xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 358309f36776ba397601ba25710e7d23ee8f55ce)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox.
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 
17 /* Verbs header. */
18 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
19 #ifdef PEDANTIC
20 #pragma GCC diagnostic ignored "-Wpedantic"
21 #endif
22 #include <infiniband/verbs.h>
23 #ifdef PEDANTIC
24 #pragma GCC diagnostic error "-Wpedantic"
25 #endif
26 
27 #include <rte_malloc.h>
28 #include <rte_ethdev_driver.h>
29 #include <rte_ethdev_pci.h>
30 #include <rte_pci.h>
31 #include <rte_bus_pci.h>
32 #include <rte_common.h>
33 #include <rte_config.h>
34 #include <rte_eal_memconfig.h>
35 #include <rte_kvargs.h>
36 
37 #include "mlx5.h"
38 #include "mlx5_utils.h"
39 #include "mlx5_rxtx.h"
40 #include "mlx5_autoconf.h"
41 #include "mlx5_defs.h"
42 #include "mlx5_glue.h"
43 
44 /* Device parameter to enable RX completion queue compression. */
45 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
46 
47 /* Device parameter to configure inline send. */
48 #define MLX5_TXQ_INLINE "txq_inline"
49 
50 /*
51  * Device parameter to configure the number of TX queues threshold for
52  * enabling inline send.
53  */
54 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
55 
56 /* Device parameter to enable multi-packet send WQEs. */
57 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
58 
59 /* Device parameter to include 2 dsegs in the title WQEBB. */
60 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
61 
62 /* Device parameter to limit the size of inlining packet. */
63 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
64 
65 /* Device parameter to enable hardware Tx vector. */
66 #define MLX5_TX_VEC_EN "tx_vec_en"
67 
68 /* Device parameter to enable hardware Rx vector. */
69 #define MLX5_RX_VEC_EN "rx_vec_en"
70 
71 #ifndef HAVE_IBV_MLX5_MOD_MPW
72 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
73 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
74 #endif
75 
76 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
77 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
78 #endif
79 
80 /** Driver-specific log messages type. */
81 int mlx5_logtype;
82 
83 /**
84  * Retrieve integer value from environment variable.
85  *
86  * @param[in] name
87  *   Environment variable name.
88  *
89  * @return
90  *   Integer value, 0 if the variable is not set.
91  */
92 int
93 mlx5_getenv_int(const char *name)
94 {
95 	const char *val = getenv(name);
96 
97 	if (val == NULL)
98 		return 0;
99 	return atoi(val);
100 }
101 
102 /**
103  * Verbs callback to allocate a memory. This function should allocate the space
104  * according to the size provided residing inside a huge page.
105  * Please note that all allocation must respect the alignment from libmlx5
106  * (i.e. currently sysconf(_SC_PAGESIZE)).
107  *
108  * @param[in] size
109  *   The size in bytes of the memory to allocate.
110  * @param[in] data
111  *   A pointer to the callback data.
112  *
113  * @return
114  *   Allocated buffer, NULL otherwise and rte_errno is set.
115  */
116 static void *
117 mlx5_alloc_verbs_buf(size_t size, void *data)
118 {
119 	struct priv *priv = data;
120 	void *ret;
121 	size_t alignment = sysconf(_SC_PAGESIZE);
122 	unsigned int socket = SOCKET_ID_ANY;
123 
124 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
125 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
126 
127 		socket = ctrl->socket;
128 	} else if (priv->verbs_alloc_ctx.type ==
129 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
130 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
131 
132 		socket = ctrl->socket;
133 	}
134 	assert(data != NULL);
135 	ret = rte_malloc_socket(__func__, size, alignment, socket);
136 	if (!ret && size)
137 		rte_errno = ENOMEM;
138 	return ret;
139 }
140 
141 /**
142  * Verbs callback to free a memory.
143  *
144  * @param[in] ptr
145  *   A pointer to the memory to free.
146  * @param[in] data
147  *   A pointer to the callback data.
148  */
149 static void
150 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
151 {
152 	assert(data != NULL);
153 	rte_free(ptr);
154 }
155 
156 /**
157  * DPDK callback to close the device.
158  *
159  * Destroy all queues and objects, free memory.
160  *
161  * @param dev
162  *   Pointer to Ethernet device structure.
163  */
164 static void
165 mlx5_dev_close(struct rte_eth_dev *dev)
166 {
167 	struct priv *priv = dev->data->dev_private;
168 	unsigned int i;
169 	int ret;
170 
171 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
172 		dev->data->port_id,
173 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
174 	/* In case mlx5_dev_stop() has not been called. */
175 	mlx5_dev_interrupt_handler_uninstall(dev);
176 	mlx5_traffic_disable(dev);
177 	/* Prevent crashes when queues are still in use. */
178 	dev->rx_pkt_burst = removed_rx_burst;
179 	dev->tx_pkt_burst = removed_tx_burst;
180 	if (priv->rxqs != NULL) {
181 		/* XXX race condition if mlx5_rx_burst() is still running. */
182 		usleep(1000);
183 		for (i = 0; (i != priv->rxqs_n); ++i)
184 			mlx5_rxq_release(dev, i);
185 		priv->rxqs_n = 0;
186 		priv->rxqs = NULL;
187 	}
188 	if (priv->txqs != NULL) {
189 		/* XXX race condition if mlx5_tx_burst() is still running. */
190 		usleep(1000);
191 		for (i = 0; (i != priv->txqs_n); ++i)
192 			mlx5_txq_release(dev, i);
193 		priv->txqs_n = 0;
194 		priv->txqs = NULL;
195 	}
196 	if (priv->pd != NULL) {
197 		assert(priv->ctx != NULL);
198 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
199 		claim_zero(mlx5_glue->close_device(priv->ctx));
200 	} else
201 		assert(priv->ctx == NULL);
202 	if (priv->rss_conf.rss_key != NULL)
203 		rte_free(priv->rss_conf.rss_key);
204 	if (priv->reta_idx != NULL)
205 		rte_free(priv->reta_idx);
206 	if (priv->primary_socket)
207 		mlx5_socket_uninit(dev);
208 	ret = mlx5_hrxq_ibv_verify(dev);
209 	if (ret)
210 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
211 			dev->data->port_id);
212 	ret = mlx5_ind_table_ibv_verify(dev);
213 	if (ret)
214 		DRV_LOG(WARNING, "port %u some indirection table still remain",
215 			dev->data->port_id);
216 	ret = mlx5_rxq_ibv_verify(dev);
217 	if (ret)
218 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
219 			dev->data->port_id);
220 	ret = mlx5_rxq_verify(dev);
221 	if (ret)
222 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
223 			dev->data->port_id);
224 	ret = mlx5_txq_ibv_verify(dev);
225 	if (ret)
226 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
227 			dev->data->port_id);
228 	ret = mlx5_txq_verify(dev);
229 	if (ret)
230 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
231 			dev->data->port_id);
232 	ret = mlx5_flow_verify(dev);
233 	if (ret)
234 		DRV_LOG(WARNING, "port %u some flows still remain",
235 			dev->data->port_id);
236 	ret = mlx5_mr_verify(dev);
237 	if (ret)
238 		DRV_LOG(WARNING, "port %u some memory region still remain",
239 			dev->data->port_id);
240 	memset(priv, 0, sizeof(*priv));
241 }
242 
243 const struct eth_dev_ops mlx5_dev_ops = {
244 	.dev_configure = mlx5_dev_configure,
245 	.dev_start = mlx5_dev_start,
246 	.dev_stop = mlx5_dev_stop,
247 	.dev_set_link_down = mlx5_set_link_down,
248 	.dev_set_link_up = mlx5_set_link_up,
249 	.dev_close = mlx5_dev_close,
250 	.promiscuous_enable = mlx5_promiscuous_enable,
251 	.promiscuous_disable = mlx5_promiscuous_disable,
252 	.allmulticast_enable = mlx5_allmulticast_enable,
253 	.allmulticast_disable = mlx5_allmulticast_disable,
254 	.link_update = mlx5_link_update,
255 	.stats_get = mlx5_stats_get,
256 	.stats_reset = mlx5_stats_reset,
257 	.xstats_get = mlx5_xstats_get,
258 	.xstats_reset = mlx5_xstats_reset,
259 	.xstats_get_names = mlx5_xstats_get_names,
260 	.dev_infos_get = mlx5_dev_infos_get,
261 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
262 	.vlan_filter_set = mlx5_vlan_filter_set,
263 	.rx_queue_setup = mlx5_rx_queue_setup,
264 	.tx_queue_setup = mlx5_tx_queue_setup,
265 	.rx_queue_release = mlx5_rx_queue_release,
266 	.tx_queue_release = mlx5_tx_queue_release,
267 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
268 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
269 	.mac_addr_remove = mlx5_mac_addr_remove,
270 	.mac_addr_add = mlx5_mac_addr_add,
271 	.mac_addr_set = mlx5_mac_addr_set,
272 	.mtu_set = mlx5_dev_set_mtu,
273 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
274 	.vlan_offload_set = mlx5_vlan_offload_set,
275 	.reta_update = mlx5_dev_rss_reta_update,
276 	.reta_query = mlx5_dev_rss_reta_query,
277 	.rss_hash_update = mlx5_rss_hash_update,
278 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
279 	.filter_ctrl = mlx5_dev_filter_ctrl,
280 	.rx_descriptor_status = mlx5_rx_descriptor_status,
281 	.tx_descriptor_status = mlx5_tx_descriptor_status,
282 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
283 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
284 	.is_removed = mlx5_is_removed,
285 };
286 
287 static const struct eth_dev_ops mlx5_dev_sec_ops = {
288 	.stats_get = mlx5_stats_get,
289 	.stats_reset = mlx5_stats_reset,
290 	.xstats_get = mlx5_xstats_get,
291 	.xstats_reset = mlx5_xstats_reset,
292 	.xstats_get_names = mlx5_xstats_get_names,
293 	.dev_infos_get = mlx5_dev_infos_get,
294 	.rx_descriptor_status = mlx5_rx_descriptor_status,
295 	.tx_descriptor_status = mlx5_tx_descriptor_status,
296 };
297 
298 /* Available operators in flow isolated mode. */
299 const struct eth_dev_ops mlx5_dev_ops_isolate = {
300 	.dev_configure = mlx5_dev_configure,
301 	.dev_start = mlx5_dev_start,
302 	.dev_stop = mlx5_dev_stop,
303 	.dev_set_link_down = mlx5_set_link_down,
304 	.dev_set_link_up = mlx5_set_link_up,
305 	.dev_close = mlx5_dev_close,
306 	.link_update = mlx5_link_update,
307 	.stats_get = mlx5_stats_get,
308 	.stats_reset = mlx5_stats_reset,
309 	.xstats_get = mlx5_xstats_get,
310 	.xstats_reset = mlx5_xstats_reset,
311 	.xstats_get_names = mlx5_xstats_get_names,
312 	.dev_infos_get = mlx5_dev_infos_get,
313 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
314 	.vlan_filter_set = mlx5_vlan_filter_set,
315 	.rx_queue_setup = mlx5_rx_queue_setup,
316 	.tx_queue_setup = mlx5_tx_queue_setup,
317 	.rx_queue_release = mlx5_rx_queue_release,
318 	.tx_queue_release = mlx5_tx_queue_release,
319 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
320 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
321 	.mac_addr_remove = mlx5_mac_addr_remove,
322 	.mac_addr_add = mlx5_mac_addr_add,
323 	.mac_addr_set = mlx5_mac_addr_set,
324 	.mtu_set = mlx5_dev_set_mtu,
325 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
326 	.vlan_offload_set = mlx5_vlan_offload_set,
327 	.filter_ctrl = mlx5_dev_filter_ctrl,
328 	.rx_descriptor_status = mlx5_rx_descriptor_status,
329 	.tx_descriptor_status = mlx5_tx_descriptor_status,
330 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
331 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
332 	.is_removed = mlx5_is_removed,
333 };
334 
335 static struct {
336 	struct rte_pci_addr pci_addr; /* associated PCI address */
337 	uint32_t ports; /* physical ports bitfield. */
338 } mlx5_dev[32];
339 
340 /**
341  * Get device index in mlx5_dev[] from PCI bus address.
342  *
343  * @param[in] pci_addr
344  *   PCI bus address to look for.
345  *
346  * @return
347  *   mlx5_dev[] index on success, -1 on failure.
348  */
349 static int
350 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
351 {
352 	unsigned int i;
353 	int ret = -1;
354 
355 	assert(pci_addr != NULL);
356 	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
357 		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
358 		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
359 		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
360 		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
361 			return i;
362 		if ((mlx5_dev[i].ports == 0) && (ret == -1))
363 			ret = i;
364 	}
365 	return ret;
366 }
367 
368 /**
369  * Verify and store value for device argument.
370  *
371  * @param[in] key
372  *   Key argument to verify.
373  * @param[in] val
374  *   Value associated with key.
375  * @param opaque
376  *   User data.
377  *
378  * @return
379  *   0 on success, a negative errno value otherwise and rte_errno is set.
380  */
381 static int
382 mlx5_args_check(const char *key, const char *val, void *opaque)
383 {
384 	struct mlx5_dev_config *config = opaque;
385 	unsigned long tmp;
386 
387 	errno = 0;
388 	tmp = strtoul(val, NULL, 0);
389 	if (errno) {
390 		rte_errno = errno;
391 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
392 		return -rte_errno;
393 	}
394 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
395 		config->cqe_comp = !!tmp;
396 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
397 		config->txq_inline = tmp;
398 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
399 		config->txqs_inline = tmp;
400 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
401 		config->mps = !!tmp ? config->mps : 0;
402 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
403 		config->mpw_hdr_dseg = !!tmp;
404 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
405 		config->inline_max_packet_sz = tmp;
406 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
407 		config->tx_vec_en = !!tmp;
408 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
409 		config->rx_vec_en = !!tmp;
410 	} else {
411 		DRV_LOG(WARNING, "%s: unknown parameter", key);
412 		rte_errno = EINVAL;
413 		return -rte_errno;
414 	}
415 	return 0;
416 }
417 
418 /**
419  * Parse device parameters.
420  *
421  * @param config
422  *   Pointer to device configuration structure.
423  * @param devargs
424  *   Device arguments structure.
425  *
426  * @return
427  *   0 on success, a negative errno value otherwise and rte_errno is set.
428  */
429 static int
430 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
431 {
432 	const char **params = (const char *[]){
433 		MLX5_RXQ_CQE_COMP_EN,
434 		MLX5_TXQ_INLINE,
435 		MLX5_TXQS_MIN_INLINE,
436 		MLX5_TXQ_MPW_EN,
437 		MLX5_TXQ_MPW_HDR_DSEG_EN,
438 		MLX5_TXQ_MAX_INLINE_LEN,
439 		MLX5_TX_VEC_EN,
440 		MLX5_RX_VEC_EN,
441 		NULL,
442 	};
443 	struct rte_kvargs *kvlist;
444 	int ret = 0;
445 	int i;
446 
447 	if (devargs == NULL)
448 		return 0;
449 	/* Following UGLY cast is done to pass checkpatch. */
450 	kvlist = rte_kvargs_parse(devargs->args, params);
451 	if (kvlist == NULL)
452 		return 0;
453 	/* Process parameters. */
454 	for (i = 0; (params[i] != NULL); ++i) {
455 		if (rte_kvargs_count(kvlist, params[i])) {
456 			ret = rte_kvargs_process(kvlist, params[i],
457 						 mlx5_args_check, config);
458 			if (ret) {
459 				rte_errno = EINVAL;
460 				rte_kvargs_free(kvlist);
461 				return -rte_errno;
462 			}
463 		}
464 	}
465 	rte_kvargs_free(kvlist);
466 	return 0;
467 }
468 
469 static struct rte_pci_driver mlx5_driver;
470 
471 /*
472  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
473  * local resource used by both primary and secondary to avoid duplicate
474  * reservation.
475  * The space has to be available on both primary and secondary process,
476  * TXQ UAR maps to this area using fixed mmap w/o double check.
477  */
478 static void *uar_base;
479 
480 /**
481  * Reserve UAR address space for primary process.
482  *
483  * @param[in] dev
484  *   Pointer to Ethernet device.
485  *
486  * @return
487  *   0 on success, a negative errno value otherwise and rte_errno is set.
488  */
489 static int
490 mlx5_uar_init_primary(struct rte_eth_dev *dev)
491 {
492 	struct priv *priv = dev->data->dev_private;
493 	void *addr = (void *)0;
494 	int i;
495 	const struct rte_mem_config *mcfg;
496 
497 	if (uar_base) { /* UAR address space mapped. */
498 		priv->uar_base = uar_base;
499 		return 0;
500 	}
501 	/* find out lower bound of hugepage segments */
502 	mcfg = rte_eal_get_configuration()->mem_config;
503 	for (i = 0; i < RTE_MAX_MEMSEG && mcfg->memseg[i].addr; i++) {
504 		if (addr)
505 			addr = RTE_MIN(addr, mcfg->memseg[i].addr);
506 		else
507 			addr = mcfg->memseg[i].addr;
508 	}
509 	/* keep distance to hugepages to minimize potential conflicts. */
510 	addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
511 	/* anonymous mmap, no real memory consumption. */
512 	addr = mmap(addr, MLX5_UAR_SIZE,
513 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
514 	if (addr == MAP_FAILED) {
515 		DRV_LOG(ERR,
516 			"port %u failed to reserve UAR address space, please"
517 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
518 			dev->data->port_id);
519 		rte_errno = ENOMEM;
520 		return -rte_errno;
521 	}
522 	/* Accept either same addr or a new addr returned from mmap if target
523 	 * range occupied.
524 	 */
525 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
526 		dev->data->port_id, addr);
527 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
528 	uar_base = addr; /* process local, don't reserve again. */
529 	return 0;
530 }
531 
532 /**
533  * Reserve UAR address space for secondary process, align with
534  * primary process.
535  *
536  * @param[in] dev
537  *   Pointer to Ethernet device.
538  *
539  * @return
540  *   0 on success, a negative errno value otherwise and rte_errno is set.
541  */
542 static int
543 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
544 {
545 	struct priv *priv = dev->data->dev_private;
546 	void *addr;
547 
548 	assert(priv->uar_base);
549 	if (uar_base) { /* already reserved. */
550 		assert(uar_base == priv->uar_base);
551 		return 0;
552 	}
553 	/* anonymous mmap, no real memory consumption. */
554 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
555 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
556 	if (addr == MAP_FAILED) {
557 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
558 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
559 		rte_errno = ENXIO;
560 		return -rte_errno;
561 	}
562 	if (priv->uar_base != addr) {
563 		DRV_LOG(ERR,
564 			"port %u UAR address %p size %llu occupied, please"
565 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
566 			" --base-virtaddr",
567 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
568 		rte_errno = ENXIO;
569 		return -rte_errno;
570 	}
571 	uar_base = addr; /* process local, don't reserve again */
572 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
573 		dev->data->port_id, addr);
574 	return 0;
575 }
576 
577 /**
578  * DPDK callback to register a PCI device.
579  *
580  * This function creates an Ethernet device for each port of a given
581  * PCI device.
582  *
583  * @param[in] pci_drv
584  *   PCI driver structure (mlx5_driver).
585  * @param[in] pci_dev
586  *   PCI device information.
587  *
588  * @return
589  *   0 on success, a negative errno value otherwise and rte_errno is set.
590  */
591 static int
592 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
593 	       struct rte_pci_device *pci_dev)
594 {
595 	struct ibv_device **list = NULL;
596 	struct ibv_device *ibv_dev;
597 	int err = 0;
598 	struct ibv_context *attr_ctx = NULL;
599 	struct ibv_device_attr_ex device_attr;
600 	unsigned int mps;
601 	unsigned int cqe_comp;
602 	unsigned int tunnel_en = 0;
603 	int idx;
604 	int i;
605 	struct mlx5dv_context attrs_out = {0};
606 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
607 	struct ibv_counter_set_description cs_desc;
608 #endif
609 
610 	assert(pci_drv == &mlx5_driver);
611 	/* Get mlx5_dev[] index. */
612 	idx = mlx5_dev_idx(&pci_dev->addr);
613 	if (idx == -1) {
614 		DRV_LOG(ERR, "this driver cannot support any more adapters");
615 		err = ENOMEM;
616 		goto error;
617 	}
618 	DRV_LOG(DEBUG, "using driver device index %d", idx);
619 	/* Save PCI address. */
620 	mlx5_dev[idx].pci_addr = pci_dev->addr;
621 	list = mlx5_glue->get_device_list(&i);
622 	if (list == NULL) {
623 		assert(errno);
624 		err = errno;
625 		if (errno == ENOSYS)
626 			DRV_LOG(ERR,
627 				"cannot list devices, is ib_uverbs loaded?");
628 		goto error;
629 	}
630 	assert(i >= 0);
631 	/*
632 	 * For each listed device, check related sysfs entry against
633 	 * the provided PCI ID.
634 	 */
635 	while (i != 0) {
636 		struct rte_pci_addr pci_addr;
637 
638 		--i;
639 		DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
640 		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
641 			continue;
642 		if ((pci_dev->addr.domain != pci_addr.domain) ||
643 		    (pci_dev->addr.bus != pci_addr.bus) ||
644 		    (pci_dev->addr.devid != pci_addr.devid) ||
645 		    (pci_dev->addr.function != pci_addr.function))
646 			continue;
647 		DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
648 			list[i]->name);
649 		attr_ctx = mlx5_glue->open_device(list[i]);
650 		rte_errno = errno;
651 		err = rte_errno;
652 		break;
653 	}
654 	if (attr_ctx == NULL) {
655 		mlx5_glue->free_device_list(list);
656 		switch (err) {
657 		case 0:
658 			DRV_LOG(ERR,
659 				"cannot access device, is mlx5_ib loaded?");
660 			err = ENODEV;
661 			goto error;
662 		case EINVAL:
663 			DRV_LOG(ERR,
664 				"cannot use device, are drivers up to date?");
665 			goto error;
666 		}
667 	}
668 	ibv_dev = list[i];
669 	DRV_LOG(DEBUG, "device opened");
670 	/*
671 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
672 	 * as all ConnectX-5 devices.
673 	 */
674 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
675 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
676 #endif
677 	mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
678 	if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
679 		if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
680 			DRV_LOG(DEBUG, "enhanced MPW is supported");
681 			mps = MLX5_MPW_ENHANCED;
682 		} else {
683 			DRV_LOG(DEBUG, "MPW is supported");
684 			mps = MLX5_MPW;
685 		}
686 	} else {
687 		DRV_LOG(DEBUG, "MPW isn't supported");
688 		mps = MLX5_MPW_DISABLED;
689 	}
690 	if (RTE_CACHE_LINE_SIZE == 128 &&
691 	    !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
692 		cqe_comp = 0;
693 	else
694 		cqe_comp = 1;
695 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
696 	if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
697 		tunnel_en = ((attrs_out.tunnel_offloads_caps &
698 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
699 			     (attrs_out.tunnel_offloads_caps &
700 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
701 	}
702 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
703 		tunnel_en ? "" : "not ");
704 #else
705 	DRV_LOG(WARNING,
706 		"tunnel offloading disabled due to old OFED/rdma-core version");
707 #endif
708 	if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr)) {
709 		err = errno;
710 		goto error;
711 	}
712 	DRV_LOG(INFO, "%u port(s) detected",
713 		device_attr.orig_attr.phys_port_cnt);
714 	for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
715 		char name[RTE_ETH_NAME_MAX_LEN];
716 		int len;
717 		uint32_t port = i + 1; /* ports are indexed from one */
718 		uint32_t test = (1 << i);
719 		struct ibv_context *ctx = NULL;
720 		struct ibv_port_attr port_attr;
721 		struct ibv_pd *pd = NULL;
722 		struct priv *priv = NULL;
723 		struct rte_eth_dev *eth_dev = NULL;
724 		struct ibv_device_attr_ex device_attr_ex;
725 		struct ether_addr mac;
726 		struct mlx5_dev_config config = {
727 			.cqe_comp = cqe_comp,
728 			.mps = mps,
729 			.tunnel_en = tunnel_en,
730 			.tx_vec_en = 1,
731 			.rx_vec_en = 1,
732 			.mpw_hdr_dseg = 0,
733 			.txq_inline = MLX5_ARG_UNSET,
734 			.txqs_inline = MLX5_ARG_UNSET,
735 			.inline_max_packet_sz = MLX5_ARG_UNSET,
736 		};
737 
738 		len = snprintf(name, sizeof(name), PCI_PRI_FMT,
739 			 pci_dev->addr.domain, pci_dev->addr.bus,
740 			 pci_dev->addr.devid, pci_dev->addr.function);
741 		if (device_attr.orig_attr.phys_port_cnt > 1)
742 			snprintf(name + len, sizeof(name), " port %u", i);
743 		mlx5_dev[idx].ports |= test;
744 		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
745 			eth_dev = rte_eth_dev_attach_secondary(name);
746 			if (eth_dev == NULL) {
747 				DRV_LOG(ERR, "can not attach rte ethdev");
748 				rte_errno = ENOMEM;
749 				err = rte_errno;
750 				goto error;
751 			}
752 			eth_dev->device = &pci_dev->device;
753 			eth_dev->dev_ops = &mlx5_dev_sec_ops;
754 			err = mlx5_uar_init_secondary(eth_dev);
755 			if (err)
756 				goto error;
757 			/* Receive command fd from primary process */
758 			err = mlx5_socket_connect(eth_dev);
759 			if (err)
760 				goto error;
761 			/* Remap UAR for Tx queues. */
762 			err = mlx5_tx_uar_remap(eth_dev, err);
763 			if (err)
764 				goto error;
765 			/*
766 			 * Ethdev pointer is still required as input since
767 			 * the primary device is not accessible from the
768 			 * secondary process.
769 			 */
770 			eth_dev->rx_pkt_burst =
771 				mlx5_select_rx_function(eth_dev);
772 			eth_dev->tx_pkt_burst =
773 				mlx5_select_tx_function(eth_dev);
774 			continue;
775 		}
776 		DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
777 		ctx = mlx5_glue->open_device(ibv_dev);
778 		if (ctx == NULL) {
779 			err = ENODEV;
780 			goto port_error;
781 		}
782 		/* Check port status. */
783 		err = mlx5_glue->query_port(ctx, port, &port_attr);
784 		if (err) {
785 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
786 			goto port_error;
787 		}
788 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
789 			DRV_LOG(ERR,
790 				"port %d is not configured in Ethernet mode",
791 				port);
792 			err = EINVAL;
793 			goto port_error;
794 		}
795 		if (port_attr.state != IBV_PORT_ACTIVE)
796 			DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
797 				port,
798 				mlx5_glue->port_state_str(port_attr.state),
799 				port_attr.state);
800 		/* Allocate protection domain. */
801 		pd = mlx5_glue->alloc_pd(ctx);
802 		if (pd == NULL) {
803 			DRV_LOG(ERR, "PD allocation failure");
804 			err = ENOMEM;
805 			goto port_error;
806 		}
807 		mlx5_dev[idx].ports |= test;
808 		/* from rte_ethdev.c */
809 		priv = rte_zmalloc("ethdev private structure",
810 				   sizeof(*priv),
811 				   RTE_CACHE_LINE_SIZE);
812 		if (priv == NULL) {
813 			DRV_LOG(ERR, "priv allocation failure");
814 			err = ENOMEM;
815 			goto port_error;
816 		}
817 		priv->ctx = ctx;
818 		strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
819 			sizeof(priv->ibdev_path));
820 		priv->device_attr = device_attr;
821 		priv->port = port;
822 		priv->pd = pd;
823 		priv->mtu = ETHER_MTU;
824 		err = mlx5_args(&config, pci_dev->device.devargs);
825 		if (err) {
826 			DRV_LOG(ERR, "failed to process device arguments: %s",
827 				strerror(err));
828 			goto port_error;
829 		}
830 		if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
831 			DRV_LOG(ERR, "ibv_query_device_ex() failed");
832 			err = errno;
833 			goto port_error;
834 		}
835 		config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
836 				    IBV_DEVICE_RAW_IP_CSUM);
837 		DRV_LOG(DEBUG, "checksum offloading is %ssupported",
838 			(config.hw_csum ? "" : "not "));
839 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
840 		config.flow_counter_en = !!(device_attr.max_counter_sets);
841 		mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
842 		DRV_LOG(DEBUG,
843 			"counter type = %d, num of cs = %ld, attributes = %d",
844 			cs_desc.counter_type, cs_desc.num_of_cs,
845 			cs_desc.attributes);
846 #endif
847 		config.ind_table_max_size =
848 			device_attr_ex.rss_caps.max_rwq_indirection_table_size;
849 		/* Remove this check once DPDK supports larger/variable
850 		 * indirection tables. */
851 		if (config.ind_table_max_size >
852 				(unsigned int)ETH_RSS_RETA_SIZE_512)
853 			config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
854 		DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
855 			config.ind_table_max_size);
856 		config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
857 					 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
858 		DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
859 			(config.hw_vlan_strip ? "" : "not "));
860 
861 		config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
862 					 IBV_RAW_PACKET_CAP_SCATTER_FCS);
863 		DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
864 			(config.hw_fcs_strip ? "" : "not "));
865 
866 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
867 		config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
868 #endif
869 		DRV_LOG(DEBUG,
870 			"hardware Rx end alignment padding is %ssupported",
871 			(config.hw_padding ? "" : "not "));
872 		config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
873 			      (device_attr_ex.tso_caps.supported_qpts &
874 			      (1 << IBV_QPT_RAW_PACKET)));
875 		if (config.tso)
876 			config.tso_max_payload_sz =
877 					device_attr_ex.tso_caps.max_tso;
878 		if (config.mps && !mps) {
879 			DRV_LOG(ERR,
880 				"multi-packet send not supported on this device"
881 				" (" MLX5_TXQ_MPW_EN ")");
882 			err = ENOTSUP;
883 			goto port_error;
884 		}
885 		DRV_LOG(INFO, "%s MPS is %s",
886 			config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
887 			config.mps != MLX5_MPW_DISABLED ? "enabled" :
888 			"disabled");
889 		if (config.cqe_comp && !cqe_comp) {
890 			DRV_LOG(WARNING, "Rx CQE compression isn't supported");
891 			config.cqe_comp = 0;
892 		}
893 		eth_dev = rte_eth_dev_allocate(name);
894 		if (eth_dev == NULL) {
895 			DRV_LOG(ERR, "can not allocate rte ethdev");
896 			err = ENOMEM;
897 			goto port_error;
898 		}
899 		eth_dev->data->dev_private = priv;
900 		priv->dev = eth_dev;
901 		eth_dev->data->mac_addrs = priv->mac;
902 		eth_dev->device = &pci_dev->device;
903 		rte_eth_copy_pci_info(eth_dev, pci_dev);
904 		eth_dev->device->driver = &mlx5_driver.driver;
905 		err = mlx5_uar_init_primary(eth_dev);
906 		if (err)
907 			goto port_error;
908 		/* Configure the first MAC address by default. */
909 		if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
910 			DRV_LOG(ERR,
911 				"port %u cannot get MAC address, is mlx5_en"
912 				" loaded? (errno: %s)",
913 				eth_dev->data->port_id, strerror(errno));
914 			err = ENODEV;
915 			goto port_error;
916 		}
917 		DRV_LOG(INFO,
918 			"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
919 			eth_dev->data->port_id,
920 			mac.addr_bytes[0], mac.addr_bytes[1],
921 			mac.addr_bytes[2], mac.addr_bytes[3],
922 			mac.addr_bytes[4], mac.addr_bytes[5]);
923 #ifndef NDEBUG
924 		{
925 			char ifname[IF_NAMESIZE];
926 
927 			if (mlx5_get_ifname(eth_dev, &ifname) == 0)
928 				DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
929 					eth_dev->data->port_id, ifname);
930 			else
931 				DRV_LOG(DEBUG, "port %u ifname is unknown",
932 					eth_dev->data->port_id);
933 		}
934 #endif
935 		/* Get actual MTU if possible. */
936 		err = mlx5_get_mtu(eth_dev, &priv->mtu);
937 		if (err)
938 			goto port_error;
939 		DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
940 			priv->mtu);
941 		/*
942 		 * Initialize burst functions to prevent crashes before link-up.
943 		 */
944 		eth_dev->rx_pkt_burst = removed_rx_burst;
945 		eth_dev->tx_pkt_burst = removed_tx_burst;
946 		eth_dev->dev_ops = &mlx5_dev_ops;
947 		/* Register MAC address. */
948 		claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
949 		TAILQ_INIT(&priv->flows);
950 		TAILQ_INIT(&priv->ctrl_flows);
951 		/* Hint libmlx5 to use PMD allocator for data plane resources */
952 		struct mlx5dv_ctx_allocators alctr = {
953 			.alloc = &mlx5_alloc_verbs_buf,
954 			.free = &mlx5_free_verbs_buf,
955 			.data = priv,
956 		};
957 		mlx5_glue->dv_set_context_attr(ctx,
958 					       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
959 					       (void *)((uintptr_t)&alctr));
960 		/* Bring Ethernet device up. */
961 		DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
962 			eth_dev->data->port_id);
963 		mlx5_set_link_up(eth_dev);
964 		/* Store device configuration on private structure. */
965 		priv->config = config;
966 		continue;
967 port_error:
968 		if (priv)
969 			rte_free(priv);
970 		if (pd)
971 			claim_zero(mlx5_glue->dealloc_pd(pd));
972 		if (ctx)
973 			claim_zero(mlx5_glue->close_device(ctx));
974 		break;
975 	}
976 	/*
977 	 * XXX if something went wrong in the loop above, there is a resource
978 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
979 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
980 	 * way to enumerate the registered ethdevs to free the previous ones.
981 	 */
982 	/* no port found, complain */
983 	if (!mlx5_dev[idx].ports) {
984 		rte_errno = ENODEV;
985 		err = rte_errno;
986 	}
987 error:
988 	if (attr_ctx)
989 		claim_zero(mlx5_glue->close_device(attr_ctx));
990 	if (list)
991 		mlx5_glue->free_device_list(list);
992 	if (err) {
993 		rte_errno = err;
994 		return -rte_errno;
995 	}
996 	return 0;
997 }
998 
999 static const struct rte_pci_id mlx5_pci_id_map[] = {
1000 	{
1001 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1002 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1003 	},
1004 	{
1005 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1006 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1007 	},
1008 	{
1009 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1010 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1011 	},
1012 	{
1013 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1014 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1015 	},
1016 	{
1017 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1018 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1019 	},
1020 	{
1021 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1022 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1023 	},
1024 	{
1025 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1026 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1027 	},
1028 	{
1029 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1030 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1031 	},
1032 	{
1033 		.vendor_id = 0
1034 	}
1035 };
1036 
1037 static struct rte_pci_driver mlx5_driver = {
1038 	.driver = {
1039 		.name = MLX5_DRIVER_NAME
1040 	},
1041 	.id_table = mlx5_pci_id_map,
1042 	.probe = mlx5_pci_probe,
1043 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1044 };
1045 
1046 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1047 
1048 /**
1049  * Suffix RTE_EAL_PMD_PATH with "-glue".
1050  *
1051  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1052  * suffixing its last component.
1053  *
1054  * @param buf[out]
1055  *   Output buffer, should be large enough otherwise NULL is returned.
1056  * @param size
1057  *   Size of @p out.
1058  *
1059  * @return
1060  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1061  */
1062 static char *
1063 mlx5_glue_path(char *buf, size_t size)
1064 {
1065 	static const char *const bad[] = { "/", ".", "..", NULL };
1066 	const char *path = RTE_EAL_PMD_PATH;
1067 	size_t len = strlen(path);
1068 	size_t off;
1069 	int i;
1070 
1071 	while (len && path[len - 1] == '/')
1072 		--len;
1073 	for (off = len; off && path[off - 1] != '/'; --off)
1074 		;
1075 	for (i = 0; bad[i]; ++i)
1076 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1077 			goto error;
1078 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1079 	if (i == -1 || (size_t)i >= size)
1080 		goto error;
1081 	return buf;
1082 error:
1083 	DRV_LOG(ERR,
1084 		"unable to append \"-glue\" to last component of"
1085 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1086 		" please re-configure DPDK");
1087 	return NULL;
1088 }
1089 
1090 /**
1091  * Initialization routine for run-time dependency on rdma-core.
1092  */
1093 static int
1094 mlx5_glue_init(void)
1095 {
1096 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1097 	const char *path[] = {
1098 		/*
1099 		 * A basic security check is necessary before trusting
1100 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1101 		 */
1102 		(geteuid() == getuid() && getegid() == getgid() ?
1103 		 getenv("MLX5_GLUE_PATH") : NULL),
1104 		/*
1105 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1106 		 * variant, otherwise let dlopen() look up libraries on its
1107 		 * own.
1108 		 */
1109 		(*RTE_EAL_PMD_PATH ?
1110 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1111 	};
1112 	unsigned int i = 0;
1113 	void *handle = NULL;
1114 	void **sym;
1115 	const char *dlmsg;
1116 
1117 	while (!handle && i != RTE_DIM(path)) {
1118 		const char *end;
1119 		size_t len;
1120 		int ret;
1121 
1122 		if (!path[i]) {
1123 			++i;
1124 			continue;
1125 		}
1126 		end = strpbrk(path[i], ":;");
1127 		if (!end)
1128 			end = path[i] + strlen(path[i]);
1129 		len = end - path[i];
1130 		ret = 0;
1131 		do {
1132 			char name[ret + 1];
1133 
1134 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1135 				       (int)len, path[i],
1136 				       (!len || *(end - 1) == '/') ? "" : "/");
1137 			if (ret == -1)
1138 				break;
1139 			if (sizeof(name) != (size_t)ret + 1)
1140 				continue;
1141 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1142 				name);
1143 			handle = dlopen(name, RTLD_LAZY);
1144 			break;
1145 		} while (1);
1146 		path[i] = end + 1;
1147 		if (!*end)
1148 			++i;
1149 	}
1150 	if (!handle) {
1151 		rte_errno = EINVAL;
1152 		dlmsg = dlerror();
1153 		if (dlmsg)
1154 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1155 		goto glue_error;
1156 	}
1157 	sym = dlsym(handle, "mlx5_glue");
1158 	if (!sym || !*sym) {
1159 		rte_errno = EINVAL;
1160 		dlmsg = dlerror();
1161 		if (dlmsg)
1162 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1163 		goto glue_error;
1164 	}
1165 	mlx5_glue = *sym;
1166 	return 0;
1167 glue_error:
1168 	if (handle)
1169 		dlclose(handle);
1170 	DRV_LOG(WARNING,
1171 		"cannot initialize PMD due to missing run-time dependency on"
1172 		" rdma-core libraries (libibverbs, libmlx5)");
1173 	return -rte_errno;
1174 }
1175 
1176 #endif
1177 
1178 /**
1179  * Driver initialization routine.
1180  */
1181 RTE_INIT(rte_mlx5_pmd_init);
1182 static void
1183 rte_mlx5_pmd_init(void)
1184 {
1185 	/* Build the static table for ptype conversion. */
1186 	mlx5_set_ptype_table();
1187 	/*
1188 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1189 	 * huge pages. Calling ibv_fork_init() during init allows
1190 	 * applications to use fork() safely for purposes other than
1191 	 * using this PMD, which is not supported in forked processes.
1192 	 */
1193 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1194 	/* Match the size of Rx completion entry to the size of a cacheline. */
1195 	if (RTE_CACHE_LINE_SIZE == 128)
1196 		setenv("MLX5_CQE_SIZE", "128", 0);
1197 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1198 	if (mlx5_glue_init())
1199 		return;
1200 	assert(mlx5_glue);
1201 #endif
1202 #ifndef NDEBUG
1203 	/* Glue structure must not contain any NULL pointers. */
1204 	{
1205 		unsigned int i;
1206 
1207 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1208 			assert(((const void *const *)mlx5_glue)[i]);
1209 	}
1210 #endif
1211 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1212 		DRV_LOG(ERR,
1213 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1214 			mlx5_glue->version, MLX5_GLUE_VERSION);
1215 		return;
1216 	}
1217 	mlx5_glue->fork_init();
1218 	rte_pci_register(&mlx5_driver);
1219 }
1220 
1221 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1222 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1223 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1224 
1225 /** Initialize driver log type. */
1226 RTE_INIT(vdev_netvsc_init_log)
1227 {
1228 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1229 	if (mlx5_logtype >= 0)
1230 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1231 }
1232