xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 54ad947eda42042d2bdae69b57d0c7c8e291d9ec)
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_eal_memconfig.h>
36 #include <rte_kvargs.h>
37 #include <rte_rwlock.h>
38 #include <rte_spinlock.h>
39 #include <rte_string_fns.h>
40 #include <rte_alarm.h>
41 
42 #include "mlx5.h"
43 #include "mlx5_utils.h"
44 #include "mlx5_rxtx.h"
45 #include "mlx5_autoconf.h"
46 #include "mlx5_defs.h"
47 #include "mlx5_glue.h"
48 #include "mlx5_mr.h"
49 #include "mlx5_flow.h"
50 
51 /* Device parameter to enable RX completion queue compression. */
52 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
53 
54 /* Device parameter to enable RX completion entry padding to 128B. */
55 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
56 
57 /* Device parameter to enable padding Rx packet to cacheline size. */
58 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
59 
60 /* Device parameter to enable Multi-Packet Rx queue. */
61 #define MLX5_RX_MPRQ_EN "mprq_en"
62 
63 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
64 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
65 
66 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
67 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
68 
69 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
70 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
71 
72 /* Device parameter to configure inline send. Deprecated, ignored.*/
73 #define MLX5_TXQ_INLINE "txq_inline"
74 
75 /* Device parameter to limit packet size to inline with ordinary SEND. */
76 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
77 
78 /* Device parameter to configure minimal data size to inline. */
79 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
80 
81 /* Device parameter to limit packet size to inline with Enhanced MPW. */
82 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
83 
84 /*
85  * Device parameter to configure the number of TX queues threshold for
86  * enabling inline send.
87  */
88 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
89 
90 /*
91  * Device parameter to configure the number of TX queues threshold for
92  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
93  */
94 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
95 
96 /* Device parameter to enable multi-packet send WQEs. */
97 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
98 
99 /*
100  * Device parameter to include 2 dsegs in the title WQEBB.
101  * Deprecated, ignored.
102  */
103 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
104 
105 /*
106  * Device parameter to limit the size of inlining packet.
107  * Deprecated, ignored.
108  */
109 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
110 
111 /*
112  * Device parameter to enable hardware Tx vector.
113  * Deprecated, ignored (no vectorized Tx routines anymore).
114  */
115 #define MLX5_TX_VEC_EN "tx_vec_en"
116 
117 /* Device parameter to enable hardware Rx vector. */
118 #define MLX5_RX_VEC_EN "rx_vec_en"
119 
120 /* Allow L3 VXLAN flow creation. */
121 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
122 
123 /* Activate DV E-Switch flow steering. */
124 #define MLX5_DV_ESW_EN "dv_esw_en"
125 
126 /* Activate DV flow steering. */
127 #define MLX5_DV_FLOW_EN "dv_flow_en"
128 
129 /* Activate Netlink support in VF mode. */
130 #define MLX5_VF_NL_EN "vf_nl_en"
131 
132 /* Enable extending memsegs when creating a MR. */
133 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
134 
135 /* Select port representors to instantiate. */
136 #define MLX5_REPRESENTOR "representor"
137 
138 /* Device parameter to configure the maximum number of dump files per queue. */
139 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
140 
141 /* Configure timeout of LRO session (in microseconds). */
142 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
143 
144 #ifndef HAVE_IBV_MLX5_MOD_MPW
145 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
146 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
147 #endif
148 
149 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
150 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
151 #endif
152 
153 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
154 
155 /* Shared memory between primary and secondary processes. */
156 struct mlx5_shared_data *mlx5_shared_data;
157 
158 /* Spinlock for mlx5_shared_data allocation. */
159 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
160 
161 /* Process local data for secondary processes. */
162 static struct mlx5_local_data mlx5_local_data;
163 
164 /** Driver-specific log messages type. */
165 int mlx5_logtype;
166 
167 /** Data associated with devices to spawn. */
168 struct mlx5_dev_spawn_data {
169 	uint32_t ifindex; /**< Network interface index. */
170 	uint32_t max_port; /**< IB device maximal port index. */
171 	uint32_t ibv_port; /**< IB device physical port index. */
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 	sh->pci_dev = spawn->pci_dev;
377 	pthread_mutex_init(&sh->intr_mutex, NULL);
378 	/*
379 	 * Setting port_id to max unallowed value means
380 	 * there is no interrupt subhandler installed for
381 	 * the given port index i.
382 	 */
383 	for (i = 0; i < sh->max_port; i++)
384 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
385 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
386 	if (sh->pd == NULL) {
387 		DRV_LOG(ERR, "PD allocation failure");
388 		err = ENOMEM;
389 		goto error;
390 	}
391 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
392 	err = mlx5_get_pdn(sh->pd, &sh->pdn);
393 	if (err) {
394 		DRV_LOG(ERR, "Fail to extract pdn from PD");
395 		goto error;
396 	}
397 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
398 	/*
399 	 * Once the device is added to the list of memory event
400 	 * callback, its global MR cache table cannot be expanded
401 	 * on the fly because of deadlock. If it overflows, lookup
402 	 * should be done by searching MR list linearly, which is slow.
403 	 *
404 	 * At this point the device is not added to the memory
405 	 * event list yet, context is just being created.
406 	 */
407 	err = mlx5_mr_btree_init(&sh->mr.cache,
408 				 MLX5_MR_BTREE_CACHE_N * 2,
409 				 sh->pci_dev->device.numa_node);
410 	if (err) {
411 		err = rte_errno;
412 		goto error;
413 	}
414 	mlx5_flow_counters_mng_init(sh);
415 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
416 exit:
417 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
418 	return sh;
419 error:
420 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
421 	assert(sh);
422 	if (sh->pd)
423 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
424 	if (sh->ctx)
425 		claim_zero(mlx5_glue->close_device(sh->ctx));
426 	rte_free(sh);
427 	assert(err > 0);
428 	rte_errno = err;
429 	return NULL;
430 }
431 
432 /**
433  * Free shared IB device context. Decrement counter and if zero free
434  * all allocated resources and close handles.
435  *
436  * @param[in] sh
437  *   Pointer to mlx5_ibv_shared object to free
438  */
439 static void
440 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
441 {
442 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
443 #ifndef NDEBUG
444 	/* Check the object presence in the list. */
445 	struct mlx5_ibv_shared *lctx;
446 
447 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
448 		if (lctx == sh)
449 			break;
450 	assert(lctx);
451 	if (lctx != sh) {
452 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
453 		goto exit;
454 	}
455 #endif
456 	assert(sh);
457 	assert(sh->refcnt);
458 	/* Secondary process should not free the shared context. */
459 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
460 	if (--sh->refcnt)
461 		goto exit;
462 	/* Release created Memory Regions. */
463 	mlx5_mr_release(sh);
464 	LIST_REMOVE(sh, next);
465 	/*
466 	 *  Ensure there is no async event handler installed.
467 	 *  Only primary process handles async device events.
468 	 **/
469 	mlx5_flow_counters_mng_close(sh);
470 	assert(!sh->intr_cnt);
471 	if (sh->intr_cnt)
472 		mlx5_intr_callback_unregister
473 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
474 	pthread_mutex_destroy(&sh->intr_mutex);
475 	if (sh->pd)
476 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
477 	if (sh->ctx)
478 		claim_zero(mlx5_glue->close_device(sh->ctx));
479 	rte_free(sh);
480 exit:
481 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
482 }
483 
484 /**
485  * Initialize DR related data within private structure.
486  * Routine checks the reference counter and does actual
487  * resources creation/initialization only if counter is zero.
488  *
489  * @param[in] priv
490  *   Pointer to the private device data structure.
491  *
492  * @return
493  *   Zero on success, positive error code otherwise.
494  */
495 static int
496 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
497 {
498 #ifdef HAVE_MLX5DV_DR
499 	struct mlx5_ibv_shared *sh = priv->sh;
500 	int err = 0;
501 	void *domain;
502 
503 	assert(sh);
504 	if (sh->dv_refcnt) {
505 		/* Shared DV/DR structures is already initialized. */
506 		sh->dv_refcnt++;
507 		priv->dr_shared = 1;
508 		return 0;
509 	}
510 	/* Reference counter is zero, we should initialize structures. */
511 	domain = mlx5_glue->dr_create_domain(sh->ctx,
512 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
513 	if (!domain) {
514 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
515 		err = errno;
516 		goto error;
517 	}
518 	sh->rx_domain = domain;
519 	domain = mlx5_glue->dr_create_domain(sh->ctx,
520 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
521 	if (!domain) {
522 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
523 		err = errno;
524 		goto error;
525 	}
526 	pthread_mutex_init(&sh->dv_mutex, NULL);
527 	sh->tx_domain = domain;
528 #ifdef HAVE_MLX5DV_DR_ESWITCH
529 	if (priv->config.dv_esw_en) {
530 		domain  = mlx5_glue->dr_create_domain
531 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
532 		if (!domain) {
533 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
534 			err = errno;
535 			goto error;
536 		}
537 		sh->fdb_domain = domain;
538 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
539 	}
540 #endif
541 	sh->dv_refcnt++;
542 	priv->dr_shared = 1;
543 	return 0;
544 
545 error:
546        /* Rollback the created objects. */
547 	if (sh->rx_domain) {
548 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
549 		sh->rx_domain = NULL;
550 	}
551 	if (sh->tx_domain) {
552 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
553 		sh->tx_domain = NULL;
554 	}
555 	if (sh->fdb_domain) {
556 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
557 		sh->fdb_domain = NULL;
558 	}
559 	if (sh->esw_drop_action) {
560 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
561 		sh->esw_drop_action = NULL;
562 	}
563 	return err;
564 #else
565 	(void)priv;
566 	return 0;
567 #endif
568 }
569 
570 /**
571  * Destroy DR related data within private structure.
572  *
573  * @param[in] priv
574  *   Pointer to the private device data structure.
575  */
576 static void
577 mlx5_free_shared_dr(struct mlx5_priv *priv)
578 {
579 #ifdef HAVE_MLX5DV_DR
580 	struct mlx5_ibv_shared *sh;
581 
582 	if (!priv->dr_shared)
583 		return;
584 	priv->dr_shared = 0;
585 	sh = priv->sh;
586 	assert(sh);
587 	assert(sh->dv_refcnt);
588 	if (sh->dv_refcnt && --sh->dv_refcnt)
589 		return;
590 	if (sh->rx_domain) {
591 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
592 		sh->rx_domain = NULL;
593 	}
594 	if (sh->tx_domain) {
595 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
596 		sh->tx_domain = NULL;
597 	}
598 #ifdef HAVE_MLX5DV_DR_ESWITCH
599 	if (sh->fdb_domain) {
600 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
601 		sh->fdb_domain = NULL;
602 	}
603 	if (sh->esw_drop_action) {
604 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
605 		sh->esw_drop_action = NULL;
606 	}
607 #endif
608 	pthread_mutex_destroy(&sh->dv_mutex);
609 #else
610 	(void)priv;
611 #endif
612 }
613 
614 /**
615  * Initialize shared data between primary and secondary process.
616  *
617  * A memzone is reserved by primary process and secondary processes attach to
618  * the memzone.
619  *
620  * @return
621  *   0 on success, a negative errno value otherwise and rte_errno is set.
622  */
623 static int
624 mlx5_init_shared_data(void)
625 {
626 	const struct rte_memzone *mz;
627 	int ret = 0;
628 
629 	rte_spinlock_lock(&mlx5_shared_data_lock);
630 	if (mlx5_shared_data == NULL) {
631 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
632 			/* Allocate shared memory. */
633 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
634 						 sizeof(*mlx5_shared_data),
635 						 SOCKET_ID_ANY, 0);
636 			if (mz == NULL) {
637 				DRV_LOG(ERR,
638 					"Cannot allocate mlx5 shared data\n");
639 				ret = -rte_errno;
640 				goto error;
641 			}
642 			mlx5_shared_data = mz->addr;
643 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
644 			rte_spinlock_init(&mlx5_shared_data->lock);
645 		} else {
646 			/* Lookup allocated shared memory. */
647 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
648 			if (mz == NULL) {
649 				DRV_LOG(ERR,
650 					"Cannot attach mlx5 shared data\n");
651 				ret = -rte_errno;
652 				goto error;
653 			}
654 			mlx5_shared_data = mz->addr;
655 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
656 		}
657 	}
658 error:
659 	rte_spinlock_unlock(&mlx5_shared_data_lock);
660 	return ret;
661 }
662 
663 /**
664  * Retrieve integer value from environment variable.
665  *
666  * @param[in] name
667  *   Environment variable name.
668  *
669  * @return
670  *   Integer value, 0 if the variable is not set.
671  */
672 int
673 mlx5_getenv_int(const char *name)
674 {
675 	const char *val = getenv(name);
676 
677 	if (val == NULL)
678 		return 0;
679 	return atoi(val);
680 }
681 
682 /**
683  * Verbs callback to allocate a memory. This function should allocate the space
684  * according to the size provided residing inside a huge page.
685  * Please note that all allocation must respect the alignment from libmlx5
686  * (i.e. currently sysconf(_SC_PAGESIZE)).
687  *
688  * @param[in] size
689  *   The size in bytes of the memory to allocate.
690  * @param[in] data
691  *   A pointer to the callback data.
692  *
693  * @return
694  *   Allocated buffer, NULL otherwise and rte_errno is set.
695  */
696 static void *
697 mlx5_alloc_verbs_buf(size_t size, void *data)
698 {
699 	struct mlx5_priv *priv = data;
700 	void *ret;
701 	size_t alignment = sysconf(_SC_PAGESIZE);
702 	unsigned int socket = SOCKET_ID_ANY;
703 
704 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
705 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
706 
707 		socket = ctrl->socket;
708 	} else if (priv->verbs_alloc_ctx.type ==
709 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
710 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
711 
712 		socket = ctrl->socket;
713 	}
714 	assert(data != NULL);
715 	ret = rte_malloc_socket(__func__, size, alignment, socket);
716 	if (!ret && size)
717 		rte_errno = ENOMEM;
718 	return ret;
719 }
720 
721 /**
722  * Verbs callback to free a memory.
723  *
724  * @param[in] ptr
725  *   A pointer to the memory to free.
726  * @param[in] data
727  *   A pointer to the callback data.
728  */
729 static void
730 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
731 {
732 	assert(data != NULL);
733 	rte_free(ptr);
734 }
735 
736 /**
737  * Initialize process private data structure.
738  *
739  * @param dev
740  *   Pointer to Ethernet device structure.
741  *
742  * @return
743  *   0 on success, a negative errno value otherwise and rte_errno is set.
744  */
745 int
746 mlx5_proc_priv_init(struct rte_eth_dev *dev)
747 {
748 	struct mlx5_priv *priv = dev->data->dev_private;
749 	struct mlx5_proc_priv *ppriv;
750 	size_t ppriv_size;
751 
752 	/*
753 	 * UAR register table follows the process private structure. BlueFlame
754 	 * registers for Tx queues are stored in the table.
755 	 */
756 	ppriv_size =
757 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
758 	ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
759 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
760 	if (!ppriv) {
761 		rte_errno = ENOMEM;
762 		return -rte_errno;
763 	}
764 	ppriv->uar_table_sz = ppriv_size;
765 	dev->process_private = ppriv;
766 	return 0;
767 }
768 
769 /**
770  * Un-initialize process private data structure.
771  *
772  * @param dev
773  *   Pointer to Ethernet device structure.
774  */
775 static void
776 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
777 {
778 	if (!dev->process_private)
779 		return;
780 	rte_free(dev->process_private);
781 	dev->process_private = NULL;
782 }
783 
784 /**
785  * DPDK callback to close the device.
786  *
787  * Destroy all queues and objects, free memory.
788  *
789  * @param dev
790  *   Pointer to Ethernet device structure.
791  */
792 static void
793 mlx5_dev_close(struct rte_eth_dev *dev)
794 {
795 	struct mlx5_priv *priv = dev->data->dev_private;
796 	unsigned int i;
797 	int ret;
798 
799 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
800 		dev->data->port_id,
801 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
802 	/* In case mlx5_dev_stop() has not been called. */
803 	mlx5_dev_interrupt_handler_uninstall(dev);
804 	mlx5_traffic_disable(dev);
805 	mlx5_flow_flush(dev, NULL);
806 	/* Prevent crashes when queues are still in use. */
807 	dev->rx_pkt_burst = removed_rx_burst;
808 	dev->tx_pkt_burst = removed_tx_burst;
809 	rte_wmb();
810 	/* Disable datapath on secondary process. */
811 	mlx5_mp_req_stop_rxtx(dev);
812 	if (priv->rxqs != NULL) {
813 		/* XXX race condition if mlx5_rx_burst() is still running. */
814 		usleep(1000);
815 		for (i = 0; (i != priv->rxqs_n); ++i)
816 			mlx5_rxq_release(dev, i);
817 		priv->rxqs_n = 0;
818 		priv->rxqs = NULL;
819 	}
820 	if (priv->txqs != NULL) {
821 		/* XXX race condition if mlx5_tx_burst() is still running. */
822 		usleep(1000);
823 		for (i = 0; (i != priv->txqs_n); ++i)
824 			mlx5_txq_release(dev, i);
825 		priv->txqs_n = 0;
826 		priv->txqs = NULL;
827 	}
828 	mlx5_proc_priv_uninit(dev);
829 	mlx5_mprq_free_mp(dev);
830 	/* Remove from memory callback device list. */
831 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
832 	assert(priv->sh);
833 	LIST_REMOVE(priv->sh, mem_event_cb);
834 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
835 	mlx5_free_shared_dr(priv);
836 	if (priv->rss_conf.rss_key != NULL)
837 		rte_free(priv->rss_conf.rss_key);
838 	if (priv->reta_idx != NULL)
839 		rte_free(priv->reta_idx);
840 	if (priv->config.vf)
841 		mlx5_nl_mac_addr_flush(dev);
842 	if (priv->nl_socket_route >= 0)
843 		close(priv->nl_socket_route);
844 	if (priv->nl_socket_rdma >= 0)
845 		close(priv->nl_socket_rdma);
846 	if (priv->sh) {
847 		/*
848 		 * Free the shared context in last turn, because the cleanup
849 		 * routines above may use some shared fields, like
850 		 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
851 		 * ifindex if Netlink fails.
852 		 */
853 		mlx5_free_shared_ibctx(priv->sh);
854 		priv->sh = NULL;
855 	}
856 	ret = mlx5_hrxq_verify(dev);
857 	if (ret)
858 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
859 			dev->data->port_id);
860 	ret = mlx5_ind_table_obj_verify(dev);
861 	if (ret)
862 		DRV_LOG(WARNING, "port %u some indirection table still remain",
863 			dev->data->port_id);
864 	ret = mlx5_rxq_obj_verify(dev);
865 	if (ret)
866 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
867 			dev->data->port_id);
868 	ret = mlx5_rxq_verify(dev);
869 	if (ret)
870 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
871 			dev->data->port_id);
872 	ret = mlx5_txq_ibv_verify(dev);
873 	if (ret)
874 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
875 			dev->data->port_id);
876 	ret = mlx5_txq_verify(dev);
877 	if (ret)
878 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
879 			dev->data->port_id);
880 	ret = mlx5_flow_verify(dev);
881 	if (ret)
882 		DRV_LOG(WARNING, "port %u some flows still remain",
883 			dev->data->port_id);
884 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
885 		unsigned int c = 0;
886 		uint16_t port_id;
887 
888 		RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
889 			struct mlx5_priv *opriv =
890 				rte_eth_devices[port_id].data->dev_private;
891 
892 			if (!opriv ||
893 			    opriv->domain_id != priv->domain_id ||
894 			    &rte_eth_devices[port_id] == dev)
895 				continue;
896 			++c;
897 		}
898 		if (!c)
899 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
900 	}
901 	memset(priv, 0, sizeof(*priv));
902 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
903 	/*
904 	 * Reset mac_addrs to NULL such that it is not freed as part of
905 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
906 	 * it is freed when dev_private is freed.
907 	 */
908 	dev->data->mac_addrs = NULL;
909 }
910 
911 const struct eth_dev_ops mlx5_dev_ops = {
912 	.dev_configure = mlx5_dev_configure,
913 	.dev_start = mlx5_dev_start,
914 	.dev_stop = mlx5_dev_stop,
915 	.dev_set_link_down = mlx5_set_link_down,
916 	.dev_set_link_up = mlx5_set_link_up,
917 	.dev_close = mlx5_dev_close,
918 	.promiscuous_enable = mlx5_promiscuous_enable,
919 	.promiscuous_disable = mlx5_promiscuous_disable,
920 	.allmulticast_enable = mlx5_allmulticast_enable,
921 	.allmulticast_disable = mlx5_allmulticast_disable,
922 	.link_update = mlx5_link_update,
923 	.stats_get = mlx5_stats_get,
924 	.stats_reset = mlx5_stats_reset,
925 	.xstats_get = mlx5_xstats_get,
926 	.xstats_reset = mlx5_xstats_reset,
927 	.xstats_get_names = mlx5_xstats_get_names,
928 	.fw_version_get = mlx5_fw_version_get,
929 	.dev_infos_get = mlx5_dev_infos_get,
930 	.read_clock = mlx5_read_clock,
931 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
932 	.vlan_filter_set = mlx5_vlan_filter_set,
933 	.rx_queue_setup = mlx5_rx_queue_setup,
934 	.tx_queue_setup = mlx5_tx_queue_setup,
935 	.rx_queue_release = mlx5_rx_queue_release,
936 	.tx_queue_release = mlx5_tx_queue_release,
937 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
938 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
939 	.mac_addr_remove = mlx5_mac_addr_remove,
940 	.mac_addr_add = mlx5_mac_addr_add,
941 	.mac_addr_set = mlx5_mac_addr_set,
942 	.set_mc_addr_list = mlx5_set_mc_addr_list,
943 	.mtu_set = mlx5_dev_set_mtu,
944 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
945 	.vlan_offload_set = mlx5_vlan_offload_set,
946 	.reta_update = mlx5_dev_rss_reta_update,
947 	.reta_query = mlx5_dev_rss_reta_query,
948 	.rss_hash_update = mlx5_rss_hash_update,
949 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
950 	.filter_ctrl = mlx5_dev_filter_ctrl,
951 	.rx_descriptor_status = mlx5_rx_descriptor_status,
952 	.tx_descriptor_status = mlx5_tx_descriptor_status,
953 	.rx_queue_count = mlx5_rx_queue_count,
954 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
955 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
956 	.is_removed = mlx5_is_removed,
957 };
958 
959 /* Available operations from secondary process. */
960 static const struct eth_dev_ops mlx5_dev_sec_ops = {
961 	.stats_get = mlx5_stats_get,
962 	.stats_reset = mlx5_stats_reset,
963 	.xstats_get = mlx5_xstats_get,
964 	.xstats_reset = mlx5_xstats_reset,
965 	.xstats_get_names = mlx5_xstats_get_names,
966 	.fw_version_get = mlx5_fw_version_get,
967 	.dev_infos_get = mlx5_dev_infos_get,
968 	.rx_descriptor_status = mlx5_rx_descriptor_status,
969 	.tx_descriptor_status = mlx5_tx_descriptor_status,
970 };
971 
972 /* Available operations in flow isolated mode. */
973 const struct eth_dev_ops mlx5_dev_ops_isolate = {
974 	.dev_configure = mlx5_dev_configure,
975 	.dev_start = mlx5_dev_start,
976 	.dev_stop = mlx5_dev_stop,
977 	.dev_set_link_down = mlx5_set_link_down,
978 	.dev_set_link_up = mlx5_set_link_up,
979 	.dev_close = mlx5_dev_close,
980 	.promiscuous_enable = mlx5_promiscuous_enable,
981 	.promiscuous_disable = mlx5_promiscuous_disable,
982 	.allmulticast_enable = mlx5_allmulticast_enable,
983 	.allmulticast_disable = mlx5_allmulticast_disable,
984 	.link_update = mlx5_link_update,
985 	.stats_get = mlx5_stats_get,
986 	.stats_reset = mlx5_stats_reset,
987 	.xstats_get = mlx5_xstats_get,
988 	.xstats_reset = mlx5_xstats_reset,
989 	.xstats_get_names = mlx5_xstats_get_names,
990 	.fw_version_get = mlx5_fw_version_get,
991 	.dev_infos_get = mlx5_dev_infos_get,
992 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
993 	.vlan_filter_set = mlx5_vlan_filter_set,
994 	.rx_queue_setup = mlx5_rx_queue_setup,
995 	.tx_queue_setup = mlx5_tx_queue_setup,
996 	.rx_queue_release = mlx5_rx_queue_release,
997 	.tx_queue_release = mlx5_tx_queue_release,
998 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
999 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1000 	.mac_addr_remove = mlx5_mac_addr_remove,
1001 	.mac_addr_add = mlx5_mac_addr_add,
1002 	.mac_addr_set = mlx5_mac_addr_set,
1003 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1004 	.mtu_set = mlx5_dev_set_mtu,
1005 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1006 	.vlan_offload_set = mlx5_vlan_offload_set,
1007 	.filter_ctrl = mlx5_dev_filter_ctrl,
1008 	.rx_descriptor_status = mlx5_rx_descriptor_status,
1009 	.tx_descriptor_status = mlx5_tx_descriptor_status,
1010 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1011 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1012 	.is_removed = mlx5_is_removed,
1013 };
1014 
1015 /**
1016  * Verify and store value for device argument.
1017  *
1018  * @param[in] key
1019  *   Key argument to verify.
1020  * @param[in] val
1021  *   Value associated with key.
1022  * @param opaque
1023  *   User data.
1024  *
1025  * @return
1026  *   0 on success, a negative errno value otherwise and rte_errno is set.
1027  */
1028 static int
1029 mlx5_args_check(const char *key, const char *val, void *opaque)
1030 {
1031 	struct mlx5_dev_config *config = opaque;
1032 	unsigned long tmp;
1033 
1034 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1035 	if (!strcmp(MLX5_REPRESENTOR, key))
1036 		return 0;
1037 	errno = 0;
1038 	tmp = strtoul(val, NULL, 0);
1039 	if (errno) {
1040 		rte_errno = errno;
1041 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1042 		return -rte_errno;
1043 	}
1044 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1045 		config->cqe_comp = !!tmp;
1046 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1047 		config->cqe_pad = !!tmp;
1048 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1049 		config->hw_padding = !!tmp;
1050 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1051 		config->mprq.enabled = !!tmp;
1052 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1053 		config->mprq.stride_num_n = tmp;
1054 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1055 		config->mprq.max_memcpy_len = tmp;
1056 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1057 		config->mprq.min_rxqs_num = tmp;
1058 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1059 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1060 				 " converted to txq_inline_max", key);
1061 		config->txq_inline_max = tmp;
1062 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1063 		config->txq_inline_max = tmp;
1064 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1065 		config->txq_inline_min = tmp;
1066 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1067 		config->txq_inline_mpw = tmp;
1068 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1069 		config->txqs_inline = tmp;
1070 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1071 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1072 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1073 		config->mps = !!tmp;
1074 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1075 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1076 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1077 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1078 				 " converted to txq_inline_mpw", key);
1079 		config->txq_inline_mpw = tmp;
1080 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1081 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1082 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1083 		config->rx_vec_en = !!tmp;
1084 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1085 		config->l3_vxlan_en = !!tmp;
1086 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1087 		config->vf_nl_en = !!tmp;
1088 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1089 		config->dv_esw_en = !!tmp;
1090 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1091 		config->dv_flow_en = !!tmp;
1092 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1093 		config->mr_ext_memseg_en = !!tmp;
1094 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1095 		config->max_dump_files_num = tmp;
1096 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1097 		config->lro.timeout = tmp;
1098 	} else {
1099 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1100 		rte_errno = EINVAL;
1101 		return -rte_errno;
1102 	}
1103 	return 0;
1104 }
1105 
1106 /**
1107  * Parse device parameters.
1108  *
1109  * @param config
1110  *   Pointer to device configuration structure.
1111  * @param devargs
1112  *   Device arguments structure.
1113  *
1114  * @return
1115  *   0 on success, a negative errno value otherwise and rte_errno is set.
1116  */
1117 static int
1118 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1119 {
1120 	const char **params = (const char *[]){
1121 		MLX5_RXQ_CQE_COMP_EN,
1122 		MLX5_RXQ_CQE_PAD_EN,
1123 		MLX5_RXQ_PKT_PAD_EN,
1124 		MLX5_RX_MPRQ_EN,
1125 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1126 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1127 		MLX5_RXQS_MIN_MPRQ,
1128 		MLX5_TXQ_INLINE,
1129 		MLX5_TXQ_INLINE_MIN,
1130 		MLX5_TXQ_INLINE_MAX,
1131 		MLX5_TXQ_INLINE_MPW,
1132 		MLX5_TXQS_MIN_INLINE,
1133 		MLX5_TXQS_MAX_VEC,
1134 		MLX5_TXQ_MPW_EN,
1135 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1136 		MLX5_TXQ_MAX_INLINE_LEN,
1137 		MLX5_TX_VEC_EN,
1138 		MLX5_RX_VEC_EN,
1139 		MLX5_L3_VXLAN_EN,
1140 		MLX5_VF_NL_EN,
1141 		MLX5_DV_ESW_EN,
1142 		MLX5_DV_FLOW_EN,
1143 		MLX5_MR_EXT_MEMSEG_EN,
1144 		MLX5_REPRESENTOR,
1145 		MLX5_MAX_DUMP_FILES_NUM,
1146 		MLX5_LRO_TIMEOUT_USEC,
1147 		NULL,
1148 	};
1149 	struct rte_kvargs *kvlist;
1150 	int ret = 0;
1151 	int i;
1152 
1153 	if (devargs == NULL)
1154 		return 0;
1155 	/* Following UGLY cast is done to pass checkpatch. */
1156 	kvlist = rte_kvargs_parse(devargs->args, params);
1157 	if (kvlist == NULL) {
1158 		rte_errno = EINVAL;
1159 		return -rte_errno;
1160 	}
1161 	/* Process parameters. */
1162 	for (i = 0; (params[i] != NULL); ++i) {
1163 		if (rte_kvargs_count(kvlist, params[i])) {
1164 			ret = rte_kvargs_process(kvlist, params[i],
1165 						 mlx5_args_check, config);
1166 			if (ret) {
1167 				rte_errno = EINVAL;
1168 				rte_kvargs_free(kvlist);
1169 				return -rte_errno;
1170 			}
1171 		}
1172 	}
1173 	rte_kvargs_free(kvlist);
1174 	return 0;
1175 }
1176 
1177 static struct rte_pci_driver mlx5_driver;
1178 
1179 /**
1180  * PMD global initialization.
1181  *
1182  * Independent from individual device, this function initializes global
1183  * per-PMD data structures distinguishing primary and secondary processes.
1184  * Hence, each initialization is called once per a process.
1185  *
1186  * @return
1187  *   0 on success, a negative errno value otherwise and rte_errno is set.
1188  */
1189 static int
1190 mlx5_init_once(void)
1191 {
1192 	struct mlx5_shared_data *sd;
1193 	struct mlx5_local_data *ld = &mlx5_local_data;
1194 	int ret = 0;
1195 
1196 	if (mlx5_init_shared_data())
1197 		return -rte_errno;
1198 	sd = mlx5_shared_data;
1199 	assert(sd);
1200 	rte_spinlock_lock(&sd->lock);
1201 	switch (rte_eal_process_type()) {
1202 	case RTE_PROC_PRIMARY:
1203 		if (sd->init_done)
1204 			break;
1205 		LIST_INIT(&sd->mem_event_cb_list);
1206 		rte_rwlock_init(&sd->mem_event_rwlock);
1207 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1208 						mlx5_mr_mem_event_cb, NULL);
1209 		ret = mlx5_mp_init_primary();
1210 		if (ret)
1211 			goto out;
1212 		sd->init_done = true;
1213 		break;
1214 	case RTE_PROC_SECONDARY:
1215 		if (ld->init_done)
1216 			break;
1217 		ret = mlx5_mp_init_secondary();
1218 		if (ret)
1219 			goto out;
1220 		++sd->secondary_cnt;
1221 		ld->init_done = true;
1222 		break;
1223 	default:
1224 		break;
1225 	}
1226 out:
1227 	rte_spinlock_unlock(&sd->lock);
1228 	return ret;
1229 }
1230 
1231 /**
1232  * Configures the minimal amount of data to inline into WQE
1233  * while sending packets.
1234  *
1235  * - the txq_inline_min has the maximal priority, if this
1236  *   key is specified in devargs
1237  * - if DevX is enabled the inline mode is queried from the
1238  *   device (HCA attributes and NIC vport context if needed).
1239  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1240  *   and none (0 bytes) for other NICs
1241  *
1242  * @param spawn
1243  *   Verbs device parameters (name, port, switch_info) to spawn.
1244  * @param config
1245  *   Device configuration parameters.
1246  */
1247 static void
1248 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1249 		    struct mlx5_dev_config *config)
1250 {
1251 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1252 		/* Application defines size of inlined data explicitly. */
1253 		switch (spawn->pci_dev->id.device_id) {
1254 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1255 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1256 			if (config->txq_inline_min <
1257 				       (int)MLX5_INLINE_HSIZE_L2) {
1258 				DRV_LOG(DEBUG,
1259 					"txq_inline_mix aligned to minimal"
1260 					" ConnectX-4 required value %d",
1261 					(int)MLX5_INLINE_HSIZE_L2);
1262 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1263 			}
1264 			break;
1265 		}
1266 		goto exit;
1267 	}
1268 	if (config->hca_attr.eth_net_offloads) {
1269 		/* We have DevX enabled, inline mode queried successfully. */
1270 		switch (config->hca_attr.wqe_inline_mode) {
1271 		case MLX5_CAP_INLINE_MODE_L2:
1272 			/* outer L2 header must be inlined. */
1273 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1274 			goto exit;
1275 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1276 			/* No inline data are required by NIC. */
1277 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1278 			config->hw_vlan_insert =
1279 				config->hca_attr.wqe_vlan_insert;
1280 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1281 			goto exit;
1282 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1283 			/* inline mode is defined by NIC vport context. */
1284 			if (!config->hca_attr.eth_virt)
1285 				break;
1286 			switch (config->hca_attr.vport_inline_mode) {
1287 			case MLX5_INLINE_MODE_NONE:
1288 				config->txq_inline_min =
1289 					MLX5_INLINE_HSIZE_NONE;
1290 				goto exit;
1291 			case MLX5_INLINE_MODE_L2:
1292 				config->txq_inline_min =
1293 					MLX5_INLINE_HSIZE_L2;
1294 				goto exit;
1295 			case MLX5_INLINE_MODE_IP:
1296 				config->txq_inline_min =
1297 					MLX5_INLINE_HSIZE_L3;
1298 				goto exit;
1299 			case MLX5_INLINE_MODE_TCP_UDP:
1300 				config->txq_inline_min =
1301 					MLX5_INLINE_HSIZE_L4;
1302 				goto exit;
1303 			case MLX5_INLINE_MODE_INNER_L2:
1304 				config->txq_inline_min =
1305 					MLX5_INLINE_HSIZE_INNER_L2;
1306 				goto exit;
1307 			case MLX5_INLINE_MODE_INNER_IP:
1308 				config->txq_inline_min =
1309 					MLX5_INLINE_HSIZE_INNER_L3;
1310 				goto exit;
1311 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1312 				config->txq_inline_min =
1313 					MLX5_INLINE_HSIZE_INNER_L4;
1314 				goto exit;
1315 			}
1316 		}
1317 	}
1318 	/*
1319 	 * We get here if we are unable to deduce
1320 	 * inline data size with DevX. Try PCI ID
1321 	 * to determine old NICs.
1322 	 */
1323 	switch (spawn->pci_dev->id.device_id) {
1324 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1325 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1326 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1327 		config->hw_vlan_insert = 0;
1328 		break;
1329 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1330 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1331 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1332 		config->hw_vlan_insert = 0;
1333 		break;
1334 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1335 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1336 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1337 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1338 		/*
1339 		 * These NICs support VLAN insertion from WQE and
1340 		 * report the wqe_vlan_insert flag. But there is the bug
1341 		 * and PFC control may be broken, so disable feature.
1342 		 */
1343 		config->hw_vlan_insert = 0;
1344 		break;
1345 	default:
1346 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1347 		break;
1348 	}
1349 exit:
1350 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1351 }
1352 
1353 /**
1354  * Allocate page of door-bells and register it using DevX API.
1355  *
1356  * @param [in] dev
1357  *   Pointer to Ethernet device.
1358  *
1359  * @return
1360  *   Pointer to new page on success, NULL otherwise.
1361  */
1362 static struct mlx5_devx_dbr_page *
1363 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1364 {
1365 	struct mlx5_priv *priv = dev->data->dev_private;
1366 	struct mlx5_devx_dbr_page *page;
1367 
1368 	/* Allocate space for door-bell page and management data. */
1369 	page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1370 				 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1371 	if (!page) {
1372 		DRV_LOG(ERR, "port %u cannot allocate dbr page",
1373 			dev->data->port_id);
1374 		return NULL;
1375 	}
1376 	/* Register allocated memory. */
1377 	page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1378 					      MLX5_DBR_PAGE_SIZE, 0);
1379 	if (!page->umem) {
1380 		DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1381 			dev->data->port_id);
1382 		rte_free(page);
1383 		return NULL;
1384 	}
1385 	return page;
1386 }
1387 
1388 /**
1389  * Find the next available door-bell, allocate new page if needed.
1390  *
1391  * @param [in] dev
1392  *   Pointer to Ethernet device.
1393  * @param [out] dbr_page
1394  *   Door-bell page containing the page data.
1395  *
1396  * @return
1397  *   Door-bell address offset on success, a negative error value otherwise.
1398  */
1399 int64_t
1400 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1401 {
1402 	struct mlx5_priv *priv = dev->data->dev_private;
1403 	struct mlx5_devx_dbr_page *page = NULL;
1404 	uint32_t i, j;
1405 
1406 	LIST_FOREACH(page, &priv->dbrpgs, next)
1407 		if (page->dbr_count < MLX5_DBR_PER_PAGE)
1408 			break;
1409 	if (!page) { /* No page with free door-bell exists. */
1410 		page = mlx5_alloc_dbr_page(dev);
1411 		if (!page) /* Failed to allocate new page. */
1412 			return (-1);
1413 		LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1414 	}
1415 	/* Loop to find bitmap part with clear bit. */
1416 	for (i = 0;
1417 	     i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1418 	     i++)
1419 		; /* Empty. */
1420 	/* Find the first clear bit. */
1421 	j = rte_bsf64(~page->dbr_bitmap[i]);
1422 	assert(i < (MLX5_DBR_PER_PAGE / 64));
1423 	page->dbr_bitmap[i] |= (1 << j);
1424 	page->dbr_count++;
1425 	*dbr_page = page;
1426 	return (((i * 64) + j) * sizeof(uint64_t));
1427 }
1428 
1429 /**
1430  * Release a door-bell record.
1431  *
1432  * @param [in] dev
1433  *   Pointer to Ethernet device.
1434  * @param [in] umem_id
1435  *   UMEM ID of page containing the door-bell record to release.
1436  * @param [in] offset
1437  *   Offset of door-bell record in page.
1438  *
1439  * @return
1440  *   0 on success, a negative error value otherwise.
1441  */
1442 int32_t
1443 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1444 {
1445 	struct mlx5_priv *priv = dev->data->dev_private;
1446 	struct mlx5_devx_dbr_page *page = NULL;
1447 	int ret = 0;
1448 
1449 	LIST_FOREACH(page, &priv->dbrpgs, next)
1450 		/* Find the page this address belongs to. */
1451 		if (page->umem->umem_id == umem_id)
1452 			break;
1453 	if (!page)
1454 		return -EINVAL;
1455 	page->dbr_count--;
1456 	if (!page->dbr_count) {
1457 		/* Page not used, free it and remove from list. */
1458 		LIST_REMOVE(page, next);
1459 		if (page->umem)
1460 			ret = -mlx5_glue->devx_umem_dereg(page->umem);
1461 		rte_free(page);
1462 	} else {
1463 		/* Mark in bitmap that this door-bell is not in use. */
1464 		offset /= MLX5_DBR_SIZE;
1465 		int i = offset / 64;
1466 		int j = offset % 64;
1467 
1468 		page->dbr_bitmap[i] &= ~(1 << j);
1469 	}
1470 	return ret;
1471 }
1472 
1473 /**
1474  * Spawn an Ethernet device from Verbs information.
1475  *
1476  * @param dpdk_dev
1477  *   Backing DPDK device.
1478  * @param spawn
1479  *   Verbs device parameters (name, port, switch_info) to spawn.
1480  * @param config
1481  *   Device configuration parameters.
1482  *
1483  * @return
1484  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1485  *   is set. The following errors are defined:
1486  *
1487  *   EBUSY: device is not supposed to be spawned.
1488  *   EEXIST: device is already spawned
1489  */
1490 static struct rte_eth_dev *
1491 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1492 	       struct mlx5_dev_spawn_data *spawn,
1493 	       struct mlx5_dev_config config)
1494 {
1495 	const struct mlx5_switch_info *switch_info = &spawn->info;
1496 	struct mlx5_ibv_shared *sh = NULL;
1497 	struct ibv_port_attr port_attr;
1498 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1499 	struct rte_eth_dev *eth_dev = NULL;
1500 	struct mlx5_priv *priv = NULL;
1501 	int err = 0;
1502 	unsigned int hw_padding = 0;
1503 	unsigned int mps;
1504 	unsigned int cqe_comp;
1505 	unsigned int cqe_pad = 0;
1506 	unsigned int tunnel_en = 0;
1507 	unsigned int mpls_en = 0;
1508 	unsigned int swp = 0;
1509 	unsigned int mprq = 0;
1510 	unsigned int mprq_min_stride_size_n = 0;
1511 	unsigned int mprq_max_stride_size_n = 0;
1512 	unsigned int mprq_min_stride_num_n = 0;
1513 	unsigned int mprq_max_stride_num_n = 0;
1514 	struct rte_ether_addr mac;
1515 	char name[RTE_ETH_NAME_MAX_LEN];
1516 	int own_domain_id = 0;
1517 	uint16_t port_id;
1518 	unsigned int i;
1519 
1520 	/* Determine if this port representor is supposed to be spawned. */
1521 	if (switch_info->representor && dpdk_dev->devargs) {
1522 		struct rte_eth_devargs eth_da;
1523 
1524 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1525 		if (err) {
1526 			rte_errno = -err;
1527 			DRV_LOG(ERR, "failed to process device arguments: %s",
1528 				strerror(rte_errno));
1529 			return NULL;
1530 		}
1531 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
1532 			if (eth_da.representor_ports[i] ==
1533 			    (uint16_t)switch_info->port_name)
1534 				break;
1535 		if (i == eth_da.nb_representor_ports) {
1536 			rte_errno = EBUSY;
1537 			return NULL;
1538 		}
1539 	}
1540 	/* Build device name. */
1541 	if (!switch_info->representor)
1542 		strlcpy(name, dpdk_dev->name, sizeof(name));
1543 	else
1544 		snprintf(name, sizeof(name), "%s_representor_%u",
1545 			 dpdk_dev->name, switch_info->port_name);
1546 	/* check if the device is already spawned */
1547 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1548 		rte_errno = EEXIST;
1549 		return NULL;
1550 	}
1551 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1552 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1553 		eth_dev = rte_eth_dev_attach_secondary(name);
1554 		if (eth_dev == NULL) {
1555 			DRV_LOG(ERR, "can not attach rte ethdev");
1556 			rte_errno = ENOMEM;
1557 			return NULL;
1558 		}
1559 		eth_dev->device = dpdk_dev;
1560 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1561 		err = mlx5_proc_priv_init(eth_dev);
1562 		if (err)
1563 			return NULL;
1564 		/* Receive command fd from primary process */
1565 		err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1566 		if (err < 0)
1567 			return NULL;
1568 		/* Remap UAR for Tx queues. */
1569 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1570 		if (err)
1571 			return NULL;
1572 		/*
1573 		 * Ethdev pointer is still required as input since
1574 		 * the primary device is not accessible from the
1575 		 * secondary process.
1576 		 */
1577 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1578 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1579 		return eth_dev;
1580 	}
1581 	sh = mlx5_alloc_shared_ibctx(spawn);
1582 	if (!sh)
1583 		return NULL;
1584 	config.devx = sh->devx;
1585 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1586 	config.dest_tir = 1;
1587 #endif
1588 #ifdef HAVE_IBV_MLX5_MOD_SWP
1589 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1590 #endif
1591 	/*
1592 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1593 	 * as all ConnectX-5 devices.
1594 	 */
1595 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1596 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1597 #endif
1598 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1599 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1600 #endif
1601 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1602 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1603 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1604 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1605 			mps = MLX5_MPW_ENHANCED;
1606 		} else {
1607 			DRV_LOG(DEBUG, "MPW is supported");
1608 			mps = MLX5_MPW;
1609 		}
1610 	} else {
1611 		DRV_LOG(DEBUG, "MPW isn't supported");
1612 		mps = MLX5_MPW_DISABLED;
1613 	}
1614 #ifdef HAVE_IBV_MLX5_MOD_SWP
1615 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1616 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1617 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1618 #endif
1619 	config.swp = !!swp;
1620 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1621 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1622 		struct mlx5dv_striding_rq_caps mprq_caps =
1623 			dv_attr.striding_rq_caps;
1624 
1625 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1626 			mprq_caps.min_single_stride_log_num_of_bytes);
1627 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1628 			mprq_caps.max_single_stride_log_num_of_bytes);
1629 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1630 			mprq_caps.min_single_wqe_log_num_of_strides);
1631 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1632 			mprq_caps.max_single_wqe_log_num_of_strides);
1633 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1634 			mprq_caps.supported_qpts);
1635 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1636 		mprq = 1;
1637 		mprq_min_stride_size_n =
1638 			mprq_caps.min_single_stride_log_num_of_bytes;
1639 		mprq_max_stride_size_n =
1640 			mprq_caps.max_single_stride_log_num_of_bytes;
1641 		mprq_min_stride_num_n =
1642 			mprq_caps.min_single_wqe_log_num_of_strides;
1643 		mprq_max_stride_num_n =
1644 			mprq_caps.max_single_wqe_log_num_of_strides;
1645 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1646 						   mprq_min_stride_num_n);
1647 	}
1648 #endif
1649 	if (RTE_CACHE_LINE_SIZE == 128 &&
1650 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1651 		cqe_comp = 0;
1652 	else
1653 		cqe_comp = 1;
1654 	config.cqe_comp = cqe_comp;
1655 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1656 	/* Whether device supports 128B Rx CQE padding. */
1657 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1658 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1659 #endif
1660 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1661 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1662 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1663 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1664 			     (dv_attr.tunnel_offloads_caps &
1665 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1666 	}
1667 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1668 		tunnel_en ? "" : "not ");
1669 #else
1670 	DRV_LOG(WARNING,
1671 		"tunnel offloading disabled due to old OFED/rdma-core version");
1672 #endif
1673 	config.tunnel_en = tunnel_en;
1674 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1675 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1676 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1677 		   (dv_attr.tunnel_offloads_caps &
1678 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1679 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1680 		mpls_en ? "" : "not ");
1681 #else
1682 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1683 		" old OFED/rdma-core version or firmware configuration");
1684 #endif
1685 	config.mpls_en = mpls_en;
1686 	/* Check port status. */
1687 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1688 	if (err) {
1689 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1690 		goto error;
1691 	}
1692 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1693 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1694 		err = EINVAL;
1695 		goto error;
1696 	}
1697 	if (port_attr.state != IBV_PORT_ACTIVE)
1698 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1699 			mlx5_glue->port_state_str(port_attr.state),
1700 			port_attr.state);
1701 	/* Allocate private eth device data. */
1702 	priv = rte_zmalloc("ethdev private structure",
1703 			   sizeof(*priv),
1704 			   RTE_CACHE_LINE_SIZE);
1705 	if (priv == NULL) {
1706 		DRV_LOG(ERR, "priv allocation failure");
1707 		err = ENOMEM;
1708 		goto error;
1709 	}
1710 	priv->sh = sh;
1711 	priv->ibv_port = spawn->ibv_port;
1712 	priv->mtu = RTE_ETHER_MTU;
1713 #ifndef RTE_ARCH_64
1714 	/* Initialize UAR access locks for 32bit implementations. */
1715 	rte_spinlock_init(&priv->uar_lock_cq);
1716 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1717 		rte_spinlock_init(&priv->uar_lock[i]);
1718 #endif
1719 	/* Some internal functions rely on Netlink sockets, open them now. */
1720 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1721 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1722 	priv->nl_sn = 0;
1723 	priv->representor = !!switch_info->representor;
1724 	priv->master = !!switch_info->master;
1725 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1726 	/*
1727 	 * Currently we support single E-Switch per PF configurations
1728 	 * only and vport_id field contains the vport index for
1729 	 * associated VF, which is deduced from representor port name.
1730 	 * For example, let's have the IB device port 10, it has
1731 	 * attached network device eth0, which has port name attribute
1732 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1733 	 * as 3 (2+1). This assigning schema should be changed if the
1734 	 * multiple E-Switch instances per PF configurations or/and PCI
1735 	 * subfunctions are added.
1736 	 */
1737 	priv->vport_id = switch_info->representor ?
1738 			 switch_info->port_name + 1 : -1;
1739 	/* representor_id field keeps the unmodified port/VF index. */
1740 	priv->representor_id = switch_info->representor ?
1741 			       switch_info->port_name : -1;
1742 	/*
1743 	 * Look for sibling devices in order to reuse their switch domain
1744 	 * if any, otherwise allocate one.
1745 	 */
1746 	RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1747 		const struct mlx5_priv *opriv =
1748 			rte_eth_devices[port_id].data->dev_private;
1749 
1750 		if (!opriv ||
1751 			opriv->domain_id ==
1752 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1753 			continue;
1754 		priv->domain_id = opriv->domain_id;
1755 		break;
1756 	}
1757 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1758 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1759 		if (err) {
1760 			err = rte_errno;
1761 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1762 				strerror(rte_errno));
1763 			goto error;
1764 		}
1765 		own_domain_id = 1;
1766 	}
1767 	err = mlx5_args(&config, dpdk_dev->devargs);
1768 	if (err) {
1769 		err = rte_errno;
1770 		DRV_LOG(ERR, "failed to process device arguments: %s",
1771 			strerror(rte_errno));
1772 		goto error;
1773 	}
1774 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1775 			    IBV_DEVICE_RAW_IP_CSUM);
1776 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1777 		(config.hw_csum ? "" : "not "));
1778 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1779 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1780 	DRV_LOG(DEBUG, "counters are not supported");
1781 #endif
1782 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1783 	if (config.dv_flow_en) {
1784 		DRV_LOG(WARNING, "DV flow is not supported");
1785 		config.dv_flow_en = 0;
1786 	}
1787 #endif
1788 	config.ind_table_max_size =
1789 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1790 	/*
1791 	 * Remove this check once DPDK supports larger/variable
1792 	 * indirection tables.
1793 	 */
1794 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1795 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1796 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1797 		config.ind_table_max_size);
1798 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1799 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1800 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1801 		(config.hw_vlan_strip ? "" : "not "));
1802 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1803 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1804 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1805 		(config.hw_fcs_strip ? "" : "not "));
1806 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1807 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1808 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1809 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1810 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1811 #endif
1812 	if (config.hw_padding && !hw_padding) {
1813 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1814 		config.hw_padding = 0;
1815 	} else if (config.hw_padding) {
1816 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1817 	}
1818 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1819 		      (sh->device_attr.tso_caps.supported_qpts &
1820 		       (1 << IBV_QPT_RAW_PACKET)));
1821 	if (config.tso)
1822 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1823 	/*
1824 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1825 	 * by default.
1826 	 */
1827 	if (config.mps == MLX5_ARG_UNSET)
1828 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1829 							  MLX5_MPW_DISABLED;
1830 	else
1831 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1832 	DRV_LOG(INFO, "%sMPS is %s",
1833 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1834 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1835 	if (config.cqe_comp && !cqe_comp) {
1836 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1837 		config.cqe_comp = 0;
1838 	}
1839 	if (config.cqe_pad && !cqe_pad) {
1840 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1841 		config.cqe_pad = 0;
1842 	} else if (config.cqe_pad) {
1843 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1844 	}
1845 	if (config.devx) {
1846 		priv->counter_fallback = 0;
1847 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
1848 		if (err) {
1849 			err = -err;
1850 			goto error;
1851 		}
1852 		if (!config.hca_attr.flow_counters_dump)
1853 			priv->counter_fallback = 1;
1854 #ifndef HAVE_IBV_DEVX_ASYNC
1855 		priv->counter_fallback = 1;
1856 #endif
1857 		if (priv->counter_fallback)
1858 			DRV_LOG(INFO, "Use fall-back DV counter management\n");
1859 		/* Check for LRO support. */
1860 		if (config.dest_tir && config.hca_attr.lro_cap) {
1861 			/* TBD check tunnel lro caps. */
1862 			config.lro.supported = config.hca_attr.lro_cap;
1863 			DRV_LOG(DEBUG, "Device supports LRO");
1864 			/*
1865 			 * If LRO timeout is not configured by application,
1866 			 * use the minimal supported value.
1867 			 */
1868 			if (!config.lro.timeout)
1869 				config.lro.timeout =
1870 				config.hca_attr.lro_timer_supported_periods[0];
1871 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1872 				config.lro.timeout);
1873 		}
1874 	}
1875 	if (config.mprq.enabled && mprq) {
1876 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1877 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1878 			config.mprq.stride_num_n =
1879 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1880 					mprq_min_stride_num_n);
1881 			DRV_LOG(WARNING,
1882 				"the number of strides"
1883 				" for Multi-Packet RQ is out of range,"
1884 				" setting default value (%u)",
1885 				1 << config.mprq.stride_num_n);
1886 		}
1887 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1888 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1889 	} else if (config.mprq.enabled && !mprq) {
1890 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1891 		config.mprq.enabled = 0;
1892 	}
1893 	if (config.max_dump_files_num == 0)
1894 		config.max_dump_files_num = 128;
1895 	eth_dev = rte_eth_dev_allocate(name);
1896 	if (eth_dev == NULL) {
1897 		DRV_LOG(ERR, "can not allocate rte ethdev");
1898 		err = ENOMEM;
1899 		goto error;
1900 	}
1901 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1902 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1903 	if (priv->representor) {
1904 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1905 		eth_dev->data->representor_id = priv->representor_id;
1906 	}
1907 	/*
1908 	 * Store associated network device interface index. This index
1909 	 * is permanent throughout the lifetime of device. So, we may store
1910 	 * the ifindex here and use the cached value further.
1911 	 */
1912 	assert(spawn->ifindex);
1913 	priv->if_index = spawn->ifindex;
1914 	eth_dev->data->dev_private = priv;
1915 	priv->dev_data = eth_dev->data;
1916 	eth_dev->data->mac_addrs = priv->mac;
1917 	eth_dev->device = dpdk_dev;
1918 	/* Configure the first MAC address by default. */
1919 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1920 		DRV_LOG(ERR,
1921 			"port %u cannot get MAC address, is mlx5_en"
1922 			" loaded? (errno: %s)",
1923 			eth_dev->data->port_id, strerror(rte_errno));
1924 		err = ENODEV;
1925 		goto error;
1926 	}
1927 	DRV_LOG(INFO,
1928 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1929 		eth_dev->data->port_id,
1930 		mac.addr_bytes[0], mac.addr_bytes[1],
1931 		mac.addr_bytes[2], mac.addr_bytes[3],
1932 		mac.addr_bytes[4], mac.addr_bytes[5]);
1933 #ifndef NDEBUG
1934 	{
1935 		char ifname[IF_NAMESIZE];
1936 
1937 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1938 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1939 				eth_dev->data->port_id, ifname);
1940 		else
1941 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1942 				eth_dev->data->port_id);
1943 	}
1944 #endif
1945 	/* Get actual MTU if possible. */
1946 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1947 	if (err) {
1948 		err = rte_errno;
1949 		goto error;
1950 	}
1951 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1952 		priv->mtu);
1953 	/* Initialize burst functions to prevent crashes before link-up. */
1954 	eth_dev->rx_pkt_burst = removed_rx_burst;
1955 	eth_dev->tx_pkt_burst = removed_tx_burst;
1956 	eth_dev->dev_ops = &mlx5_dev_ops;
1957 	/* Register MAC address. */
1958 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1959 	if (config.vf && config.vf_nl_en)
1960 		mlx5_nl_mac_addr_sync(eth_dev);
1961 	TAILQ_INIT(&priv->flows);
1962 	TAILQ_INIT(&priv->ctrl_flows);
1963 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1964 	struct mlx5dv_ctx_allocators alctr = {
1965 		.alloc = &mlx5_alloc_verbs_buf,
1966 		.free = &mlx5_free_verbs_buf,
1967 		.data = priv,
1968 	};
1969 	mlx5_glue->dv_set_context_attr(sh->ctx,
1970 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1971 				       (void *)((uintptr_t)&alctr));
1972 	/* Bring Ethernet device up. */
1973 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1974 		eth_dev->data->port_id);
1975 	mlx5_set_link_up(eth_dev);
1976 	/*
1977 	 * Even though the interrupt handler is not installed yet,
1978 	 * interrupts will still trigger on the async_fd from
1979 	 * Verbs context returned by ibv_open_device().
1980 	 */
1981 	mlx5_link_update(eth_dev, 0);
1982 #ifdef HAVE_MLX5DV_DR_ESWITCH
1983 	if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1984 	      (switch_info->representor || switch_info->master)))
1985 		config.dv_esw_en = 0;
1986 #else
1987 	config.dv_esw_en = 0;
1988 #endif
1989 	/* Detect minimal data bytes to inline. */
1990 	mlx5_set_min_inline(spawn, &config);
1991 	/* Store device configuration on private structure. */
1992 	priv->config = config;
1993 	if (config.dv_flow_en) {
1994 		err = mlx5_alloc_shared_dr(priv);
1995 		if (err)
1996 			goto error;
1997 	}
1998 	/* Supported Verbs flow priority number detection. */
1999 	err = mlx5_flow_discover_priorities(eth_dev);
2000 	if (err < 0) {
2001 		err = -err;
2002 		goto error;
2003 	}
2004 	priv->config.flow_prio = err;
2005 	/* Add device to memory callback list. */
2006 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
2007 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
2008 			 sh, mem_event_cb);
2009 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
2010 	return eth_dev;
2011 error:
2012 	if (priv) {
2013 		if (priv->sh)
2014 			mlx5_free_shared_dr(priv);
2015 		if (priv->nl_socket_route >= 0)
2016 			close(priv->nl_socket_route);
2017 		if (priv->nl_socket_rdma >= 0)
2018 			close(priv->nl_socket_rdma);
2019 		if (own_domain_id)
2020 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2021 		rte_free(priv);
2022 		if (eth_dev != NULL)
2023 			eth_dev->data->dev_private = NULL;
2024 	}
2025 	if (eth_dev != NULL) {
2026 		/* mac_addrs must not be freed alone because part of dev_private */
2027 		eth_dev->data->mac_addrs = NULL;
2028 		rte_eth_dev_release_port(eth_dev);
2029 	}
2030 	if (sh)
2031 		mlx5_free_shared_ibctx(sh);
2032 	assert(err > 0);
2033 	rte_errno = err;
2034 	return NULL;
2035 }
2036 
2037 /**
2038  * Comparison callback to sort device data.
2039  *
2040  * This is meant to be used with qsort().
2041  *
2042  * @param a[in]
2043  *   Pointer to pointer to first data object.
2044  * @param b[in]
2045  *   Pointer to pointer to second data object.
2046  *
2047  * @return
2048  *   0 if both objects are equal, less than 0 if the first argument is less
2049  *   than the second, greater than 0 otherwise.
2050  */
2051 static int
2052 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2053 {
2054 	const struct mlx5_switch_info *si_a =
2055 		&((const struct mlx5_dev_spawn_data *)a)->info;
2056 	const struct mlx5_switch_info *si_b =
2057 		&((const struct mlx5_dev_spawn_data *)b)->info;
2058 	int ret;
2059 
2060 	/* Master device first. */
2061 	ret = si_b->master - si_a->master;
2062 	if (ret)
2063 		return ret;
2064 	/* Then representor devices. */
2065 	ret = si_b->representor - si_a->representor;
2066 	if (ret)
2067 		return ret;
2068 	/* Unidentified devices come last in no specific order. */
2069 	if (!si_a->representor)
2070 		return 0;
2071 	/* Order representors by name. */
2072 	return si_a->port_name - si_b->port_name;
2073 }
2074 
2075 /**
2076  * DPDK callback to register a PCI device.
2077  *
2078  * This function spawns Ethernet devices out of a given PCI device.
2079  *
2080  * @param[in] pci_drv
2081  *   PCI driver structure (mlx5_driver).
2082  * @param[in] pci_dev
2083  *   PCI device information.
2084  *
2085  * @return
2086  *   0 on success, a negative errno value otherwise and rte_errno is set.
2087  */
2088 static int
2089 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2090 	       struct rte_pci_device *pci_dev)
2091 {
2092 	struct ibv_device **ibv_list;
2093 	/*
2094 	 * Number of found IB Devices matching with requested PCI BDF.
2095 	 * nd != 1 means there are multiple IB devices over the same
2096 	 * PCI device and we have representors and master.
2097 	 */
2098 	unsigned int nd = 0;
2099 	/*
2100 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2101 	 * we have the single multiport IB device, and there may be
2102 	 * representors attached to some of found ports.
2103 	 */
2104 	unsigned int np = 0;
2105 	/*
2106 	 * Number of DPDK ethernet devices to Spawn - either over
2107 	 * multiple IB devices or multiple ports of single IB device.
2108 	 * Actually this is the number of iterations to spawn.
2109 	 */
2110 	unsigned int ns = 0;
2111 	struct mlx5_dev_config dev_config;
2112 	int ret;
2113 
2114 	ret = mlx5_init_once();
2115 	if (ret) {
2116 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2117 			strerror(rte_errno));
2118 		return -rte_errno;
2119 	}
2120 	assert(pci_drv == &mlx5_driver);
2121 	errno = 0;
2122 	ibv_list = mlx5_glue->get_device_list(&ret);
2123 	if (!ibv_list) {
2124 		rte_errno = errno ? errno : ENOSYS;
2125 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2126 		return -rte_errno;
2127 	}
2128 	/*
2129 	 * First scan the list of all Infiniband devices to find
2130 	 * matching ones, gathering into the list.
2131 	 */
2132 	struct ibv_device *ibv_match[ret + 1];
2133 	int nl_route = -1;
2134 	int nl_rdma = -1;
2135 	unsigned int i;
2136 
2137 	while (ret-- > 0) {
2138 		struct rte_pci_addr pci_addr;
2139 
2140 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2141 		if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
2142 			continue;
2143 		if (pci_dev->addr.domain != pci_addr.domain ||
2144 		    pci_dev->addr.bus != pci_addr.bus ||
2145 		    pci_dev->addr.devid != pci_addr.devid ||
2146 		    pci_dev->addr.function != pci_addr.function)
2147 			continue;
2148 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2149 			ibv_list[ret]->name);
2150 		ibv_match[nd++] = ibv_list[ret];
2151 	}
2152 	ibv_match[nd] = NULL;
2153 	if (!nd) {
2154 		/* No device matches, just complain and bail out. */
2155 		mlx5_glue->free_device_list(ibv_list);
2156 		DRV_LOG(WARNING,
2157 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2158 			" are kernel drivers loaded?",
2159 			pci_dev->addr.domain, pci_dev->addr.bus,
2160 			pci_dev->addr.devid, pci_dev->addr.function);
2161 		rte_errno = ENOENT;
2162 		ret = -rte_errno;
2163 		return ret;
2164 	}
2165 	nl_route = mlx5_nl_init(NETLINK_ROUTE);
2166 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2167 	if (nd == 1) {
2168 		/*
2169 		 * Found single matching device may have multiple ports.
2170 		 * Each port may be representor, we have to check the port
2171 		 * number and check the representors existence.
2172 		 */
2173 		if (nl_rdma >= 0)
2174 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2175 		if (!np)
2176 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2177 					 " ports number", ibv_match[0]->name);
2178 	}
2179 	/*
2180 	 * Now we can determine the maximal
2181 	 * amount of devices to be spawned.
2182 	 */
2183 	struct mlx5_dev_spawn_data list[np ? np : nd];
2184 
2185 	if (np > 1) {
2186 		/*
2187 		 * Single IB device with multiple ports found,
2188 		 * it may be E-Switch master device and representors.
2189 		 * We have to perform identification trough the ports.
2190 		 */
2191 		assert(nl_rdma >= 0);
2192 		assert(ns == 0);
2193 		assert(nd == 1);
2194 		for (i = 1; i <= np; ++i) {
2195 			list[ns].max_port = np;
2196 			list[ns].ibv_port = i;
2197 			list[ns].ibv_dev = ibv_match[0];
2198 			list[ns].eth_dev = NULL;
2199 			list[ns].pci_dev = pci_dev;
2200 			list[ns].ifindex = mlx5_nl_ifindex
2201 					(nl_rdma, list[ns].ibv_dev->name, i);
2202 			if (!list[ns].ifindex) {
2203 				/*
2204 				 * No network interface index found for the
2205 				 * specified port, it means there is no
2206 				 * representor on this port. It's OK,
2207 				 * there can be disabled ports, for example
2208 				 * if sriov_numvfs < sriov_totalvfs.
2209 				 */
2210 				continue;
2211 			}
2212 			ret = -1;
2213 			if (nl_route >= 0)
2214 				ret = mlx5_nl_switch_info
2215 					       (nl_route,
2216 						list[ns].ifindex,
2217 						&list[ns].info);
2218 			if (ret || (!list[ns].info.representor &&
2219 				    !list[ns].info.master)) {
2220 				/*
2221 				 * We failed to recognize representors with
2222 				 * Netlink, let's try to perform the task
2223 				 * with sysfs.
2224 				 */
2225 				ret =  mlx5_sysfs_switch_info
2226 						(list[ns].ifindex,
2227 						 &list[ns].info);
2228 			}
2229 			if (!ret && (list[ns].info.representor ^
2230 				     list[ns].info.master))
2231 				ns++;
2232 		}
2233 		if (!ns) {
2234 			DRV_LOG(ERR,
2235 				"unable to recognize master/representors"
2236 				" on the IB device with multiple ports");
2237 			rte_errno = ENOENT;
2238 			ret = -rte_errno;
2239 			goto exit;
2240 		}
2241 	} else {
2242 		/*
2243 		 * The existence of several matching entries (nd > 1) means
2244 		 * port representors have been instantiated. No existing Verbs
2245 		 * call nor sysfs entries can tell them apart, this can only
2246 		 * be done through Netlink calls assuming kernel drivers are
2247 		 * recent enough to support them.
2248 		 *
2249 		 * In the event of identification failure through Netlink,
2250 		 * try again through sysfs, then:
2251 		 *
2252 		 * 1. A single IB device matches (nd == 1) with single
2253 		 *    port (np=0/1) and is not a representor, assume
2254 		 *    no switch support.
2255 		 *
2256 		 * 2. Otherwise no safe assumptions can be made;
2257 		 *    complain louder and bail out.
2258 		 */
2259 		np = 1;
2260 		for (i = 0; i != nd; ++i) {
2261 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2262 			list[ns].max_port = 1;
2263 			list[ns].ibv_port = 1;
2264 			list[ns].ibv_dev = ibv_match[i];
2265 			list[ns].eth_dev = NULL;
2266 			list[ns].pci_dev = pci_dev;
2267 			list[ns].ifindex = 0;
2268 			if (nl_rdma >= 0)
2269 				list[ns].ifindex = mlx5_nl_ifindex
2270 					(nl_rdma, list[ns].ibv_dev->name, 1);
2271 			if (!list[ns].ifindex) {
2272 				char ifname[IF_NAMESIZE];
2273 
2274 				/*
2275 				 * Netlink failed, it may happen with old
2276 				 * ib_core kernel driver (before 4.16).
2277 				 * We can assume there is old driver because
2278 				 * here we are processing single ports IB
2279 				 * devices. Let's try sysfs to retrieve
2280 				 * the ifindex. The method works for
2281 				 * master device only.
2282 				 */
2283 				if (nd > 1) {
2284 					/*
2285 					 * Multiple devices found, assume
2286 					 * representors, can not distinguish
2287 					 * master/representor and retrieve
2288 					 * ifindex via sysfs.
2289 					 */
2290 					continue;
2291 				}
2292 				ret = mlx5_get_master_ifname
2293 					(ibv_match[i]->ibdev_path, &ifname);
2294 				if (!ret)
2295 					list[ns].ifindex =
2296 						if_nametoindex(ifname);
2297 				if (!list[ns].ifindex) {
2298 					/*
2299 					 * No network interface index found
2300 					 * for the specified device, it means
2301 					 * there it is neither representor
2302 					 * nor master.
2303 					 */
2304 					continue;
2305 				}
2306 			}
2307 			ret = -1;
2308 			if (nl_route >= 0)
2309 				ret = mlx5_nl_switch_info
2310 					       (nl_route,
2311 						list[ns].ifindex,
2312 						&list[ns].info);
2313 			if (ret || (!list[ns].info.representor &&
2314 				    !list[ns].info.master)) {
2315 				/*
2316 				 * We failed to recognize representors with
2317 				 * Netlink, let's try to perform the task
2318 				 * with sysfs.
2319 				 */
2320 				ret =  mlx5_sysfs_switch_info
2321 						(list[ns].ifindex,
2322 						 &list[ns].info);
2323 			}
2324 			if (!ret && (list[ns].info.representor ^
2325 				     list[ns].info.master)) {
2326 				ns++;
2327 			} else if ((nd == 1) &&
2328 				   !list[ns].info.representor &&
2329 				   !list[ns].info.master) {
2330 				/*
2331 				 * Single IB device with
2332 				 * one physical port and
2333 				 * attached network device.
2334 				 * May be SRIOV is not enabled
2335 				 * or there is no representors.
2336 				 */
2337 				DRV_LOG(INFO, "no E-Switch support detected");
2338 				ns++;
2339 				break;
2340 			}
2341 		}
2342 		if (!ns) {
2343 			DRV_LOG(ERR,
2344 				"unable to recognize master/representors"
2345 				" on the multiple IB devices");
2346 			rte_errno = ENOENT;
2347 			ret = -rte_errno;
2348 			goto exit;
2349 		}
2350 	}
2351 	assert(ns);
2352 	/*
2353 	 * Sort list to probe devices in natural order for users convenience
2354 	 * (i.e. master first, then representors from lowest to highest ID).
2355 	 */
2356 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2357 	/* Default configuration. */
2358 	dev_config = (struct mlx5_dev_config){
2359 		.hw_padding = 0,
2360 		.mps = MLX5_ARG_UNSET,
2361 		.rx_vec_en = 1,
2362 		.txq_inline_max = MLX5_ARG_UNSET,
2363 		.txq_inline_min = MLX5_ARG_UNSET,
2364 		.txq_inline_mpw = MLX5_ARG_UNSET,
2365 		.txqs_inline = MLX5_ARG_UNSET,
2366 		.vf_nl_en = 1,
2367 		.mr_ext_memseg_en = 1,
2368 		.mprq = {
2369 			.enabled = 0, /* Disabled by default. */
2370 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
2371 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
2372 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
2373 		},
2374 		.dv_esw_en = 1,
2375 	};
2376 	/* Device specific configuration. */
2377 	switch (pci_dev->id.device_id) {
2378 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2379 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2380 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2381 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2382 		dev_config.vf = 1;
2383 		break;
2384 	default:
2385 		break;
2386 	}
2387 	for (i = 0; i != ns; ++i) {
2388 		uint32_t restore;
2389 
2390 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2391 						 &list[i],
2392 						 dev_config);
2393 		if (!list[i].eth_dev) {
2394 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2395 				break;
2396 			/* Device is disabled or already spawned. Ignore it. */
2397 			continue;
2398 		}
2399 		restore = list[i].eth_dev->data->dev_flags;
2400 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2401 		/* Restore non-PCI flags cleared by the above call. */
2402 		list[i].eth_dev->data->dev_flags |= restore;
2403 		rte_eth_dev_probing_finish(list[i].eth_dev);
2404 	}
2405 	if (i != ns) {
2406 		DRV_LOG(ERR,
2407 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2408 			" encountering an error: %s",
2409 			pci_dev->addr.domain, pci_dev->addr.bus,
2410 			pci_dev->addr.devid, pci_dev->addr.function,
2411 			strerror(rte_errno));
2412 		ret = -rte_errno;
2413 		/* Roll back. */
2414 		while (i--) {
2415 			if (!list[i].eth_dev)
2416 				continue;
2417 			mlx5_dev_close(list[i].eth_dev);
2418 			/* mac_addrs must not be freed because in dev_private */
2419 			list[i].eth_dev->data->mac_addrs = NULL;
2420 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2421 		}
2422 		/* Restore original error. */
2423 		rte_errno = -ret;
2424 	} else {
2425 		ret = 0;
2426 	}
2427 exit:
2428 	/*
2429 	 * Do the routine cleanup:
2430 	 * - close opened Netlink sockets
2431 	 * - free the Infiniband device list
2432 	 */
2433 	if (nl_rdma >= 0)
2434 		close(nl_rdma);
2435 	if (nl_route >= 0)
2436 		close(nl_route);
2437 	assert(ibv_list);
2438 	mlx5_glue->free_device_list(ibv_list);
2439 	return ret;
2440 }
2441 
2442 /**
2443  * DPDK callback to remove a PCI device.
2444  *
2445  * This function removes all Ethernet devices belong to a given PCI device.
2446  *
2447  * @param[in] pci_dev
2448  *   Pointer to the PCI device.
2449  *
2450  * @return
2451  *   0 on success, the function cannot fail.
2452  */
2453 static int
2454 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2455 {
2456 	uint16_t port_id;
2457 
2458 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2459 		rte_eth_dev_close(port_id);
2460 	return 0;
2461 }
2462 
2463 static const struct rte_pci_id mlx5_pci_id_map[] = {
2464 	{
2465 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2466 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2467 	},
2468 	{
2469 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2470 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2471 	},
2472 	{
2473 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2474 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2475 	},
2476 	{
2477 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2478 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2479 	},
2480 	{
2481 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2482 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2483 	},
2484 	{
2485 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2486 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2487 	},
2488 	{
2489 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2490 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2491 	},
2492 	{
2493 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2494 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2495 	},
2496 	{
2497 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2498 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2499 	},
2500 	{
2501 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2502 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2503 	},
2504 	{
2505 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2506 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2507 	},
2508 	{
2509 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2510 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2511 	},
2512 	{
2513 		.vendor_id = 0
2514 	}
2515 };
2516 
2517 static struct rte_pci_driver mlx5_driver = {
2518 	.driver = {
2519 		.name = MLX5_DRIVER_NAME
2520 	},
2521 	.id_table = mlx5_pci_id_map,
2522 	.probe = mlx5_pci_probe,
2523 	.remove = mlx5_pci_remove,
2524 	.dma_map = mlx5_dma_map,
2525 	.dma_unmap = mlx5_dma_unmap,
2526 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2527 		     RTE_PCI_DRV_PROBE_AGAIN,
2528 };
2529 
2530 #ifdef RTE_IBVERBS_LINK_DLOPEN
2531 
2532 /**
2533  * Suffix RTE_EAL_PMD_PATH with "-glue".
2534  *
2535  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2536  * suffixing its last component.
2537  *
2538  * @param buf[out]
2539  *   Output buffer, should be large enough otherwise NULL is returned.
2540  * @param size
2541  *   Size of @p out.
2542  *
2543  * @return
2544  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2545  */
2546 static char *
2547 mlx5_glue_path(char *buf, size_t size)
2548 {
2549 	static const char *const bad[] = { "/", ".", "..", NULL };
2550 	const char *path = RTE_EAL_PMD_PATH;
2551 	size_t len = strlen(path);
2552 	size_t off;
2553 	int i;
2554 
2555 	while (len && path[len - 1] == '/')
2556 		--len;
2557 	for (off = len; off && path[off - 1] != '/'; --off)
2558 		;
2559 	for (i = 0; bad[i]; ++i)
2560 		if (!strncmp(path + off, bad[i], (int)(len - off)))
2561 			goto error;
2562 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2563 	if (i == -1 || (size_t)i >= size)
2564 		goto error;
2565 	return buf;
2566 error:
2567 	DRV_LOG(ERR,
2568 		"unable to append \"-glue\" to last component of"
2569 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2570 		" please re-configure DPDK");
2571 	return NULL;
2572 }
2573 
2574 /**
2575  * Initialization routine for run-time dependency on rdma-core.
2576  */
2577 static int
2578 mlx5_glue_init(void)
2579 {
2580 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2581 	const char *path[] = {
2582 		/*
2583 		 * A basic security check is necessary before trusting
2584 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2585 		 */
2586 		(geteuid() == getuid() && getegid() == getgid() ?
2587 		 getenv("MLX5_GLUE_PATH") : NULL),
2588 		/*
2589 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2590 		 * variant, otherwise let dlopen() look up libraries on its
2591 		 * own.
2592 		 */
2593 		(*RTE_EAL_PMD_PATH ?
2594 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2595 	};
2596 	unsigned int i = 0;
2597 	void *handle = NULL;
2598 	void **sym;
2599 	const char *dlmsg;
2600 
2601 	while (!handle && i != RTE_DIM(path)) {
2602 		const char *end;
2603 		size_t len;
2604 		int ret;
2605 
2606 		if (!path[i]) {
2607 			++i;
2608 			continue;
2609 		}
2610 		end = strpbrk(path[i], ":;");
2611 		if (!end)
2612 			end = path[i] + strlen(path[i]);
2613 		len = end - path[i];
2614 		ret = 0;
2615 		do {
2616 			char name[ret + 1];
2617 
2618 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2619 				       (int)len, path[i],
2620 				       (!len || *(end - 1) == '/') ? "" : "/");
2621 			if (ret == -1)
2622 				break;
2623 			if (sizeof(name) != (size_t)ret + 1)
2624 				continue;
2625 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2626 				name);
2627 			handle = dlopen(name, RTLD_LAZY);
2628 			break;
2629 		} while (1);
2630 		path[i] = end + 1;
2631 		if (!*end)
2632 			++i;
2633 	}
2634 	if (!handle) {
2635 		rte_errno = EINVAL;
2636 		dlmsg = dlerror();
2637 		if (dlmsg)
2638 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2639 		goto glue_error;
2640 	}
2641 	sym = dlsym(handle, "mlx5_glue");
2642 	if (!sym || !*sym) {
2643 		rte_errno = EINVAL;
2644 		dlmsg = dlerror();
2645 		if (dlmsg)
2646 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2647 		goto glue_error;
2648 	}
2649 	mlx5_glue = *sym;
2650 	return 0;
2651 glue_error:
2652 	if (handle)
2653 		dlclose(handle);
2654 	DRV_LOG(WARNING,
2655 		"cannot initialize PMD due to missing run-time dependency on"
2656 		" rdma-core libraries (libibverbs, libmlx5)");
2657 	return -rte_errno;
2658 }
2659 
2660 #endif
2661 
2662 /**
2663  * Driver initialization routine.
2664  */
2665 RTE_INIT(rte_mlx5_pmd_init)
2666 {
2667 	/* Initialize driver log type. */
2668 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
2669 	if (mlx5_logtype >= 0)
2670 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2671 
2672 	/* Build the static tables for Verbs conversion. */
2673 	mlx5_set_ptype_table();
2674 	mlx5_set_cksum_table();
2675 	mlx5_set_swp_types_table();
2676 	/*
2677 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2678 	 * huge pages. Calling ibv_fork_init() during init allows
2679 	 * applications to use fork() safely for purposes other than
2680 	 * using this PMD, which is not supported in forked processes.
2681 	 */
2682 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2683 	/* Match the size of Rx completion entry to the size of a cacheline. */
2684 	if (RTE_CACHE_LINE_SIZE == 128)
2685 		setenv("MLX5_CQE_SIZE", "128", 0);
2686 	/*
2687 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2688 	 * cleanup all the Verbs resources even when the device was removed.
2689 	 */
2690 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2691 #ifdef RTE_IBVERBS_LINK_DLOPEN
2692 	if (mlx5_glue_init())
2693 		return;
2694 	assert(mlx5_glue);
2695 #endif
2696 #ifndef NDEBUG
2697 	/* Glue structure must not contain any NULL pointers. */
2698 	{
2699 		unsigned int i;
2700 
2701 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2702 			assert(((const void *const *)mlx5_glue)[i]);
2703 	}
2704 #endif
2705 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2706 		DRV_LOG(ERR,
2707 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
2708 			mlx5_glue->version, MLX5_GLUE_VERSION);
2709 		return;
2710 	}
2711 	mlx5_glue->fork_init();
2712 	rte_pci_register(&mlx5_driver);
2713 }
2714 
2715 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2716 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2717 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
2718