIZE_L3;
1347 				goto exit;
1348 			case MLX5_INLINE_MODE_TCP_UDP:
1349 				config->txq_inline_min =
1350 					MLX5_INLINE_HSIZE_L4;
1351 				goto exit;
1352 			case MLX5_INLINE_MODE_INNER_L2:
1353 				config->txq_inline_min =
1354 					MLX5_INLINE_HSIZE_INNER_L2;
1355 				goto exit;
1356 			case MLX5_INLINE_MODE_INNER_IP:
1357 				config->txq_inline_min =
1358 					MLX5_INLINE_HSIZE_INNER_L3;
1359 				goto exit;
1360 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1361 				config->txq_inline_min =
1362 					MLX5_INLINE_HSIZE_INNER_L4;
1363 				goto exit;
1364 			}
1365 		}
1366 	}
1367 	/*
1368 	 * We get here if we are unable to deduce
1369 	 * inline data size with DevX. Try PCI ID
1370 	 * to determine old NICs.
1371 	 */
1372 	switch (spawn->pci_dev->id.device_id) {
1373 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1374 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1375 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1376 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1377 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1378 		config->hw_vlan_insert = 0;
1379 		break;
1380 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1381 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1382 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1383 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1384 		/*
1385 		 * These NICs support VLAN insertion from WQE and
1386 		 * report the wqe_vlan_insert flag. But there is the bug
1387 		 * and PFC control may be broken, so disable feature.
1388 		 */
1389 		config->hw_vlan_insert = 0;
1390 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1391 		break;
1392 	default:
1393 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1394 		break;
1395 	}
1396 exit:
1397 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1398 }
1399 
1400 /**
1401  * Allocate page of door-bells and register it using DevX API.
1402  *
1403  * @param [in] dev
1404  *   Pointer to Ethernet device.
1405  *
1406  * @return
1407  *   Pointer to new page on success, NULL otherwise.
1408  */
1409 static struct mlx5_devx_dbr_page *
1410 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1411 {
1412 	struct mlx5_priv *priv = dev->data->dev_private;
1413 	struct mlx5_devx_dbr_page *page;
1414 
1415 	/* Allocate space for door-bell page and management data. */
1416 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1417 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1418 	if (!page) {
1419 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1420 			dev->data->port_id);
1421 		return NULL;
1422 	}
1423 	/* Register allocated memory. */
1424 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1425 					      MLX5_DBR_PAGE_SIZE, 0);
1426 	if (!page->umem) {
1427 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1428 			dev->data->port_id);
1429 		rte_free(page);
1430 		return NULL;
1431 	}
1432 	return page;
1433 }
1434 
1435 /**
1436  * Find the next available door-bell, allocate new page if needed.
1437  *
1438  * @param [in] dev
1439  *   Pointer to Ethernet device.
1440  * @param [out] dbr_page
1441  *   Door-bell page containing the page data.
1442  *
1443  * @return
1444  *   Door-bell address offset on success, a negative error value otherwise.
1445  */
1446 int64_t
1447 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1448 {
1449 	struct mlx5_priv *priv = dev->data->dev_private;
1450 	struct mlx5_devx_dbr_page *page = NULL;
1451 	uint32_t i, j;
1452 
1453 	LIST_FOREACH(page, &priv->dbrpgs, next)
1454 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1455 			break;
1456 	if (!page) { /* No page with free door-bell exists. */
1457 		page = mlx5_alloc_dbr_page(dev);
1458 		if (!page) /* Failed to allocate new page. */
1459 			return (-1);
1460 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1461 	}
1462 	/* Loop to find bitmap part with clear bit. */
1463 	for (i = 0;
1464 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1465 	     i++)
1466 		; /* Empty. */
1467 	/* Find the first clear bit. */
1468 	j = rte_bsf64(~page->dbr_bitmap[i]);
1469 	assert(i < (MLX5_DBR_PER_PAGE / 64));
1470 	page->dbr_bitmap[i] |= (1 << j);
1471 	page->dbr_count++;
1472 	*dbr_page = page;
1473 	return (((i * 64) + j) * sizeof(uint64_t));
1474 }
1475 
1476 /**
1477  * Release a door-bell record.
1478  *
1479  * @param [in] dev
1480  *   Pointer to Ethernet device.
1481  * @param [in] umem_id
1482  *   UMEM ID of page containing the door-bell record to release.
1483  * @param [in] offset
1484  *   Offset of door-bell record in page.
1485  *
1486  * @return
1487  *   0 on success, a negative error value otherwise.
1488  */
1489 int32_t
1490 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1491 {
1492 	struct mlx5_priv *priv = dev->data->dev_private;
1493 	struct mlx5_devx_dbr_page *page = NULL;
1494 	int ret = 0;
1495 
1496 	LIST_FOREACH(page, &priv->dbrpgs, next)
1497 		/* Find the page this address belongs to. */
1498 		if (page->umem->umem_id == umem_id)
1499 			break;
1500 	if (!page)
1501 		return -EINVAL;
1502 	page->dbr_count--;
1503 	if (!page->dbr_count) {
1504 		/* Page not used, free it and remove from list. */
1505 		LIST_REMOVE(page, next);
1506 		if (page->umem)
1507 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
1508 		rte_free(page);
1509 	} else {
1510 		/* Mark in bitmap that this door-bell is not in use. */
1511 		offset /= MLX5_DBR_SIZE;
1512 		int i = offset / 64;
1513 		int j = offset % 64;
1514 
1515 		page->dbr_bitmap[i] &= ~(1 << j);
1516 	}
1517 	return ret;
1518 }
1519 
1520 /**
1521  * Check sibling device configurations.
1522  *
1523  * Sibling devices sharing the Infiniband device context
1524  * should have compatible configurations. This regards
1525  * representors and bonding slaves.
1526  *
1527  * @param priv
1528  *   Private device descriptor.
1529  * @param config
1530  *   Configuration of the device is going to be created.
1531  *
1532  * @return
1533  *   0 on success, EINVAL otherwise
1534  */
1535 static int
1536 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1537 			      struct mlx5_dev_config *config)
1538 {
1539 	struct mlx5_ibv_shared *sh = priv->sh;
1540 	struct mlx5_dev_config *sh_conf = NULL;
1541 	uint16_t port_id;
1542 
1543 	assert(sh);
1544 	/* Nothing to compare for the single/first device. */
1545 	if (sh->refcnt == 1)
1546 		return 0;
1547 	/* Find the device with shared context. */
1548 	MLX5_ETH_FOREACH_DEV(port_id) {
1549 		struct mlx5_priv *opriv =
1550 			rte_eth_devices[port_id].data->dev_private;
1551 
1552 		if (opriv && opriv != priv && opriv->sh == sh) {
1553 			sh_conf = &opriv->config;
1554 			break;
1555 		}
1556 	}
1557 	if (!sh_conf)
1558 		return 0;
1559 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1560 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1561 			     " for shared %s context", sh->ibdev_name);
1562 		rte_errno = EINVAL;
1563 		return rte_errno;
1564 	}
1565 	return 0;
1566 }
1567 /**
1568  * Spawn an Ethernet device from Verbs information.
1569  *
1570  * @param dpdk_dev
1571  *   Backing DPDK device.
1572  * @param spawn
1573  *   Verbs device parameters (name, port, switch_info) to spawn.
1574  * @param config
1575  *   Device configuration parameters.
1576  *
1577  * @return
1578  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1579  *   is set. The following errors are defined:
1580  *
1581  *   EBUSY: device is not supposed to be spawned.
1582  *   EEXIST: device is already spawned
1583  */
1584 static struct rte_eth_dev *
1585 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1586 	       struct mlx5_dev_spawn_data *spawn,
1587 	       struct mlx5_dev_config config)
1588 {
1589 	const struct mlx5_switch_info *switch_info = &spawn->info;
1590 	struct mlx5_ibv_shared *sh = NULL;
1591 	struct ibv_port_attr port_attr;
1592 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1593 	struct rte_eth_dev *eth_dev = NULL;
1594 	struct mlx5_priv *priv = NULL;
1595 	int err = 0;
1596 	unsigned int hw_padding = 0;
1597 	unsigned int mps;
1598 	unsigned int cqe_comp;
1599 	unsigned int cqe_pad = 0;
1600 	unsigned int tunnel_en = 0;
1601 	unsigned int mpls_en = 0;
1602 	unsigned int swp = 0;
1603 	unsigned int mprq = 0;
1604 	unsigned int mprq_min_stride_size_n = 0;
1605 	unsigned int mprq_max_stride_size_n = 0;
1606 	unsigned int mprq_min_stride_num_n = 0;
1607 	unsigned int mprq_max_stride_num_n = 0;
1608 	struct rte_ether_addr mac;
1609 	char name[RTE_ETH_NAME_MAX_LEN];
1610 	int own_domain_id = 0;
1611 	uint16_t port_id;
1612 	unsigned int i;
1613 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1614 	struct mlx5dv_devx_port devx_port;
1615 #endif
1616 
1617 	/* Determine if this port representor is supposed to be spawned. */
1618 	if (switch_info->representor && dpdk_dev->devargs) {
1619 		struct rte_eth_devargs eth_da;
1620 
1621 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1622 		if (err) {
1623 			rte_errno = -err;
1624 			DRV_LOG(ERR, "failed to process device arguments: %s",
1625 				strerror(rte_errno));
1626 			return NULL;
1627 		}
1628 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
1629 			if (eth_da.representor_ports[i] ==
1630 			    (uint16_t)switch_info->port_name)
1631 				break;
1632 		if (i == eth_da.nb_representor_ports) {
1633 			rte_errno = EBUSY;
1634 			return NULL;
1635 		}
1636 	}
1637 	/* Build device name. */
1638 	if (spawn->pf_bond <  0) {
1639 		/* Single device. */
1640 		if (!switch_info->representor)
1641 			strlcpy(name, dpdk_dev->name, sizeof(name));
1642 		else
1643 			snprintf(name, sizeof(name), "%s_representor_%u",
1644 				 dpdk_dev->name, switch_info->port_name);
1645 	} else {
1646 		/* Bonding device. */
1647 		if (!switch_info->representor)
1648 			snprintf(name, sizeof(name), "%s_%s",
1649 				 dpdk_dev->name, spawn->ibv_dev->name);
1650 		else
1651 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
1652 				 dpdk_dev->name, spawn->ibv_dev->name,
1653 				 switch_info->port_name);
1654 	}
1655 	/* check if the device is already spawned */
1656 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1657 		rte_errno = EEXIST;
1658 		return NULL;
1659 	}
1660 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1661 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1662 		eth_dev = rte_eth_dev_attach_secondary(name);
1663 		if (eth_dev == NULL) {
1664 			DRV_LOG(ERR, "can not attach rte ethdev");
1665 			rte_errno = ENOMEM;
1666 			return NULL;
1667 		}
1668 		eth_dev->device = dpdk_dev;
1669 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1670 		err = mlx5_proc_priv_init(eth_dev);
1671 		if (err)
1672 			return NULL;
1673 		/* Receive command fd from primary process */
1674 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1675 		if (err < 0)
1676 			return NULL;
1677 		/* Remap UAR for Tx queues. */
1678 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1679 		if (err)
1680 			return NULL;
1681 		/*
1682 		 * Ethdev pointer is still required as input since
1683 		 * the primary device is not accessible from the
1684 		 * secondary process.
1685 		 */
1686 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1687 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1688 		return eth_dev;
1689 	}
1690 	sh = mlx5_alloc_shared_ibctx(spawn);
1691 	if (!sh)
1692 		return NULL;
1693 	config.devx = sh->devx;
1694 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1695 	config.dest_tir = 1;
1696 #endif
1697 #ifdef HAVE_IBV_MLX5_MOD_SWP
1698 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1699 #endif
1700 	/*
1701 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1702 	 * as all ConnectX-5 devices.
1703 	 */
1704 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1705 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1706 #endif
1707 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1708 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1709 #endif
1710 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1711 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1712 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1713 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1714 			mps = MLX5_MPW_ENHANCED;
1715 		} else {
1716 			DRV_LOG(DEBUG, "MPW is supported");
1717 			mps = MLX5_MPW;
1718 		}
1719 	} else {
1720 		DRV_LOG(DEBUG, "MPW isn't supported");
1721 		mps = MLX5_MPW_DISABLED;
1722 	}
1723 #ifdef HAVE_IBV_MLX5_MOD_SWP
1724 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1725 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1726 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1727 #endif
1728 	config.swp = !!swp;
1729 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1730 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1731 		struct mlx5dv_striding_rq_caps mprq_caps =
1732 			dv_attr.striding_rq_caps;
1733 
1734 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1735 			mprq_caps.min_single_stride_log_num_of_bytes);
1736 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1737 			mprq_caps.max_single_stride_log_num_of_bytes);
1738 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1739 			mprq_caps.min_single_wqe_log_num_of_strides);
1740 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1741 			mprq_caps.max_single_wqe_log_num_of_strides);
1742 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1743 			mprq_caps.supported_qpts);
1744 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1745 		mprq = 1;
1746 		mprq_min_stride_size_n =
1747 			mprq_caps.min_single_stride_log_num_of_bytes;
1748 		mprq_max_stride_size_n =
1749 			mprq_caps.max_single_stride_log_num_of_bytes;
1750 		mprq_min_stride_num_n =
1751 			mprq_caps.min_single_wqe_log_num_of_strides;
1752 		mprq_max_stride_num_n =
1753 			mprq_caps.max_single_wqe_log_num_of_strides;
1754 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1755 						   mprq_min_stride_num_n);
1756 	}
1757 #endif
1758 	if (RTE_CACHE_LINE_SIZE == 128 &&
1759 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1760 		cqe_comp = 0;
1761 	else
1762 		cqe_comp = 1;
1763 	config.cqe_comp = cqe_comp;
1764 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1765 	/* Whether device supports 128B Rx CQE padding. */
1766 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1767 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1768 #endif
1769 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1770 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1771 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1772 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1773 			     (dv_attr.tunnel_offloads_caps &
1774 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1775 	}
1776 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1777 		tunnel_en ? "" : "not ");
1778 #else
1779 	DRV_LOG(WARNING,
1780 		"tunnel offloading disabled due to old OFED/rdma-core version");
1781 #endif
1782 	config.tunnel_en = tunnel_en;
1783 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1784 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1785 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1786 		   (dv_attr.tunnel_offloads_caps &
1787 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1788 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1789 		mpls_en ? "" : "not ");
1790 #else
1791 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1792 		" old OFED/rdma-core version or firmware configuration");
1793 #endif
1794 	config.mpls_en = mpls_en;
1795 	/* Check port status. */
1796 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1797 	if (err) {
1798 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1799 		goto error;
1800 	}
1801 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1802 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1803 		err = EINVAL;
1804 		goto error;
1805 	}
1806 	if (port_attr.state != IBV_PORT_ACTIVE)
1807 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1808 			mlx5_glue->port_state_str(port_attr.state),
1809 			port_attr.state);
1810 	/* Allocate private eth device data. */
1811 	priv = rte_zmalloc("ethdev private structure",
1812 			   sizeof(*priv),
1813 			   RTE_CACHE_LINE_SIZE);
1814 	if (priv == NULL) {
1815 		DRV_LOG(ERR, "priv allocation failure");
1816 		err = ENOMEM;
1817 		goto error;
1818 	}
1819 	priv->sh = sh;
1820 	priv->ibv_port = spawn->ibv_port;
1821 	priv->pci_dev = spawn->pci_dev;
1822 	priv->mtu = RTE_ETHER_MTU;
1823 #ifndef RTE_ARCH_64
1824 	/* Initialize UAR access locks for 32bit implementations. */
1825 	rte_spinlock_init(&priv->uar_lock_cq);
1826 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1827 		rte_spinlock_init(&priv->uar_lock[i]);
1828 #endif
1829 	/* Some internal functions rely on Netlink sockets, open them now. */
1830 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1831 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1832 	priv->nl_sn = 0;
1833 	priv->representor = !!switch_info->representor;
1834 	priv->master = !!switch_info->master;
1835 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1836 	priv->vport_meta_tag = 0;
1837 	priv->vport_meta_mask = 0;
1838 	priv->pf_bond = spawn->pf_bond;
1839 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1840 	/*
1841 	 * The DevX port query API is implemented. E-Switch may use
1842 	 * either vport or reg_c[0] metadata register to match on
1843 	 * vport index. The engaged part of metadata register is
1844 	 * defined by mask.
1845 	 */
1846 	devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1847 			      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1848 	err = mlx5dv_query_devx_port(sh->ctx, spawn->ibv_port, &devx_port);
1849 	if (err) {
1850 		DRV_LOG(WARNING, "can't query devx port %d on device %s\n",
1851 			spawn->ibv_port, spawn->ibv_dev->name);
1852 		devx_port.comp_mask = 0;
1853 	}
1854 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1855 		priv->vport_meta_tag = devx_port.reg_c_0.value;
1856 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
1857 		if (!priv->vport_meta_mask) {
1858 			DRV_LOG(ERR, "vport zero mask for port %d"
1859 				     " on bonding device %s\n",
1860 				     spawn->ibv_port, spawn->ibv_dev->name);
1861 			err = ENOTSUP;
1862 			goto error;
1863 		}
1864 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1865 			DRV_LOG(ERR, "invalid vport tag for port %d"
1866 				     " on bonding device %s\n",
1867 				     spawn->ibv_port, spawn->ibv_dev->name);
1868 			err = ENOTSUP;
1869 			goto error;
1870 		}
1871 	} else if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1872 		priv->vport_id = devx_port.vport_num;
1873 	} else if (spawn->pf_bond >= 0) {
1874 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1875 			     " on bonding device %s\n",
1876 			     spawn->ibv_port, spawn->ibv_dev->name);
1877 		err = ENOTSUP;
1878 		goto error;
1879 	} else {
1880 		/* Suppose vport index in compatible way. */
1881 		priv->vport_id = switch_info->representor ?
1882 				 switch_info->port_name + 1 : -1;
1883 	}
1884 #else
1885 	/*
1886 	 * Kernel/rdma_core support single E-Switch per PF configurations
1887 	 * only and vport_id field contains the vport index for
1888 	 * associated VF, which is deduced from representor port name.
1889 	 * For example, let's have the IB device port 10, it has
1890 	 * attached network device eth0, which has port name attribute
1891 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1892 	 * as 3 (2+1). This assigning schema should be changed if the
1893 	 * multiple E-Switch instances per PF configurations or/and PCI
1894 	 * subfunctions are added.
1895 	 */
1896 	priv->vport_id = switch_info->representor ?
1897 			 switch_info->port_name + 1 : -1;
1898 #endif
1899 	/* representor_id field keeps the unmodified VF index. */
1900 	priv->representor_id = switch_info->representor ?
1901 			       switch_info->port_name : -1;
1902 	/*
1903 	 * Look for sibling devices in order to reuse their switch domain
1904 	 * if any, otherwise allocate one.
1905 	 */
1906 	MLX5_ETH_FOREACH_DEV(port_id) {
1907 		const struct mlx5_priv *opriv =
1908 			rte_eth_devices[port_id].data->dev_private;
1909 
1910 		if (!opriv ||
1911 		    opriv->sh != priv->sh ||
1912 			opriv->domain_id ==
1913 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1914 			continue;
1915 		priv->domain_id = opriv->domain_id;
1916 		break;
1917 	}
1918 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1919 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1920 		if (err) {
1921 			err = rte_errno;
1922 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1923 				strerror(rte_errno));
1924 			goto error;
1925 		}
1926 		own_domain_id = 1;
1927 	}
1928 	err = mlx5_args(&config, dpdk_dev->devargs);
1929 	if (err) {
1930 		err = rte_errno;
1931 		DRV_LOG(ERR, "failed to process device arguments: %s",
1932 			strerror(rte_errno));
1933 		goto error;
1934 	}
1935 	err = mlx5_dev_check_sibling_config(priv, &config);
1936 	if (err)
1937 		goto error;
1938 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1939 			    IBV_DEVICE_RAW_IP_CSUM);
1940 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1941 		(config.hw_csum ? "" : "not "));
1942 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1943 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1944 	DRV_LOG(DEBUG, "counters are not supported");
1945 #endif
1946 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1947 	if (config.dv_flow_en) {
1948 		DRV_LOG(WARNING, "DV flow is not supported");
1949 		config.dv_flow_en = 0;
1950 	}
1951 #endif
1952 	config.ind_table_max_size =
1953 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1954 	/*
1955 	 * Remove this check once DPDK supports larger/variable
1956 	 * indirection tables.
1957 	 */
1958 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1959 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1960 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1961 		config.ind_table_max_size);
1962 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1963 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1964 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1965 		(config.hw_vlan_strip ? "" : "not "));
1966 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1967 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1968 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1969 		(config.hw_fcs_strip ? "" : "not "));
1970 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1971 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1972 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1973 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1974 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1975 #endif
1976 	if (config.hw_padding && !hw_padding) {
1977 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1978 		config.hw_padding = 0;
1979 	} else if (config.hw_padding) {
1980 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1981 	}
1982 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1983 		      (sh->device_attr.tso_caps.supported_qpts &
1984 		       (1 << IBV_QPT_RAW_PACKET)));
1985 	if (config.tso)
1986 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1987 	/*
1988 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1989 	 * by default.
1990 	 */
1991 	if (config.mps == MLX5_ARG_UNSET)
1992 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1993 							  MLX5_MPW_DISABLED;
1994 	else
1995 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1996 	DRV_LOG(INFO, "%sMPS is %s",
1997 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1998 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1999 	if (config.cqe_comp && !cqe_comp) {
2000 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2001 		config.cqe_comp = 0;
2002 	}
2003 	if (config.cqe_pad && !cqe_pad) {
2004 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2005 		config.cqe_pad = 0;
2006 	} else if (config.cqe_pad) {
2007 		DRV_LOG(INFO, "Rx CQE padding is enabled");
2008 	}
2009 	if (config.devx) {
2010 		priv->counter_fallback = 0;
2011 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2012 		if (err) {
2013 			err = -err;
2014 			goto error;
2015 		}
2016 		if (!config.hca_attr.flow_counters_dump)
2017 			priv->counter_fallback = 1;
2018 #ifndef HAVE_IBV_DEVX_ASYNC
2019 		priv->counter_fallback = 1;
2020 #endif
2021 		if (priv->counter_fallback)
2022 			DRV_LOG(INFO, "Use fall-back DV counter management\n");
2023 		/* Check for LRO support. */
2024 		if (config.dest_tir && config.hca_attr.lro_cap) {
2025 			/* TBD check tunnel lro caps. */
2026 			config.lro.supported = config.hca_attr.lro_cap;
2027 			DRV_LOG(DEBUG, "Device supports LRO");
2028 			/*
2029 			 * If LRO timeout is not configured by application,
2030 			 * use the minimal supported value.
2031 			 */
2032 			if (!config.lro.timeout)
2033 				config.lro.timeout =
2034 				config.hca_attr.lro_timer_supported_periods[0];
2035 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2036 				config.lro.timeout);
2037 		}
2038 	}
2039 	if (config.mprq.enabled && mprq) {
2040 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2041 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
2042 			config.mprq.stride_num_n =
2043 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2044 					mprq_min_stride_num_n);
2045 			DRV_LOG(WARNING,
2046 				"the number of strides"
2047 				" for Multi-Packet RQ is out of range,"
2048 				" setting default value (%u)",
2049 				1 << config.mprq.stride_num_n);
2050 		}
2051 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2052 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2053 	} else if (config.mprq.enabled && !mprq) {
2054 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2055 		config.mprq.enabled = 0;
2056 	}
2057 	if (config.max_dump_files_num == 0)
2058 		config.max_dump_files_num = 128;
2059 	eth_dev = rte_eth_dev_allocate(name);
2060 	if (eth_dev == NULL) {
2061 		DRV_LOG(ERR, "can not allocate rte ethdev");
2062 		err = ENOMEM;
2063 		goto error;
2064 	}
2065 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2066 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2067 	if (priv->representor) {
2068 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2069 		eth_dev->data->representor_id = priv->representor_id;
2070 	}
2071 	/*
2072 	 * Store associated network device interface index. This index
2073 	 * is permanent throughout the lifetime of device. So, we may store
2074 	 * the ifindex here and use the cached value further.
2075 	 */
2076 	assert(spawn->ifindex);
2077 	priv->if_index = spawn->ifindex;
2078 	eth_dev->data->dev_private = priv;
2079 	priv->dev_data = eth_dev->data;
2080 	eth_dev->data->mac_addrs = priv->mac;
2081 	eth_dev->device = dpdk_dev;
2082 	/* Configure the first MAC address by default. */
2083 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2084 		DRV_LOG(ERR,
2085 			"port %u cannot get MAC address, is mlx5_en"
2086 			" loaded? (errno: %s)",
2087 			eth_dev->data->port_id, strerror(rte_errno));
2088 		err = ENODEV;
2089 		goto error;
2090 	}
2091 	DRV_LOG(INFO,
2092 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2093 		eth_dev->data->port_id,
2094 		mac.addr_bytes[0], mac.addr_bytes[1],
2095 		mac.addr_bytes[2], mac.addr_bytes[3],
2096 		mac.addr_bytes[4], mac.addr_bytes[5]);
2097 #ifndef NDEBUG
2098 	{
2099 		char ifname[IF_NAMESIZE];
2100 
2101 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2102 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2103 				eth_dev->data->port_id, ifname);
2104 		else
2105 			DRV_LOG(DEBUG, "port %u ifname is unknown",
2106 				eth_dev->data->port_id);
2107 	}
2108 #endif
2109 	/* Get actual MTU if possible. */
2110 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
2111 	if (err) {
2112 		err = rte_errno;
2113 		goto error;
2114 	}
2115 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2116 		priv->mtu);
2117 	/* Initialize burst functions to prevent crashes before link-up. */
2118 	eth_dev->rx_pkt_burst = removed_rx_burst;
2119 	eth_dev->tx_pkt_burst = removed_tx_burst;
2120 	eth_dev->dev_ops = &mlx5_dev_ops;
2121 	/* Register MAC address. */
2122 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2123 	if (config.vf && config.vf_nl_en)
2124 		mlx5_nl_mac_addr_sync(eth_dev);
2125 	TAILQ_INIT(&priv->flows);
2126 	TAILQ_INIT(&priv->ctrl_flows);
2127 	/* Hint libmlx5 to use PMD allocator for data plane resources */
2128 	struct mlx5dv_ctx_allocators alctr = {
2129 		.alloc = &mlx5_alloc_verbs_buf,
2130 		.free = &mlx5_free_verbs_buf,
2131 		.data = priv,
2132 	};
2133 	mlx5_glue->dv_set_context_attr(sh->ctx,
2134 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2135 				       (void *)((uintptr_t)&alctr));
2136 	/* Bring Ethernet device up. */
2137 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2138 		eth_dev->data->port_id);
2139 	mlx5_set_link_up(eth_dev);
2140 	/*
2141 	 * Even though the interrupt handler is not installed yet,
2142 	 * interrupts will still trigger on the async_fd from
2143 	 * Verbs context returned by ibv_open_device().
2144 	 */
2145 	mlx5_link_update(eth_dev, 0);
2146 #ifdef HAVE_MLX5DV_DR_ESWITCH
2147 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2148 	      (switch_info->representor || switch_info->master)))
2149 		config.dv_esw_en = 0;
2150 #else
2151 	config.dv_esw_en = 0;
2152 #endif
2153 	/* Detect minimal data bytes to inline. */
2154 	mlx5_set_min_inline(spawn, &config);
2155 	/* Store device configuration on private structure. */
2156 	priv->config = config;
2157 	/* Create context for virtual machine VLAN workaround. */
2158 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2159 	if (config.dv_flow_en) {
2160 		err = mlx5_alloc_shared_dr(priv);
2161 		if (err)
2162 			goto error;
2163 	}
2164 	/* Supported Verbs flow priority number detection. */
2165 	err = mlx5_flow_discover_priorities(eth_dev);
2166 	if (err < 0) {
2167 		err = -err;
2168 		goto error;
2169 	}
2170 	priv->config.flow_prio = err;
2171 	return eth_dev;
2172 error:
2173 	if (priv) {
2174 		if (priv->sh)
2175 			mlx5_free_shared_dr(priv);
2176 		if (priv->nl_socket_route >= 0)
2177 			close(priv->nl_socket_route);
2178 		if (priv->nl_socket_rdma >= 0)
2179 			close(priv->nl_socket_rdma);
2180 		if (priv->vmwa_context)
2181 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
2182 		if (own_domain_id)
2183 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2184 		rte_free(priv);
2185 		if (eth_dev != NULL)
2186 			eth_dev->data->dev_private = NULL;
2187 	}
2188 	if (eth_dev != NULL) {
2189 		/* mac_addrs must not be freed alone because part of dev_private */
2190 		eth_dev->data->mac_addrs = NULL;
2191 		rte_eth_dev_release_port(eth_dev);
2192 	}
2193 	if (sh)
2194 		mlx5_free_shared_ibctx(sh);
2195 	assert(err > 0);
2196 	rte_errno = err;
2197 	return NULL;
2198 }
2199 
2200 /**
2201  * Comparison callback to sort device data.
2202  *
2203  * This is meant to be used with qsort().
2204  *
2205  * @param a[in]
2206  *   Pointer to pointer to first data object.
2207  * @param b[in]
2208  *   Pointer to pointer to second data object.
2209  *
2210  * @return
2211  *   0 if both objects are equal, less than 0 if the first argument is less
2212  *   than the second, greater than 0 otherwise.
2213  */
2214 static int
2215 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2216 {
2217 	const struct mlx5_switch_info *si_a =
2218 		&((const struct mlx5_dev_spawn_data *)a)->info;
2219 	const struct mlx5_switch_info *si_b =
2220 		&((const struct mlx5_dev_spawn_data *)b)->info;
2221 	int ret;
2222 
2223 	/* Master device first. */
2224 	ret = si_b->master - si_a->master;
2225 	if (ret)
2226 		return ret;
2227 	/* Then representor devices. */
2228 	ret = si_b->representor - si_a->representor;
2229 	if (ret)
2230 		return ret;
2231 	/* Unidentified devices come last in no specific order. */
2232 	if (!si_a->representor)
2233 		return 0;
2234 	/* Order representors by name. */
2235 	return si_a->port_name - si_b->port_name;
2236 }
2237 
2238 /**
2239  * Match PCI information for possible slaves of bonding device.
2240  *
2241  * @param[in] ibv_dev
2242  *   Pointer to Infiniband device structure.
2243  * @param[in] pci_dev
2244  *   Pointer to PCI device structure to match PCI address.
2245  * @param[in] nl_rdma
2246  *   Netlink RDMA group socket handle.
2247  *
2248  * @return
2249  *   negative value if no bonding device found, otherwise
2250  *   positive index of slave PF in bonding.
2251  */
2252 static int
2253 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2254 			   const struct rte_pci_device *pci_dev,
2255 			   int nl_rdma)
2256 {
2257 	char ifname[IF_NAMESIZE + 1];
2258 	unsigned int ifindex;
2259 	unsigned int np, i;
2260 	FILE *file = NULL;
2261 	int pf = -1;
2262 
2263 	/*
2264 	 * Try to get master device name. If something goes
2265 	 * wrong suppose the lack of kernel support and no
2266 	 * bonding devices.
2267 	 */
2268 	if (nl_rdma < 0)
2269 		return -1;
2270 	if (!strstr(ibv_dev->name, "bond"))
2271 		return -1;
2272 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2273 	if (!np)
2274 		return -1;
2275 	/*
2276 	 * The Master device might not be on the predefined
2277 	 * port (not on port index 1, it is not garanted),
2278 	 * we have to scan all Infiniband device port and
2279 	 * find master.
2280 	 */
2281 	for (i = 1; i <= np; ++i) {
2282 		/* Check whether Infiniband port is populated. */
2283 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2284 		if (!ifindex)
2285 			continue;
2286 		if (!if_indextoname(ifindex, ifname))
2287 			continue;
2288 		/* Try to read bonding slave names from sysfs. */
2289 		MKSTR(slaves,
2290 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2291 		file = fopen(slaves, "r");
2292 		if (file)
2293 			break;
2294 	}
2295 	if (!file)
2296 		return -1;
2297 	/* Use safe format to check maximal buffer length. */
2298 	assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2299 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2300 		char tmp_str[IF_NAMESIZE + 32];
2301 		struct rte_pci_addr pci_addr;
2302 		struct mlx5_switch_info	info;
2303 
2304 		/* Process slave interface names in the loop. */
2305 		snprintf(tmp_str, sizeof(tmp_str),
2306 			 "/sys/class/net/%s", ifname);
2307 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2308 			DRV_LOG(WARNING, "can not get PCI address"
2309 					 " for netdev \"%s\"", ifname);
2310 			continue;
2311 		}
2312 		if (pci_dev->addr.domain != pci_addr.domain ||
2313 		    pci_dev->addr.bus != pci_addr.bus ||
2314 		    pci_dev->addr.devid != pci_addr.devid ||
2315 		    pci_dev->addr.function != pci_addr.function)
2316 			continue;
2317 		/* Slave interface PCI address match found. */
2318 		fclose(file);
2319 		snprintf(tmp_str, sizeof(tmp_str),
2320 			 "/sys/class/net/%s/phys_port_name", ifname);
2321 		file = fopen(tmp_str, "rb");
2322 		if (!file)
2323 			break;
2324 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2325 		if (fscanf(file, "%32s", tmp_str) == 1)
2326 			mlx5_translate_port_name(tmp_str, &info);
2327 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2328 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2329 			pf = info.port_name;
2330 		break;
2331 	}
2332 	if (file)
2333 		fclose(file);
2334 	return pf;
2335 }
2336 
2337 /**
2338  * DPDK callback to register a PCI device.
2339  *
2340  * This function spawns Ethernet devices out of a given PCI device.
2341  *
2342  * @param[in] pci_drv
2343  *   PCI driver structure (mlx5_driver).
2344  * @param[in] pci_dev
2345  *   PCI device information.
2346  *
2347  * @return
2348  *   0 on success, a negative errno value otherwise and rte_errno is set.
2349  */
2350 static int
2351 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2352 	       struct rte_pci_device *pci_dev)
2353 {
2354 	struct ibv_device **ibv_list;
2355 	/*
2356 	 * Number of found IB Devices matching with requested PCI BDF.
2357 	 * nd != 1 means there are multiple IB devices over the same
2358 	 * PCI device and we have representors and master.
2359 	 */
2360 	unsigned int nd = 0;
2361 	/*
2362 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2363 	 * we have the single multiport IB device, and there may be
2364 	 * representors attached to some of found ports.
2365 	 */
2366 	unsigned int np = 0;
2367 	/*
2368 	 * Number of DPDK ethernet devices to Spawn - either over
2369 	 * multiple IB devices or multiple ports of single IB device.
2370 	 * Actually this is the number of iterations to spawn.
2371 	 */
2372 	unsigned int ns = 0;
2373 	/*
2374 	 * Bonding device
2375 	 *   < 0 - no bonding device (single one)
2376 	 *  >= 0 - bonding device (value is slave PF index)
2377 	 */
2378 	int bd = -1;
2379 	struct mlx5_dev_spawn_data *list = NULL;
2380 	struct mlx5_dev_config dev_config;
2381 	int ret;
2382 
2383 	ret = mlx5_init_once();
2384 	if (ret) {
2385 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2386 			strerror(rte_errno));
2387 		return -rte_errno;
2388 	}
2389 	assert(pci_drv == &mlx5_driver);
2390 	errno = 0;
2391 	ibv_list = mlx5_glue->get_device_list(&ret);
2392 	if (!ibv_list) {
2393 		rte_errno = errno ? errno : ENOSYS;
2394 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2395 		return -rte_errno;
2396 	}
2397 	/*
2398 	 * First scan the list of all Infiniband devices to find
2399 	 * matching ones, gathering into the list.
2400 	 */
2401 	struct ibv_device *ibv_match[ret + 1];
2402 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2403 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2404 	unsigned int i;
2405 
2406 	while (ret-- > 0) {
2407 		struct rte_pci_addr pci_addr;
2408 
2409 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2410 		bd = mlx5_device_bond_pci_match
2411 				(ibv_list[ret], pci_dev, nl_rdma);
2412 		if (bd >= 0) {
2413 			/*
2414 			 * Bonding device detected. Only one match is allowed,
2415 			 * the bonding is supported over multi-port IB device,
2416 			 * there should be no matches on representor PCI
2417 			 * functions or non VF LAG bonding devices with
2418 			 * specified address.
2419 			 */
2420 			if (nd) {
2421 				DRV_LOG(ERR,
2422 					"multiple PCI match on bonding device"
2423 					"\"%s\" found", ibv_list[ret]->name);
2424 				rte_errno = ENOENT;
2425 				ret = -rte_errno;
2426 				goto exit;
2427 			}
2428 			DRV_LOG(INFO, "PCI information matches for"
2429 				      " slave %d bonding device \"%s\"",
2430 				      bd, ibv_list[ret]->name);
2431 			ibv_match[nd++] = ibv_list[ret];
2432 			break;
2433 		}
2434 		if (mlx5_dev_to_pci_addr
2435 			(ibv_list[ret]->ibdev_path, &pci_addr))
2436 			continue;
2437 		if (pci_dev->addr.domain != pci_addr.domain ||
2438 		    pci_dev->addr.bus != pci_addr.bus ||
2439 		    pci_dev->addr.devid != pci_addr.devid ||
2440 		    pci_dev->addr.function != pci_addr.function)
2441 			continue;
2442 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2443 			ibv_list[ret]->name);
2444 		ibv_match[nd++] = ibv_list[ret];
2445 	}
2446 	ibv_match[nd] = NULL;
2447 	if (!nd) {
2448 		/* No device matches, just complain and bail out. */
2449 		DRV_LOG(WARNING,
2450 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2451 			" are kernel drivers loaded?",
2452 			pci_dev->addr.domain, pci_dev->addr.bus,
2453 			pci_dev->addr.devid, pci_dev->addr.function);
2454 		rte_errno = ENOENT;
2455 		ret = -rte_errno;
2456 		goto exit;
2457 	}
2458 	if (nd == 1) {
2459 		/*
2460 		 * Found single matching device may have multiple ports.
2461 		 * Each port may be representor, we have to check the port
2462 		 * number and check the representors existence.
2463 		 */
2464 		if (nl_rdma >= 0)
2465 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2466 		if (!np)
2467 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2468 					 " ports number", ibv_match[0]->name);
2469 		if (bd >= 0 && !np) {
2470 			DRV_LOG(ERR, "can not get ports"
2471 				     " for bonding device");
2472 			rte_errno = ENOENT;
2473 			ret = -rte_errno;
2474 			goto exit;
2475 		}
2476 	}
2477 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2478 	if (bd >= 0) {
2479 		/*
2480 		 * This may happen if there is VF LAG kernel support and
2481 		 * application is compiled with older rdma_core library.
2482 		 */
2483 		DRV_LOG(ERR,
2484 			"No kernel/verbs support for VF LAG bonding found.");
2485 		rte_errno = ENOTSUP;
2486 		ret = -rte_errno;
2487 		goto exit;
2488 	}
2489 #endif
2490 	/*
2491 	 * Now we can determine the maximal
2492 	 * amount of devices to be spawned.
2493 	 */
2494 	list = rte_zmalloc("device spawn data",
2495 			 sizeof(struct mlx5_dev_spawn_data) *
2496 			 (np ? np : nd),
2497 			 RTE_CACHE_LINE_SIZE);
2498 	if (!list) {
2499 		DRV_LOG(ERR, "spawn data array allocation failure");
2500 		rte_errno = ENOMEM;
2501 		ret = -rte_errno;
2502 		goto exit;
2503 	}
2504 	if (bd >= 0 || np > 1) {
2505 		/*
2506 		 * Single IB device with multiple ports found,
2507 		 * it may be E-Switch master device and representors.
2508 		 * We have to perform identification trough the ports.
2509 		 */
2510 		assert(nl_rdma >= 0);
2511 		assert(ns == 0);
2512 		assert(nd == 1);
2513 		assert(np);
2514 		for (i = 1; i <= np; ++i) {
2515 			list[ns].max_port = np;
2516 			list[ns].ibv_port = i;
2517 			list[ns].ibv_dev = ibv_match[0];
2518 			list[ns].eth_dev = NULL;
2519 			list[ns].pci_dev = pci_dev;
2520 			list[ns].pf_bond = bd;
2521 			list[ns].ifindex = mlx5_nl_ifindex
2522 					(nl_rdma, list[ns].ibv_dev->name, i);
2523 			if (!list[ns].ifindex) {
2524 				/*
2525 				 * No network interface index found for the
2526 				 * specified port, it means there is no
2527 				 * representor on this port. It's OK,
2528 				 * there can be disabled ports, for example
2529 				 * if sriov_numvfs < sriov_totalvfs.
2530 				 */
2531 				continue;
2532 			}
2533 			ret = -1;
2534 			if (nl_route >= 0)
2535 				ret = mlx5_nl_switch_info
2536 					       (nl_route,
2537 						list[ns].ifindex,
2538 						&list[ns].info);
2539 			if (ret || (!list[ns].info.representor &&
2540 				    !list[ns].info.master)) {
2541 				/*
2542 				 * We failed to recognize representors with
2543 				 * Netlink, let's try to perform the task
2544 				 * with sysfs.
2545 				 */
2546 				ret =  mlx5_sysfs_switch_info
2547 						(list[ns].ifindex,
2548 						 &list[ns].info);
2549 			}
2550 			if (!ret && bd >= 0) {
2551 				switch (list[ns].info.name_type) {
2552 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2553 					if (list[ns].info.port_name == bd)
2554 						ns++;
2555 					break;
2556 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2557 					if (list[ns].info.pf_num == bd)
2558 						ns++;
2559 					break;
2560 				default:
2561 					break;
2562 				}
2563 				continue;
2564 			}
2565 			if (!ret && (list[ns].info.representor ^
2566 				     list[ns].info.master))
2567 				ns++;
2568 		}
2569 		if (!ns) {
2570 			DRV_LOG(ERR,
2571 				"unable to recognize master/representors"
2572 				" on the IB device with multiple ports");
2573 			rte_errno = ENOENT;
2574 			ret = -rte_errno;
2575 			goto exit;
2576 		}
2577 	} else {
2578 		/*
2579 		 * The existence of several matching entries (nd > 1) means
2580 		 * port representors have been instantiated. No existing Verbs
2581 		 * call nor sysfs entries can tell them apart, this can only
2582 		 * be done through Netlink calls assuming kernel drivers are
2583 		 * recent enough to support them.
2584 		 *
2585 		 * In the event of identification failure through Netlink,
2586 		 * try again through sysfs, then:
2587 		 *
2588 		 * 1. A single IB device matches (nd == 1) with single
2589 		 *    port (np=0/1) and is not a representor, assume
2590 		 *    no switch support.
2591 		 *
2592 		 * 2. Otherwise no safe assumptions can be made;
2593 		 *    complain louder and bail out.
2594 		 */
2595 		np = 1;
2596 		for (i = 0; i != nd; ++i) {
2597 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2598 			list[ns].max_port = 1;
2599 			list[ns].ibv_port = 1;
2600 			list[ns].ibv_dev = ibv_match[i];
2601 			list[ns].eth_dev = NULL;
2602 			list[ns].pci_dev = pci_dev;
2603 			list[ns].pf_bond = -1;
2604 			list[ns].ifindex = 0;
2605 			if (nl_rdma >= 0)
2606 				list[ns].ifindex = mlx5_nl_ifindex
2607 					(nl_rdma, list[ns].ibv_dev->name, 1);
2608 			if (!list[ns].ifindex) {
2609 				char ifname[IF_NAMESIZE];
2610 
2611 				/*
2612 				 * Netlink failed, it may happen with old
2613 				 * ib_core kernel driver (before 4.16).
2614 				 * We can assume there is old driver because
2615 				 * here we are processing single ports IB
2616 				 * devices. Let's try sysfs to retrieve
2617 				 * the ifindex. The method works for
2618 				 * master device only.
2619 				 */
2620 				if (nd > 1) {
2621 					/*
2622 					 * Multiple devices found, assume
2623 					 * representors, can not distinguish
2624 					 * master/representor and retrieve
2625 					 * ifindex via sysfs.
2626 					 */
2627 					continue;
2628 				}
2629 				ret = mlx5_get_master_ifname
2630 					(ibv_match[i]->ibdev_path, &ifname);
2631 				if (!ret)
2632 					list[ns].ifindex =
2633 						if_nametoindex(ifname);
2634 				if (!list[ns].ifindex) {
2635 					/*
2636 					 * No network interface index found
2637 					 * for the specified device, it means
2638 					 * there it is neither representor
2639 					 * nor master.
2640 					 */
2641 					continue;
2642 				}
2643 			}
2644 			ret = -1;
2645 			if (nl_route >= 0)
2646 				ret = mlx5_nl_switch_info
2647 					       (nl_route,
2648 						list[ns].ifindex,
2649 						&list[ns].info);
2650 			if (ret || (!list[ns].info.representor &&
2651 				    !list[ns].info.master)) {
2652 				/*
2653 				 * We failed to recognize representors with
2654 				 * Netlink, let's try to perform the task
2655 				 * with sysfs.
2656 				 */
2657 				ret =  mlx5_sysfs_switch_info
2658 						(list[ns].ifindex,
2659 						 &list[ns].info);
2660 			}
2661 			if (!ret && (list[ns].info.representor ^
2662 				     list[ns].info.master)) {
2663 				ns++;
2664 			} else if ((nd == 1) &&
2665 				   !list[ns].info.representor &&
2666 				   !list[ns].info.master) {
2667 				/*
2668 				 * Single IB device with
2669 				 * one physical port and
2670 				 * attached network device.
2671 				 * May be SRIOV is not enabled
2672 				 * or there is no representors.
2673 				 */
2674 				DRV_LOG(INFO, "no E-Switch support detected");
2675 				ns++;
2676 				break;
2677 			}
2678 		}
2679 		if (!ns) {
2680 			DRV_LOG(ERR,
2681 				"unable to recognize master/representors"
2682 				" on the multiple IB devices");
2683 			rte_errno = ENOENT;
2684 			ret = -rte_errno;
2685 			goto exit;
2686 		}
2687 	}
2688 	assert(ns);
2689 	/*
2690 	 * Sort list to probe devices in natural order for users convenience
2691 	 * (i.e. master first, then representors from lowest to highest ID).
2692 	 */
2693 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2694 	/* Default configuration. */
2695 	dev_config = (struct mlx5_dev_config){
2696 		.hw_padding = 0,
2697 		.mps = MLX5_ARG_UNSET,
2698 		.rx_vec_en = 1,
2699 		.txq_inline_max = MLX5_ARG_UNSET,
2700 		.txq_inline_min = MLX5_ARG_UNSET,
2701 		.txq_inline_mpw = MLX5_ARG_UNSET,
2702 		.txqs_inline = MLX5_ARG_UNSET,
2703 		.vf_nl_en = 1,
2704 		.mr_ext_memseg_en = 1,
2705 		.mprq = {
2706 			.enabled = 0, /* Disabled by default. */
2707 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
2708 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
2709 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
2710 		},
2711 		.dv_esw_en = 1,
2712 	};
2713 	/* Device specific configuration. */
2714 	switch (pci_dev->id.device_id) {
2715 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2716 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2717 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2718 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2719 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2720 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2721 		dev_config.vf = 1;
2722 		break;
2723 	default:
2724 		break;
2725 	}
2726 	for (i = 0; i != ns; ++i) {
2727 		uint32_t restore;
2728 
2729 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2730 						 &list[i],
2731 						 dev_config);
2732 		if (!list[i].eth_dev) {
2733 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2734 				break;
2735 			/* Device is disabled or already spawned. Ignore it. */
2736 			continue;
2737 		}
2738 		restore = list[i].eth_dev->data->dev_flags;
2739 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2740 		/* Restore non-PCI flags cleared by the above call. */
2741 		list[i].eth_dev->data->dev_flags |= restore;
2742 		rte_eth_dev_probing_finish(list[i].eth_dev);
2743 	}
2744 	if (i != ns) {
2745 		DRV_LOG(ERR,
2746 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2747 			" encountering an error: %s",
2748 			pci_dev->addr.domain, pci_dev->addr.bus,
2749 			pci_dev->addr.devid, pci_dev->addr.function,
2750 			strerror(rte_errno));
2751 		ret = -rte_errno;
2752 		/* Roll back. */
2753 		while (i--) {
2754 			if (!list[i].eth_dev)
2755 				continue;
2756 			mlx5_dev_close(list[i].eth_dev);
2757 			/* mac_addrs must not be freed because in dev_private */
2758 			list[i].eth_dev->data->mac_addrs = NULL;
2759 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2760 		}
2761 		/* Restore original error. */
2762 		rte_errno = -ret;
2763 	} else {
2764 		ret = 0;
2765 	}
2766 exit:
2767 	/*
2768 	 * Do the routine cleanup:
2769 	 * - close opened Netlink sockets
2770 	 * - free allocated spawn data array
2771 	 * - free the Infiniband device list
2772 	 */
2773 	if (nl_rdma >= 0)
2774 		close(nl_rdma);
2775 	if (nl_route >= 0)
2776 		close(nl_route);
2777 	if (list)
2778 		rte_free(list);
2779 	assert(ibv_list);
2780 	mlx5_glue->free_device_list(ibv_list);
2781 	return ret;
2782 }
2783 
2784 uint16_t
2785 mlx5_eth_find_next(uint16_t port_id)
2786 {
2787 	while (port_id < RTE_MAX_ETHPORTS) {
2788 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2789 
2790 		if (dev->state != RTE_ETH_DEV_UNUSED &&
2791 		    dev->device &&
2792 		    dev->device->driver &&
2793 		    dev->device->driver->name &&
2794 		    !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))
2795 			break;
2796 		port_id++;
2797 	}
2798 	if (port_id >= RTE_MAX_ETHPORTS)
2799 		return RTE_MAX_ETHPORTS;
2800 	return port_id;
2801 }
2802 
2803 /**
2804  * DPDK callback to remove a PCI device.
2805  *
2806  * This function removes all Ethernet devices belong to a given PCI device.
2807  *
2808  * @param[in] pci_dev
2809  *   Pointer to the PCI device.
2810  *
2811  * @return
2812  *   0 on success, the function cannot fail.
2813  */
2814 static int
2815 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2816 {
2817 	uint16_t port_id;
2818 
2819 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2820 		rte_eth_dev_close(port_id);
2821 	return 0;
2822 }
2823 
2824 static const struct rte_pci_id mlx5_pci_id_map[] = {
2825 	{
2826 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2827 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2828 	},
2829 	{
2830 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2831 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2832 	},
2833 	{
2834 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2835 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2836 	},
2837 	{
2838 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2839 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2840 	},
2841 	{
2842 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2843 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2844 	},
2845 	{
2846 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2847 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2848 	},
2849 	{
2850 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2851 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2852 	},
2853 	{
2854 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2855 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2856 	},
2857 	{
2858 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2859 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2860 	},
2861 	{
2862 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2863 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2864 	},
2865 	{
2866 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2867 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2868 	},
2869 	{
2870 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2871 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2872 	},
2873 	{
2874 		.vendor_id = 0
2875 	}
2876 };
2877 
2878 static struct rte_pci_driver mlx5_driver = {
2879 	.driver = {
2880 		.name = MLX5_DRIVER_NAME
2881 	},
2882 	.id_table = mlx5_pci_id_map,
2883 	.probe = mlx5_pci_probe,
2884 	.remove = mlx5_pci_remove,
2885 	.dma_map = mlx5_dma_map,
2886 	.dma_unmap = mlx5_dma_unmap,
2887 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2888 		     RTE_PCI_DRV_PROBE_AGAIN,
2889 };
2890 
2891 #ifdef RTE_IBVERBS_LINK_DLOPEN
2892 
2893 /**
2894  * Suffix RTE_EAL_PMD_PATH with "-glue".
2895  *
2896  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2897  * suffixing its last component.
2898  *
2899  * @param buf[out]
2900  *   Output buffer, should be large enough otherwise NULL is returned.
2901  * @param size
2902  *   Size of @p out.
2903  *
2904  * @return
2905  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2906  */
2907 static char *
2908 mlx5_glue_path(char *buf, size_t size)
2909 {
2910 	static const char *const bad[] = { "/", ".", "..", NULL };
2911 	const char *path = RTE_EAL_PMD_PATH;
2912 	size_t len = strlen(path);
2913 	size_t off;
2914 	int i;
2915 
2916 	while (len && path[len - 1] == '/')
2917 		--len;
2918 	for (off = len; off && path[off - 1] != '/'; --off)
2919 		;
2920 	for (i = 0; bad[i]; ++i)
2921 		if (!strncmp(path + off, bad[i], (int)(len - off)))
2922 			goto error;
2923 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2924 	if (i == -1 || (size_t)i >= size)
2925 		goto error;
2926 	return buf;
2927 error:
2928 	DRV_LOG(ERR,
2929 		"unable to append \"-glue\" to last component of"
2930 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2931 		" please re-configure DPDK");
2932 	return NULL;
2933 }
2934 
2935 /**
2936  * Initialization routine for run-time dependency on rdma-core.
2937  */
2938 static int
2939 mlx5_glue_init(void)
2940 {
2941 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2942 	const char *path[] = {
2943 		/*
2944 		 * A basic security check is necessary before trusting
2945 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2946 		 */
2947 		(geteuid() == getuid() && getegid() == getgid() ?
2948 		 getenv("MLX5_GLUE_PATH") : NULL),
2949 		/*
2950 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2951 		 * variant, otherwise let dlopen() look up libraries on its
2952 		 * own.
2953 		 */
2954 		(*RTE_EAL_PMD_PATH ?
2955 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2956 	};
2957 	unsigned int i = 0;
2958 	void *handle = NULL;
2959 	void **sym;
2960 	const char *dlmsg;
2961 
2962 	while (!handle && i != RTE_DIM(path)) {
2963 		const char *end;
2964 		size_t len;
2965 		int ret;
2966 
2967 		if (!path[i]) {
2968 			++i;
2969 			continue;
2970 		}
2971 		end = strpbrk(path[i], ":;");
2972 		if (!end)
2973 			end = path[i] + strlen(path[i]);
2974 		len = end - path[i];
2975 		ret = 0;
2976 		do {
2977 			char name[ret + 1];
2978 
2979 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2980 				       (int)len, path[i],
2981 				       (!len || *(end - 1) == '/') ? "" : "/");
2982 			if (ret == -1)
2983 				break;
2984 			if (sizeof(name) != (size_t)ret + 1)
2985 				continue;
2986 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2987 				name);
2988 			handle = dlopen(name, RTLD_LAZY);
2989 			break;
2990 		} while (1);
2991 		path[i] = end + 1;
2992 		if (!*end)
2993 			++i;
2994 	}
2995 	if (!handle) {
2996 		rte_errno = EINVAL;
2997 		dlmsg = dlerror();
2998 		if (dlmsg)
2999 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
3000 		goto glue_error;
3001 	}
3002 	sym = dlsym(handle, "mlx5_glue");
3003 	if (!sym || !*sym) {
3004 		rte_errno = EINVAL;
3005 		dlmsg = dlerror();
3006 		if (dlmsg)
3007 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
3008 		goto glue_error;
3009 	}
3010 	mlx5_glue = *sym;
3011 	return 0;
3012 glue_error:
3013 	if (handle)
3014 		dlclose(handle);
3015 	DRV_LOG(WARNING,
3016 		"cannot initialize PMD due to missing run-time dependency on"
3017 		" rdma-core libraries (libibverbs, libmlx5)");
3018 	return -rte_errno;
3019 }
3020 
3021 #endif
3022 
3023 /**
3024  * Driver initialization routine.
3025  */
3026 RTE_INIT(rte_mlx5_pmd_init)
3027 {
3028 	/* Initialize driver log type. */
3029 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
3030 	if (mlx5_logtype >= 0)
3031 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3032 
3033 	/* Build the static tables for Verbs conversion. */
3034 	mlx5_set_ptype_table();
3035 	mlx5_set_cksum_table();
3036 	mlx5_set_swp_types_table();
3037 	/*
3038 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3039 	 * huge pages. Calling ibv_fork_init() during init allows
3040 	 * applications to use fork() safely for purposes other than
3041 	 * using this PMD, which is not supported in forked processes.
3042 	 */
3043 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3044 	/* Match the size of Rx completion entry to the size of a cacheline. */
3045 	if (RTE_CACHE_LINE_SIZE == 128)
3046 		setenv("MLX5_CQE_SIZE", "128", 0);
3047 	/*
3048 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
3049 	 * cleanup all the Verbs resources even when the device was removed.
3050 	 */
3051 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
3052 #ifdef RTE_IBVERBS_LINK_DLOPEN
3053 	if (mlx5_glue_init())
3054 		return;
3055 	assert(mlx5_glue);
3056 #endif
3057 #ifndef NDEBUG
3058 	/* Glue structure must not contain any NULL pointers. */
3059 	{
3060 		unsigned int i;
3061 
3062 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
3063 			assert(((const void *const *)mlx5_glue)[i]);
3064 	}
3065 #endif
3066 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
3067 		DRV_LOG(ERR,
3068 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
3069 			mlx5_glue->version, MLX5_GLUE_VERSION);
3070 		return;
3071 	}
3072 	mlx5_glue->fork_init();
3073 	rte_pci_register(&mlx5_driver);
3074 }
3075 
3076 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3077 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3078 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
3079