xref: /dpdk/drivers/net/mlx4/mlx4.c (revision 0d09cbc7)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2012 6WIND S.A.
3  * Copyright 2012 Mellanox Technologies, Ltd
4  */
5 
6 /**
7  * @file
8  * mlx4 driver initialization.
9  */
10 
11 #include <errno.h>
12 #include <inttypes.h>
13 #include <stddef.h>
14 #include <stdint.h>
15 #include <stdio.h>
16 #include <stdlib.h>
17 #include <string.h>
18 #include <sys/mman.h>
19 #include <unistd.h>
20 #ifdef RTE_IBVERBS_LINK_DLOPEN
21 #include <dlfcn.h>
22 #endif
23 
24 /* Verbs headers do not support -pedantic. */
25 #ifdef PEDANTIC
26 #pragma GCC diagnostic ignored "-Wpedantic"
27 #endif
28 #include <infiniband/verbs.h>
29 #ifdef PEDANTIC
30 #pragma GCC diagnostic error "-Wpedantic"
31 #endif
32 
33 #include <rte_common.h>
34 #include <rte_dev.h>
35 #include <rte_errno.h>
36 #include <rte_ethdev_driver.h>
37 #include <rte_ethdev_pci.h>
38 #include <rte_ether.h>
39 #include <rte_flow.h>
40 #include <rte_interrupts.h>
41 #include <rte_kvargs.h>
42 #include <rte_malloc.h>
43 #include <rte_mbuf.h>
44 
45 #include "mlx4.h"
46 #include "mlx4_glue.h"
47 #include "mlx4_flow.h"
48 #include "mlx4_mr.h"
49 #include "mlx4_rxtx.h"
50 #include "mlx4_utils.h"
51 
52 #ifdef MLX4_GLUE
53 const struct mlx4_glue *mlx4_glue;
54 #endif
55 
56 static const char *MZ_MLX4_PMD_SHARED_DATA = "mlx4_pmd_shared_data";
57 
58 /* Shared memory between primary and secondary processes. */
59 struct mlx4_shared_data *mlx4_shared_data;
60 
61 /* Spinlock for mlx4_shared_data allocation. */
62 static rte_spinlock_t mlx4_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
63 
64 /* Process local data for secondary processes. */
65 static struct mlx4_local_data mlx4_local_data;
66 
67 /** Configuration structure for device arguments. */
68 struct mlx4_conf {
69 	struct {
70 		uint32_t present; /**< Bit-field for existing ports. */
71 		uint32_t enabled; /**< Bit-field for user-enabled ports. */
72 	} ports;
73 	int mr_ext_memseg_en;
74 	/** Whether memseg should be extended for MR creation. */
75 };
76 
77 /* Available parameters list. */
78 const char *pmd_mlx4_init_params[] = {
79 	MLX4_PMD_PORT_KVARG,
80 	MLX4_MR_EXT_MEMSEG_EN_KVARG,
81 	NULL,
82 };
83 
84 static void mlx4_dev_stop(struct rte_eth_dev *dev);
85 
86 /**
87  * Initialize shared data between primary and secondary process.
88  *
89  * A memzone is reserved by primary process and secondary processes attach to
90  * the memzone.
91  *
92  * @return
93  *   0 on success, a negative errno value otherwise and rte_errno is set.
94  */
95 static int
96 mlx4_init_shared_data(void)
97 {
98 	const struct rte_memzone *mz;
99 	int ret = 0;
100 
101 	rte_spinlock_lock(&mlx4_shared_data_lock);
102 	if (mlx4_shared_data == NULL) {
103 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
104 			/* Allocate shared memory. */
105 			mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA,
106 						 sizeof(*mlx4_shared_data),
107 						 SOCKET_ID_ANY, 0);
108 			if (mz == NULL) {
109 				ERROR("Cannot allocate mlx4 shared data\n");
110 				ret = -rte_errno;
111 				goto error;
112 			}
113 			mlx4_shared_data = mz->addr;
114 			memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data));
115 			rte_spinlock_init(&mlx4_shared_data->lock);
116 		} else {
117 			/* Lookup allocated shared memory. */
118 			mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
119 			if (mz == NULL) {
120 				ERROR("Cannot attach mlx4 shared data\n");
121 				ret = -rte_errno;
122 				goto error;
123 			}
124 			mlx4_shared_data = mz->addr;
125 			memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
126 		}
127 	}
128 error:
129 	rte_spinlock_unlock(&mlx4_shared_data_lock);
130 	return ret;
131 }
132 
133 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
134 /**
135  * Verbs callback to allocate a memory. This function should allocate the space
136  * according to the size provided residing inside a huge page.
137  * Please note that all allocation must respect the alignment from libmlx4
138  * (i.e. currently sysconf(_SC_PAGESIZE)).
139  *
140  * @param[in] size
141  *   The size in bytes of the memory to allocate.
142  * @param[in] data
143  *   A pointer to the callback data.
144  *
145  * @return
146  *   Allocated buffer, NULL otherwise and rte_errno is set.
147  */
148 static void *
149 mlx4_alloc_verbs_buf(size_t size, void *data)
150 {
151 	struct mlx4_priv *priv = data;
152 	void *ret;
153 	size_t alignment = sysconf(_SC_PAGESIZE);
154 	unsigned int socket = SOCKET_ID_ANY;
155 
156 	if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
157 		const struct txq *txq = priv->verbs_alloc_ctx.obj;
158 
159 		socket = txq->socket;
160 	} else if (priv->verbs_alloc_ctx.type ==
161 		   MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
162 		const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
163 
164 		socket = rxq->socket;
165 	}
166 	MLX4_ASSERT(data != NULL);
167 	ret = rte_malloc_socket(__func__, size, alignment, socket);
168 	if (!ret && size)
169 		rte_errno = ENOMEM;
170 	return ret;
171 }
172 
173 /**
174  * Verbs callback to free a memory.
175  *
176  * @param[in] ptr
177  *   A pointer to the memory to free.
178  * @param[in] data
179  *   A pointer to the callback data.
180  */
181 static void
182 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
183 {
184 	MLX4_ASSERT(data != NULL);
185 	rte_free(ptr);
186 }
187 #endif
188 
189 /**
190  * Initialize process private data structure.
191  *
192  * @param dev
193  *   Pointer to Ethernet device structure.
194  *
195  * @return
196  *   0 on success, a negative errno value otherwise and rte_errno is set.
197  */
198 static int
199 mlx4_proc_priv_init(struct rte_eth_dev *dev)
200 {
201 	struct mlx4_proc_priv *ppriv;
202 	size_t ppriv_size;
203 
204 	/*
205 	 * UAR register table follows the process private structure. BlueFlame
206 	 * registers for Tx queues are stored in the table.
207 	 */
208 	ppriv_size = sizeof(struct mlx4_proc_priv) +
209 		     dev->data->nb_tx_queues * sizeof(void *);
210 	ppriv = rte_malloc_socket("mlx4_proc_priv", ppriv_size,
211 				  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
212 	if (!ppriv) {
213 		rte_errno = ENOMEM;
214 		return -rte_errno;
215 	}
216 	ppriv->uar_table_sz = ppriv_size;
217 	dev->process_private = ppriv;
218 	return 0;
219 }
220 
221 /**
222  * Un-initialize process private data structure.
223  *
224  * @param dev
225  *   Pointer to Ethernet device structure.
226  */
227 static void
228 mlx4_proc_priv_uninit(struct rte_eth_dev *dev)
229 {
230 	if (!dev->process_private)
231 		return;
232 	rte_free(dev->process_private);
233 	dev->process_private = NULL;
234 }
235 
236 /**
237  * DPDK callback for Ethernet device configuration.
238  *
239  * @param dev
240  *   Pointer to Ethernet device structure.
241  *
242  * @return
243  *   0 on success, negative errno value otherwise and rte_errno is set.
244  */
245 static int
246 mlx4_dev_configure(struct rte_eth_dev *dev)
247 {
248 	struct mlx4_priv *priv = dev->data->dev_private;
249 	struct rte_flow_error error;
250 	int ret;
251 
252 	if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
253 		dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
254 
255 	/* Prepare internal flow rules. */
256 	ret = mlx4_flow_sync(priv, &error);
257 	if (ret) {
258 		ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
259 		      " flow error type %d, cause %p, message: %s",
260 		      -ret, strerror(-ret), error.type, error.cause,
261 		      error.message ? error.message : "(unspecified)");
262 		goto exit;
263 	}
264 	ret = mlx4_intr_install(priv);
265 	if (ret) {
266 		ERROR("%p: interrupt handler installation failed",
267 		      (void *)dev);
268 		goto exit;
269 	}
270 	ret = mlx4_proc_priv_init(dev);
271 	if (ret) {
272 		ERROR("%p: process private data allocation failed",
273 		      (void *)dev);
274 		goto exit;
275 	}
276 exit:
277 	return ret;
278 }
279 
280 /**
281  * DPDK callback to start the device.
282  *
283  * Simulate device start by initializing common RSS resources and attaching
284  * all configured flows.
285  *
286  * @param dev
287  *   Pointer to Ethernet device structure.
288  *
289  * @return
290  *   0 on success, negative errno value otherwise and rte_errno is set.
291  */
292 static int
293 mlx4_dev_start(struct rte_eth_dev *dev)
294 {
295 	struct mlx4_priv *priv = dev->data->dev_private;
296 	struct rte_flow_error error;
297 	int ret;
298 
299 	if (priv->started)
300 		return 0;
301 	DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
302 	priv->started = 1;
303 	ret = mlx4_rss_init(priv);
304 	if (ret) {
305 		ERROR("%p: cannot initialize RSS resources: %s",
306 		      (void *)dev, strerror(-ret));
307 		goto err;
308 	}
309 #ifdef RTE_LIBRTE_MLX4_DEBUG
310 	mlx4_mr_dump_dev(dev);
311 #endif
312 	ret = mlx4_rxq_intr_enable(priv);
313 	if (ret) {
314 		ERROR("%p: interrupt handler installation failed",
315 		     (void *)dev);
316 		goto err;
317 	}
318 	ret = mlx4_flow_sync(priv, &error);
319 	if (ret) {
320 		ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
321 		      " flow error type %d, cause %p, message: %s",
322 		      (void *)dev,
323 		      -ret, strerror(-ret), error.type, error.cause,
324 		      error.message ? error.message : "(unspecified)");
325 		goto err;
326 	}
327 	rte_wmb();
328 	dev->tx_pkt_burst = mlx4_tx_burst;
329 	dev->rx_pkt_burst = mlx4_rx_burst;
330 	/* Enable datapath on secondary process. */
331 	mlx4_mp_req_start_rxtx(dev);
332 	return 0;
333 err:
334 	mlx4_dev_stop(dev);
335 	return ret;
336 }
337 
338 /**
339  * DPDK callback to stop the device.
340  *
341  * Simulate device stop by detaching all configured flows.
342  *
343  * @param dev
344  *   Pointer to Ethernet device structure.
345  */
346 static void
347 mlx4_dev_stop(struct rte_eth_dev *dev)
348 {
349 	struct mlx4_priv *priv = dev->data->dev_private;
350 
351 	if (!priv->started)
352 		return;
353 	DEBUG("%p: detaching flows from all RX queues", (void *)dev);
354 	priv->started = 0;
355 	dev->tx_pkt_burst = mlx4_tx_burst_removed;
356 	dev->rx_pkt_burst = mlx4_rx_burst_removed;
357 	rte_wmb();
358 	/* Disable datapath on secondary process. */
359 	mlx4_mp_req_stop_rxtx(dev);
360 	mlx4_flow_sync(priv, NULL);
361 	mlx4_rxq_intr_disable(priv);
362 	mlx4_rss_deinit(priv);
363 }
364 
365 /**
366  * DPDK callback to close the device.
367  *
368  * Destroy all queues and objects, free memory.
369  *
370  * @param dev
371  *   Pointer to Ethernet device structure.
372  */
373 static void
374 mlx4_dev_close(struct rte_eth_dev *dev)
375 {
376 	struct mlx4_priv *priv = dev->data->dev_private;
377 	unsigned int i;
378 
379 	DEBUG("%p: closing device \"%s\"",
380 	      (void *)dev,
381 	      ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
382 	dev->rx_pkt_burst = mlx4_rx_burst_removed;
383 	dev->tx_pkt_burst = mlx4_tx_burst_removed;
384 	rte_wmb();
385 	/* Disable datapath on secondary process. */
386 	mlx4_mp_req_stop_rxtx(dev);
387 	mlx4_flow_clean(priv);
388 	mlx4_rss_deinit(priv);
389 	for (i = 0; i != dev->data->nb_rx_queues; ++i)
390 		mlx4_rx_queue_release(dev->data->rx_queues[i]);
391 	for (i = 0; i != dev->data->nb_tx_queues; ++i)
392 		mlx4_tx_queue_release(dev->data->tx_queues[i]);
393 	mlx4_proc_priv_uninit(dev);
394 	mlx4_mr_release(dev);
395 	if (priv->pd != NULL) {
396 		MLX4_ASSERT(priv->ctx != NULL);
397 		claim_zero(mlx4_glue->dealloc_pd(priv->pd));
398 		claim_zero(mlx4_glue->close_device(priv->ctx));
399 	} else
400 		MLX4_ASSERT(priv->ctx == NULL);
401 	mlx4_intr_uninstall(priv);
402 	memset(priv, 0, sizeof(*priv));
403 }
404 
405 static const struct eth_dev_ops mlx4_dev_ops = {
406 	.dev_configure = mlx4_dev_configure,
407 	.dev_start = mlx4_dev_start,
408 	.dev_stop = mlx4_dev_stop,
409 	.dev_set_link_down = mlx4_dev_set_link_down,
410 	.dev_set_link_up = mlx4_dev_set_link_up,
411 	.dev_close = mlx4_dev_close,
412 	.link_update = mlx4_link_update,
413 	.promiscuous_enable = mlx4_promiscuous_enable,
414 	.promiscuous_disable = mlx4_promiscuous_disable,
415 	.allmulticast_enable = mlx4_allmulticast_enable,
416 	.allmulticast_disable = mlx4_allmulticast_disable,
417 	.mac_addr_remove = mlx4_mac_addr_remove,
418 	.mac_addr_add = mlx4_mac_addr_add,
419 	.mac_addr_set = mlx4_mac_addr_set,
420 	.set_mc_addr_list = mlx4_set_mc_addr_list,
421 	.stats_get = mlx4_stats_get,
422 	.stats_reset = mlx4_stats_reset,
423 	.fw_version_get = mlx4_fw_version_get,
424 	.dev_infos_get = mlx4_dev_infos_get,
425 	.dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
426 	.vlan_filter_set = mlx4_vlan_filter_set,
427 	.rx_queue_setup = mlx4_rx_queue_setup,
428 	.tx_queue_setup = mlx4_tx_queue_setup,
429 	.rx_queue_release = mlx4_rx_queue_release,
430 	.tx_queue_release = mlx4_tx_queue_release,
431 	.flow_ctrl_get = mlx4_flow_ctrl_get,
432 	.flow_ctrl_set = mlx4_flow_ctrl_set,
433 	.mtu_set = mlx4_mtu_set,
434 	.filter_ctrl = mlx4_filter_ctrl,
435 	.rx_queue_intr_enable = mlx4_rx_intr_enable,
436 	.rx_queue_intr_disable = mlx4_rx_intr_disable,
437 	.is_removed = mlx4_is_removed,
438 };
439 
440 /* Available operations from secondary process. */
441 static const struct eth_dev_ops mlx4_dev_sec_ops = {
442 	.stats_get = mlx4_stats_get,
443 	.stats_reset = mlx4_stats_reset,
444 	.fw_version_get = mlx4_fw_version_get,
445 	.dev_infos_get = mlx4_dev_infos_get,
446 };
447 
448 /**
449  * Get PCI information from struct ibv_device.
450  *
451  * @param device
452  *   Pointer to Ethernet device structure.
453  * @param[out] pci_addr
454  *   PCI bus address output buffer.
455  *
456  * @return
457  *   0 on success, negative errno value otherwise and rte_errno is set.
458  */
459 static int
460 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
461 			    struct rte_pci_addr *pci_addr)
462 {
463 	FILE *file;
464 	char line[32];
465 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
466 
467 	file = fopen(path, "rb");
468 	if (file == NULL) {
469 		rte_errno = errno;
470 		return -rte_errno;
471 	}
472 	while (fgets(line, sizeof(line), file) == line) {
473 		size_t len = strlen(line);
474 		int ret;
475 
476 		/* Truncate long lines. */
477 		if (len == (sizeof(line) - 1))
478 			while (line[(len - 1)] != '\n') {
479 				ret = fgetc(file);
480 				if (ret == EOF)
481 					break;
482 				line[(len - 1)] = ret;
483 			}
484 		/* Extract information. */
485 		if (sscanf(line,
486 			   "PCI_SLOT_NAME="
487 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
488 			   &pci_addr->domain,
489 			   &pci_addr->bus,
490 			   &pci_addr->devid,
491 			   &pci_addr->function) == 4) {
492 			break;
493 		}
494 	}
495 	fclose(file);
496 	return 0;
497 }
498 
499 /**
500  * Verify and store value for device argument.
501  *
502  * @param[in] key
503  *   Key argument to verify.
504  * @param[in] val
505  *   Value associated with key.
506  * @param[in, out] conf
507  *   Shared configuration data.
508  *
509  * @return
510  *   0 on success, negative errno value otherwise and rte_errno is set.
511  */
512 static int
513 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
514 {
515 	unsigned long tmp;
516 
517 	errno = 0;
518 	tmp = strtoul(val, NULL, 0);
519 	if (errno) {
520 		rte_errno = errno;
521 		WARN("%s: \"%s\" is not a valid integer", key, val);
522 		return -rte_errno;
523 	}
524 	if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
525 		uint32_t ports = rte_log2_u32(conf->ports.present + 1);
526 
527 		if (tmp >= ports) {
528 			ERROR("port index %lu outside range [0,%" PRIu32 ")",
529 			      tmp, ports);
530 			return -EINVAL;
531 		}
532 		if (!(conf->ports.present & (1 << tmp))) {
533 			rte_errno = EINVAL;
534 			ERROR("invalid port index %lu", tmp);
535 			return -rte_errno;
536 		}
537 		conf->ports.enabled |= 1 << tmp;
538 	} else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) {
539 		conf->mr_ext_memseg_en = !!tmp;
540 	} else {
541 		rte_errno = EINVAL;
542 		WARN("%s: unknown parameter", key);
543 		return -rte_errno;
544 	}
545 	return 0;
546 }
547 
548 /**
549  * Parse device parameters.
550  *
551  * @param devargs
552  *   Device arguments structure.
553  *
554  * @return
555  *   0 on success, negative errno value otherwise and rte_errno is set.
556  */
557 static int
558 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
559 {
560 	struct rte_kvargs *kvlist;
561 	unsigned int arg_count;
562 	int ret = 0;
563 	int i;
564 
565 	if (devargs == NULL)
566 		return 0;
567 	kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
568 	if (kvlist == NULL) {
569 		rte_errno = EINVAL;
570 		ERROR("failed to parse kvargs");
571 		return -rte_errno;
572 	}
573 	/* Process parameters. */
574 	for (i = 0; pmd_mlx4_init_params[i]; ++i) {
575 		arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]);
576 		while (arg_count-- > 0) {
577 			ret = rte_kvargs_process(kvlist,
578 						 pmd_mlx4_init_params[i],
579 						 (int (*)(const char *,
580 							  const char *,
581 							  void *))
582 						 mlx4_arg_parse,
583 						 conf);
584 			if (ret != 0)
585 				goto free_kvlist;
586 		}
587 	}
588 free_kvlist:
589 	rte_kvargs_free(kvlist);
590 	return ret;
591 }
592 
593 /**
594  * Interpret RSS capabilities reported by device.
595  *
596  * This function returns the set of usable Verbs RSS hash fields, kernel
597  * quirks taken into account.
598  *
599  * @param ctx
600  *   Verbs context.
601  * @param pd
602  *   Verbs protection domain.
603  * @param device_attr_ex
604  *   Extended device attributes to interpret.
605  *
606  * @return
607  *   Usable RSS hash fields mask in Verbs format.
608  */
609 static uint64_t
610 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
611 		struct ibv_device_attr_ex *device_attr_ex)
612 {
613 	uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
614 	struct ibv_cq *cq = NULL;
615 	struct ibv_wq *wq = NULL;
616 	struct ibv_rwq_ind_table *ind = NULL;
617 	struct ibv_qp *qp = NULL;
618 
619 	if (!hw_rss_sup) {
620 		WARN("no RSS capabilities reported; disabling support for UDP"
621 		     " RSS and inner VXLAN RSS");
622 		return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
623 			IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
624 			IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
625 	}
626 	if (!(hw_rss_sup & IBV_RX_HASH_INNER))
627 		return hw_rss_sup;
628 	/*
629 	 * Although reported as supported, missing code in some Linux
630 	 * versions (v4.15, v4.16) prevents the creation of hash QPs with
631 	 * inner capability.
632 	 *
633 	 * There is no choice but to attempt to instantiate a temporary RSS
634 	 * context in order to confirm its support.
635 	 */
636 	cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
637 	wq = cq ? mlx4_glue->create_wq
638 		(ctx,
639 		 &(struct ibv_wq_init_attr){
640 			.wq_type = IBV_WQT_RQ,
641 			.max_wr = 1,
642 			.max_sge = 1,
643 			.pd = pd,
644 			.cq = cq,
645 		 }) : NULL;
646 	ind = wq ? mlx4_glue->create_rwq_ind_table
647 		(ctx,
648 		 &(struct ibv_rwq_ind_table_init_attr){
649 			.log_ind_tbl_size = 0,
650 			.ind_tbl = &wq,
651 			.comp_mask = 0,
652 		 }) : NULL;
653 	qp = ind ? mlx4_glue->create_qp_ex
654 		(ctx,
655 		 &(struct ibv_qp_init_attr_ex){
656 			.comp_mask =
657 				(IBV_QP_INIT_ATTR_PD |
658 				 IBV_QP_INIT_ATTR_RX_HASH |
659 				 IBV_QP_INIT_ATTR_IND_TABLE),
660 			.qp_type = IBV_QPT_RAW_PACKET,
661 			.pd = pd,
662 			.rwq_ind_tbl = ind,
663 			.rx_hash_conf = {
664 				.rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
665 				.rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
666 				.rx_hash_key = mlx4_rss_hash_key_default,
667 				.rx_hash_fields_mask = hw_rss_sup,
668 			},
669 		 }) : NULL;
670 	if (!qp) {
671 		WARN("disabling unusable inner RSS capability due to kernel"
672 		     " quirk");
673 		hw_rss_sup &= ~IBV_RX_HASH_INNER;
674 	} else {
675 		claim_zero(mlx4_glue->destroy_qp(qp));
676 	}
677 	if (ind)
678 		claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
679 	if (wq)
680 		claim_zero(mlx4_glue->destroy_wq(wq));
681 	if (cq)
682 		claim_zero(mlx4_glue->destroy_cq(cq));
683 	return hw_rss_sup;
684 }
685 
686 static struct rte_pci_driver mlx4_driver;
687 
688 /**
689  * PMD global initialization.
690  *
691  * Independent from individual device, this function initializes global
692  * per-PMD data structures distinguishing primary and secondary processes.
693  * Hence, each initialization is called once per a process.
694  *
695  * @return
696  *   0 on success, a negative errno value otherwise and rte_errno is set.
697  */
698 static int
699 mlx4_init_once(void)
700 {
701 	struct mlx4_shared_data *sd;
702 	struct mlx4_local_data *ld = &mlx4_local_data;
703 	int ret = 0;
704 
705 	if (mlx4_init_shared_data())
706 		return -rte_errno;
707 	sd = mlx4_shared_data;
708 	MLX4_ASSERT(sd);
709 	rte_spinlock_lock(&sd->lock);
710 	switch (rte_eal_process_type()) {
711 	case RTE_PROC_PRIMARY:
712 		if (sd->init_done)
713 			break;
714 		LIST_INIT(&sd->mem_event_cb_list);
715 		rte_rwlock_init(&sd->mem_event_rwlock);
716 		rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
717 						mlx4_mr_mem_event_cb, NULL);
718 		ret = mlx4_mp_init_primary();
719 		if (ret)
720 			goto out;
721 		sd->init_done = 1;
722 		break;
723 	case RTE_PROC_SECONDARY:
724 		if (ld->init_done)
725 			break;
726 		ret = mlx4_mp_init_secondary();
727 		if (ret)
728 			goto out;
729 		++sd->secondary_cnt;
730 		ld->init_done = 1;
731 		break;
732 	default:
733 		break;
734 	}
735 out:
736 	rte_spinlock_unlock(&sd->lock);
737 	return ret;
738 }
739 
740 /**
741  * DPDK callback to register a PCI device.
742  *
743  * This function creates an Ethernet device for each port of a given
744  * PCI device.
745  *
746  * @param[in] pci_drv
747  *   PCI driver structure (mlx4_driver).
748  * @param[in] pci_dev
749  *   PCI device information.
750  *
751  * @return
752  *   0 on success, negative errno value otherwise and rte_errno is set.
753  */
754 static int
755 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
756 {
757 	struct ibv_device **list;
758 	struct ibv_device *ibv_dev;
759 	int err = 0;
760 	struct ibv_context *attr_ctx = NULL;
761 	struct ibv_device_attr device_attr;
762 	struct ibv_device_attr_ex device_attr_ex;
763 	struct mlx4_conf conf = {
764 		.ports.present = 0,
765 		.mr_ext_memseg_en = 1,
766 	};
767 	unsigned int vf;
768 	int i;
769 	char ifname[IF_NAMESIZE];
770 
771 	(void)pci_drv;
772 	err = mlx4_init_once();
773 	if (err) {
774 		ERROR("unable to init PMD global data: %s",
775 		      strerror(rte_errno));
776 		return -rte_errno;
777 	}
778 	MLX4_ASSERT(pci_drv == &mlx4_driver);
779 	list = mlx4_glue->get_device_list(&i);
780 	if (list == NULL) {
781 		rte_errno = errno;
782 		MLX4_ASSERT(rte_errno);
783 		if (rte_errno == ENOSYS)
784 			ERROR("cannot list devices, is ib_uverbs loaded?");
785 		return -rte_errno;
786 	}
787 	MLX4_ASSERT(i >= 0);
788 	/*
789 	 * For each listed device, check related sysfs entry against
790 	 * the provided PCI ID.
791 	 */
792 	while (i != 0) {
793 		struct rte_pci_addr pci_addr;
794 
795 		--i;
796 		DEBUG("checking device \"%s\"", list[i]->name);
797 		if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
798 			continue;
799 		if ((pci_dev->addr.domain != pci_addr.domain) ||
800 		    (pci_dev->addr.bus != pci_addr.bus) ||
801 		    (pci_dev->addr.devid != pci_addr.devid) ||
802 		    (pci_dev->addr.function != pci_addr.function))
803 			continue;
804 		vf = (pci_dev->id.device_id ==
805 		      PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
806 		INFO("PCI information matches, using device \"%s\" (VF: %s)",
807 		     list[i]->name, (vf ? "true" : "false"));
808 		attr_ctx = mlx4_glue->open_device(list[i]);
809 		err = errno;
810 		break;
811 	}
812 	if (attr_ctx == NULL) {
813 		mlx4_glue->free_device_list(list);
814 		switch (err) {
815 		case 0:
816 			rte_errno = ENODEV;
817 			ERROR("cannot access device, is mlx4_ib loaded?");
818 			return -rte_errno;
819 		case EINVAL:
820 			rte_errno = EINVAL;
821 			ERROR("cannot use device, are drivers up to date?");
822 			return -rte_errno;
823 		}
824 		MLX4_ASSERT(err > 0);
825 		rte_errno = err;
826 		return -rte_errno;
827 	}
828 	ibv_dev = list[i];
829 	DEBUG("device opened");
830 	if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
831 		err = ENODEV;
832 		goto error;
833 	}
834 	INFO("%u port(s) detected", device_attr.phys_port_cnt);
835 	conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
836 	if (mlx4_args(pci_dev->device.devargs, &conf)) {
837 		ERROR("failed to process device arguments");
838 		err = EINVAL;
839 		goto error;
840 	}
841 	/* Use all ports when none are defined */
842 	if (!conf.ports.enabled)
843 		conf.ports.enabled = conf.ports.present;
844 	/* Retrieve extended device attributes. */
845 	if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
846 		err = ENODEV;
847 		goto error;
848 	}
849 	MLX4_ASSERT(device_attr.max_sge >= MLX4_MAX_SGE);
850 	for (i = 0; i < device_attr.phys_port_cnt; i++) {
851 		uint32_t port = i + 1; /* ports are indexed from one */
852 		struct ibv_context *ctx = NULL;
853 		struct ibv_port_attr port_attr;
854 		struct ibv_pd *pd = NULL;
855 		struct mlx4_priv *priv = NULL;
856 		struct rte_eth_dev *eth_dev = NULL;
857 		struct rte_ether_addr mac;
858 		char name[RTE_ETH_NAME_MAX_LEN];
859 
860 		/* If port is not enabled, skip. */
861 		if (!(conf.ports.enabled & (1 << i)))
862 			continue;
863 		DEBUG("using port %u", port);
864 		ctx = mlx4_glue->open_device(ibv_dev);
865 		if (ctx == NULL) {
866 			err = ENODEV;
867 			goto port_error;
868 		}
869 		snprintf(name, sizeof(name), "%s port %u",
870 			 mlx4_glue->get_device_name(ibv_dev), port);
871 		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
872 			eth_dev = rte_eth_dev_attach_secondary(name);
873 			if (eth_dev == NULL) {
874 				ERROR("can not attach rte ethdev");
875 				rte_errno = ENOMEM;
876 				err = rte_errno;
877 				goto error;
878 			}
879 			priv = eth_dev->data->dev_private;
880 			if (!priv->verbs_alloc_ctx.enabled) {
881 				ERROR("secondary process is not supported"
882 				      " due to lack of external allocator"
883 				      " from Verbs");
884 				rte_errno = ENOTSUP;
885 				err = rte_errno;
886 				goto error;
887 			}
888 			eth_dev->device = &pci_dev->device;
889 			eth_dev->dev_ops = &mlx4_dev_sec_ops;
890 			err = mlx4_proc_priv_init(eth_dev);
891 			if (err)
892 				goto error;
893 			/* Receive command fd from primary process. */
894 			err = mlx4_mp_req_verbs_cmd_fd(eth_dev);
895 			if (err < 0) {
896 				err = rte_errno;
897 				goto error;
898 			}
899 			/* Remap UAR for Tx queues. */
900 			err = mlx4_tx_uar_init_secondary(eth_dev, err);
901 			if (err) {
902 				err = rte_errno;
903 				goto error;
904 			}
905 			/*
906 			 * Ethdev pointer is still required as input since
907 			 * the primary device is not accessible from the
908 			 * secondary process.
909 			 */
910 			eth_dev->tx_pkt_burst = mlx4_tx_burst;
911 			eth_dev->rx_pkt_burst = mlx4_rx_burst;
912 			claim_zero(mlx4_glue->close_device(ctx));
913 			rte_eth_copy_pci_info(eth_dev, pci_dev);
914 			rte_eth_dev_probing_finish(eth_dev);
915 			continue;
916 		}
917 		/* Check port status. */
918 		err = mlx4_glue->query_port(ctx, port, &port_attr);
919 		if (err) {
920 			err = ENODEV;
921 			ERROR("port query failed: %s", strerror(err));
922 			goto port_error;
923 		}
924 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
925 			err = ENOTSUP;
926 			ERROR("port %d is not configured in Ethernet mode",
927 			      port);
928 			goto port_error;
929 		}
930 		if (port_attr.state != IBV_PORT_ACTIVE)
931 			DEBUG("port %d is not active: \"%s\" (%d)",
932 			      port, mlx4_glue->port_state_str(port_attr.state),
933 			      port_attr.state);
934 		/* Make asynchronous FD non-blocking to handle interrupts. */
935 		err = mlx4_fd_set_non_blocking(ctx->async_fd);
936 		if (err) {
937 			ERROR("cannot make asynchronous FD non-blocking: %s",
938 			      strerror(err));
939 			goto port_error;
940 		}
941 		/* Allocate protection domain. */
942 		pd = mlx4_glue->alloc_pd(ctx);
943 		if (pd == NULL) {
944 			err = ENOMEM;
945 			ERROR("PD allocation failure");
946 			goto port_error;
947 		}
948 		/* from rte_ethdev.c */
949 		priv = rte_zmalloc("ethdev private structure",
950 				   sizeof(*priv),
951 				   RTE_CACHE_LINE_SIZE);
952 		if (priv == NULL) {
953 			err = ENOMEM;
954 			ERROR("priv allocation failure");
955 			goto port_error;
956 		}
957 		priv->ctx = ctx;
958 		priv->device_attr = device_attr;
959 		priv->port = port;
960 		priv->pd = pd;
961 		priv->mtu = RTE_ETHER_MTU;
962 		priv->vf = vf;
963 		priv->hw_csum =	!!(device_attr.device_cap_flags &
964 				   IBV_DEVICE_RAW_IP_CSUM);
965 		DEBUG("checksum offloading is %ssupported",
966 		      (priv->hw_csum ? "" : "not "));
967 		/* Only ConnectX-3 Pro supports tunneling. */
968 		priv->hw_csum_l2tun =
969 			priv->hw_csum &&
970 			(device_attr.vendor_part_id ==
971 			 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
972 		DEBUG("L2 tunnel checksum offloads are %ssupported",
973 		      priv->hw_csum_l2tun ? "" : "not ");
974 		priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
975 						   &device_attr_ex);
976 		DEBUG("supported RSS hash fields mask: %016" PRIx64,
977 		      priv->hw_rss_sup);
978 		priv->hw_rss_max_qps =
979 			device_attr_ex.rss_caps.max_rwq_indirection_table_size;
980 		DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
981 		priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
982 					IBV_RAW_PACKET_CAP_SCATTER_FCS);
983 		DEBUG("FCS stripping toggling is %ssupported",
984 		      priv->hw_fcs_strip ? "" : "not ");
985 		priv->tso =
986 			((device_attr_ex.tso_caps.max_tso > 0) &&
987 			 (device_attr_ex.tso_caps.supported_qpts &
988 			  (1 << IBV_QPT_RAW_PACKET)));
989 		if (priv->tso)
990 			priv->tso_max_payload_sz =
991 					device_attr_ex.tso_caps.max_tso;
992 		DEBUG("TSO is %ssupported",
993 		      priv->tso ? "" : "not ");
994 		priv->mr_ext_memseg_en = conf.mr_ext_memseg_en;
995 		/* Configure the first MAC address by default. */
996 		err = mlx4_get_mac(priv, &mac.addr_bytes);
997 		if (err) {
998 			ERROR("cannot get MAC address, is mlx4_en loaded?"
999 			      " (error: %s)", strerror(err));
1000 			goto port_error;
1001 		}
1002 		INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1003 		     priv->port,
1004 		     mac.addr_bytes[0], mac.addr_bytes[1],
1005 		     mac.addr_bytes[2], mac.addr_bytes[3],
1006 		     mac.addr_bytes[4], mac.addr_bytes[5]);
1007 		/* Register MAC address. */
1008 		priv->mac[0] = mac;
1009 
1010 		if (mlx4_get_ifname(priv, &ifname) == 0) {
1011 			DEBUG("port %u ifname is \"%s\"",
1012 			      priv->port, ifname);
1013 			priv->if_index = if_nametoindex(ifname);
1014 		} else {
1015 			DEBUG("port %u ifname is unknown", priv->port);
1016 		}
1017 
1018 		/* Get actual MTU if possible. */
1019 		mlx4_mtu_get(priv, &priv->mtu);
1020 		DEBUG("port %u MTU is %u", priv->port, priv->mtu);
1021 		eth_dev = rte_eth_dev_allocate(name);
1022 		if (eth_dev == NULL) {
1023 			err = ENOMEM;
1024 			ERROR("can not allocate rte ethdev");
1025 			goto port_error;
1026 		}
1027 		eth_dev->data->dev_private = priv;
1028 		eth_dev->data->mac_addrs = priv->mac;
1029 		eth_dev->device = &pci_dev->device;
1030 		rte_eth_copy_pci_info(eth_dev, pci_dev);
1031 		/* Initialize local interrupt handle for current port. */
1032 		memset(&priv->intr_handle, 0, sizeof(struct rte_intr_handle));
1033 		priv->intr_handle.fd = -1;
1034 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1035 		/*
1036 		 * Override ethdev interrupt handle pointer with private
1037 		 * handle instead of that of the parent PCI device used by
1038 		 * default. This prevents it from being shared between all
1039 		 * ports of the same PCI device since each of them is
1040 		 * associated its own Verbs context.
1041 		 *
1042 		 * Rx interrupts in particular require this as the PMD has
1043 		 * no control over the registration of queue interrupts
1044 		 * besides setting up eth_dev->intr_handle, the rest is
1045 		 * handled by rte_intr_rx_ctl().
1046 		 */
1047 		eth_dev->intr_handle = &priv->intr_handle;
1048 		priv->dev_data = eth_dev->data;
1049 		eth_dev->dev_ops = &mlx4_dev_ops;
1050 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
1051 		/* Hint libmlx4 to use PMD allocator for data plane resources */
1052 		err = mlx4_glue->dv_set_context_attr
1053 			(ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
1054 			 (void *)((uintptr_t)&(struct mlx4dv_ctx_allocators){
1055 				 .alloc = &mlx4_alloc_verbs_buf,
1056 				 .free = &mlx4_free_verbs_buf,
1057 				 .data = priv,
1058 			}));
1059 		if (err)
1060 			WARN("Verbs external allocator is not supported");
1061 		else
1062 			priv->verbs_alloc_ctx.enabled = 1;
1063 #endif
1064 		/* Bring Ethernet device up. */
1065 		DEBUG("forcing Ethernet interface up");
1066 		mlx4_dev_set_link_up(eth_dev);
1067 		/* Update link status once if waiting for LSC. */
1068 		if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
1069 			mlx4_link_update(eth_dev, 0);
1070 		/*
1071 		 * Once the device is added to the list of memory event
1072 		 * callback, its global MR cache table cannot be expanded
1073 		 * on the fly because of deadlock. If it overflows, lookup
1074 		 * should be done by searching MR list linearly, which is slow.
1075 		 */
1076 		err = mlx4_mr_btree_init(&priv->mr.cache,
1077 					 MLX4_MR_BTREE_CACHE_N * 2,
1078 					 eth_dev->device->numa_node);
1079 		if (err) {
1080 			/* rte_errno is already set. */
1081 			goto port_error;
1082 		}
1083 		/* Add device to memory callback list. */
1084 		rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1085 		LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list,
1086 				 priv, mem_event_cb);
1087 		rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1088 		rte_eth_dev_probing_finish(eth_dev);
1089 		continue;
1090 port_error:
1091 		rte_free(priv);
1092 		if (eth_dev != NULL)
1093 			eth_dev->data->dev_private = NULL;
1094 		if (pd)
1095 			claim_zero(mlx4_glue->dealloc_pd(pd));
1096 		if (ctx)
1097 			claim_zero(mlx4_glue->close_device(ctx));
1098 		if (eth_dev != NULL) {
1099 			/* mac_addrs must not be freed because part of dev_private */
1100 			eth_dev->data->mac_addrs = NULL;
1101 			rte_eth_dev_release_port(eth_dev);
1102 		}
1103 		break;
1104 	}
1105 	/*
1106 	 * XXX if something went wrong in the loop above, there is a resource
1107 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1108 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
1109 	 * way to enumerate the registered ethdevs to free the previous ones.
1110 	 */
1111 error:
1112 	if (attr_ctx)
1113 		claim_zero(mlx4_glue->close_device(attr_ctx));
1114 	if (list)
1115 		mlx4_glue->free_device_list(list);
1116 	if (err)
1117 		rte_errno = err;
1118 	return -err;
1119 }
1120 
1121 static const struct rte_pci_id mlx4_pci_id_map[] = {
1122 	{
1123 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1124 			       PCI_DEVICE_ID_MELLANOX_CONNECTX3)
1125 	},
1126 	{
1127 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1128 			       PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
1129 	},
1130 	{
1131 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1132 			       PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
1133 	},
1134 	{
1135 		.vendor_id = 0
1136 	}
1137 };
1138 
1139 static struct rte_pci_driver mlx4_driver = {
1140 	.driver = {
1141 		.name = MLX4_DRIVER_NAME
1142 	},
1143 	.id_table = mlx4_pci_id_map,
1144 	.probe = mlx4_pci_probe,
1145 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1146 };
1147 
1148 #ifdef RTE_IBVERBS_LINK_DLOPEN
1149 
1150 /**
1151  * Suffix RTE_EAL_PMD_PATH with "-glue".
1152  *
1153  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1154  * suffixing its last component.
1155  *
1156  * @param buf[out]
1157  *   Output buffer, should be large enough otherwise NULL is returned.
1158  * @param size
1159  *   Size of @p out.
1160  *
1161  * @return
1162  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1163  */
1164 static char *
1165 mlx4_glue_path(char *buf, size_t size)
1166 {
1167 	static const char *const bad[] = { "/", ".", "..", NULL };
1168 	const char *path = RTE_EAL_PMD_PATH;
1169 	size_t len = strlen(path);
1170 	size_t off;
1171 	int i;
1172 
1173 	while (len && path[len - 1] == '/')
1174 		--len;
1175 	for (off = len; off && path[off - 1] != '/'; --off)
1176 		;
1177 	for (i = 0; bad[i]; ++i)
1178 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1179 			goto error;
1180 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1181 	if (i == -1 || (size_t)i >= size)
1182 		goto error;
1183 	return buf;
1184 error:
1185 	ERROR("unable to append \"-glue\" to last component of"
1186 	      " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1187 	      " please re-configure DPDK");
1188 	return NULL;
1189 }
1190 
1191 /**
1192  * Initialization routine for run-time dependency on rdma-core.
1193  */
1194 static int
1195 mlx4_glue_init(void)
1196 {
1197 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1198 	const char *path[] = {
1199 		/*
1200 		 * A basic security check is necessary before trusting
1201 		 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1202 		 */
1203 		(geteuid() == getuid() && getegid() == getgid() ?
1204 		 getenv("MLX4_GLUE_PATH") : NULL),
1205 		/*
1206 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1207 		 * variant, otherwise let dlopen() look up libraries on its
1208 		 * own.
1209 		 */
1210 		(*RTE_EAL_PMD_PATH ?
1211 		 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
1212 	};
1213 	unsigned int i = 0;
1214 	void *handle = NULL;
1215 	void **sym;
1216 	const char *dlmsg;
1217 
1218 	while (!handle && i != RTE_DIM(path)) {
1219 		const char *end;
1220 		size_t len;
1221 		int ret;
1222 
1223 		if (!path[i]) {
1224 			++i;
1225 			continue;
1226 		}
1227 		end = strpbrk(path[i], ":;");
1228 		if (!end)
1229 			end = path[i] + strlen(path[i]);
1230 		len = end - path[i];
1231 		ret = 0;
1232 		do {
1233 			char name[ret + 1];
1234 
1235 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
1236 				       (int)len, path[i],
1237 				       (!len || *(end - 1) == '/') ? "" : "/");
1238 			if (ret == -1)
1239 				break;
1240 			if (sizeof(name) != (size_t)ret + 1)
1241 				continue;
1242 			DEBUG("looking for rdma-core glue as \"%s\"", name);
1243 			handle = dlopen(name, RTLD_LAZY);
1244 			break;
1245 		} while (1);
1246 		path[i] = end + 1;
1247 		if (!*end)
1248 			++i;
1249 	}
1250 	if (!handle) {
1251 		rte_errno = EINVAL;
1252 		dlmsg = dlerror();
1253 		if (dlmsg)
1254 			WARN("cannot load glue library: %s", dlmsg);
1255 		goto glue_error;
1256 	}
1257 	sym = dlsym(handle, "mlx4_glue");
1258 	if (!sym || !*sym) {
1259 		rte_errno = EINVAL;
1260 		dlmsg = dlerror();
1261 		if (dlmsg)
1262 			ERROR("cannot resolve glue symbol: %s", dlmsg);
1263 		goto glue_error;
1264 	}
1265 	mlx4_glue = *sym;
1266 	return 0;
1267 glue_error:
1268 	if (handle)
1269 		dlclose(handle);
1270 	WARN("cannot initialize PMD due to missing run-time"
1271 	     " dependency on rdma-core libraries (libibverbs,"
1272 	     " libmlx4)");
1273 	return -rte_errno;
1274 }
1275 
1276 #endif
1277 
1278 /* Initialize driver log type. */
1279 RTE_LOG_REGISTER(mlx4_logtype, pmd.net.mlx4, NOTICE)
1280 
1281 /**
1282  * Driver initialization routine.
1283  */
1284 RTE_INIT(rte_mlx4_pmd_init)
1285 {
1286 	/*
1287 	 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
1288 	 * want to get success errno value in case of calling them
1289 	 * when the device was removed.
1290 	 */
1291 	setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
1292 	/*
1293 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1294 	 * huge pages. Calling ibv_fork_init() during init allows
1295 	 * applications to use fork() safely for purposes other than
1296 	 * using this PMD, which is not supported in forked processes.
1297 	 */
1298 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1299 #ifdef RTE_IBVERBS_LINK_DLOPEN
1300 	if (mlx4_glue_init())
1301 		return;
1302 	MLX4_ASSERT(mlx4_glue);
1303 #endif
1304 #ifdef RTE_LIBRTE_MLX4_DEBUG
1305 	/* Glue structure must not contain any NULL pointers. */
1306 	{
1307 		unsigned int i;
1308 
1309 		for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
1310 			MLX4_ASSERT(((const void *const *)mlx4_glue)[i]);
1311 	}
1312 #endif
1313 	if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
1314 		ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
1315 		      mlx4_glue->version, MLX4_GLUE_VERSION);
1316 		return;
1317 	}
1318 	mlx4_glue->fork_init();
1319 	rte_pci_register(&mlx4_driver);
1320 }
1321 
1322 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1323 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1324 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1325 	"* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");
1326