xref: /dpdk/drivers/net/mlx5/mlx5.c (revision fd5baf09cdf9170e0f92a112fd0ef19c29649330)
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 
38 #include "mlx5.h"
39 #include "mlx5_utils.h"
40 #include "mlx5_rxtx.h"
41 #include "mlx5_autoconf.h"
42 #include "mlx5_defs.h"
43 #include "mlx5_glue.h"
44 
45 /* Device parameter to enable RX completion queue compression. */
46 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
47 
48 /* Device parameter to configure inline send. */
49 #define MLX5_TXQ_INLINE "txq_inline"
50 
51 /*
52  * Device parameter to configure the number of TX queues threshold for
53  * enabling inline send.
54  */
55 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
56 
57 /* Device parameter to enable multi-packet send WQEs. */
58 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
59 
60 /* Device parameter to include 2 dsegs in the title WQEBB. */
61 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
62 
63 /* Device parameter to limit the size of inlining packet. */
64 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
65 
66 /* Device parameter to enable hardware Tx vector. */
67 #define MLX5_TX_VEC_EN "tx_vec_en"
68 
69 /* Device parameter to enable hardware Rx vector. */
70 #define MLX5_RX_VEC_EN "rx_vec_en"
71 
72 /* Activate Netlink support in VF mode. */
73 #define MLX5_VF_NL_EN "vf_nl_en"
74 
75 #ifndef HAVE_IBV_MLX5_MOD_MPW
76 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
77 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
78 #endif
79 
80 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
81 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
82 #endif
83 
84 /** Driver-specific log messages type. */
85 int mlx5_logtype;
86 
87 /**
88  * Retrieve integer value from environment variable.
89  *
90  * @param[in] name
91  *   Environment variable name.
92  *
93  * @return
94  *   Integer value, 0 if the variable is not set.
95  */
96 int
97 mlx5_getenv_int(const char *name)
98 {
99 	const char *val = getenv(name);
100 
101 	if (val == NULL)
102 		return 0;
103 	return atoi(val);
104 }
105 
106 /**
107  * Verbs callback to allocate a memory. This function should allocate the space
108  * according to the size provided residing inside a huge page.
109  * Please note that all allocation must respect the alignment from libmlx5
110  * (i.e. currently sysconf(_SC_PAGESIZE)).
111  *
112  * @param[in] size
113  *   The size in bytes of the memory to allocate.
114  * @param[in] data
115  *   A pointer to the callback data.
116  *
117  * @return
118  *   Allocated buffer, NULL otherwise and rte_errno is set.
119  */
120 static void *
121 mlx5_alloc_verbs_buf(size_t size, void *data)
122 {
123 	struct priv *priv = data;
124 	void *ret;
125 	size_t alignment = sysconf(_SC_PAGESIZE);
126 	unsigned int socket = SOCKET_ID_ANY;
127 
128 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
129 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
130 
131 		socket = ctrl->socket;
132 	} else if (priv->verbs_alloc_ctx.type ==
133 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
134 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
135 
136 		socket = ctrl->socket;
137 	}
138 	assert(data != NULL);
139 	ret = rte_malloc_socket(__func__, size, alignment, socket);
140 	if (!ret && size)
141 		rte_errno = ENOMEM;
142 	return ret;
143 }
144 
145 /**
146  * Verbs callback to free a memory.
147  *
148  * @param[in] ptr
149  *   A pointer to the memory to free.
150  * @param[in] data
151  *   A pointer to the callback data.
152  */
153 static void
154 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
155 {
156 	assert(data != NULL);
157 	rte_free(ptr);
158 }
159 
160 /**
161  * DPDK callback to close the device.
162  *
163  * Destroy all queues and objects, free memory.
164  *
165  * @param dev
166  *   Pointer to Ethernet device structure.
167  */
168 static void
169 mlx5_dev_close(struct rte_eth_dev *dev)
170 {
171 	struct priv *priv = dev->data->dev_private;
172 	unsigned int i;
173 	int ret;
174 
175 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
176 		dev->data->port_id,
177 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
178 	/* In case mlx5_dev_stop() has not been called. */
179 	mlx5_dev_interrupt_handler_uninstall(dev);
180 	mlx5_traffic_disable(dev);
181 	/* Prevent crashes when queues are still in use. */
182 	dev->rx_pkt_burst = removed_rx_burst;
183 	dev->tx_pkt_burst = removed_tx_burst;
184 	if (priv->rxqs != NULL) {
185 		/* XXX race condition if mlx5_rx_burst() is still running. */
186 		usleep(1000);
187 		for (i = 0; (i != priv->rxqs_n); ++i)
188 			mlx5_rxq_release(dev, i);
189 		priv->rxqs_n = 0;
190 		priv->rxqs = NULL;
191 	}
192 	if (priv->txqs != NULL) {
193 		/* XXX race condition if mlx5_tx_burst() is still running. */
194 		usleep(1000);
195 		for (i = 0; (i != priv->txqs_n); ++i)
196 			mlx5_txq_release(dev, i);
197 		priv->txqs_n = 0;
198 		priv->txqs = NULL;
199 	}
200 	if (priv->pd != NULL) {
201 		assert(priv->ctx != NULL);
202 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
203 		claim_zero(mlx5_glue->close_device(priv->ctx));
204 	} else
205 		assert(priv->ctx == NULL);
206 	if (priv->rss_conf.rss_key != NULL)
207 		rte_free(priv->rss_conf.rss_key);
208 	if (priv->reta_idx != NULL)
209 		rte_free(priv->reta_idx);
210 	if (priv->primary_socket)
211 		mlx5_socket_uninit(dev);
212 	if (priv->config.vf)
213 		mlx5_nl_mac_addr_flush(dev);
214 	if (priv->nl_socket >= 0)
215 		close(priv->nl_socket);
216 	ret = mlx5_hrxq_ibv_verify(dev);
217 	if (ret)
218 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
219 			dev->data->port_id);
220 	ret = mlx5_ind_table_ibv_verify(dev);
221 	if (ret)
222 		DRV_LOG(WARNING, "port %u some indirection table still remain",
223 			dev->data->port_id);
224 	ret = mlx5_rxq_ibv_verify(dev);
225 	if (ret)
226 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
227 			dev->data->port_id);
228 	ret = mlx5_rxq_verify(dev);
229 	if (ret)
230 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
231 			dev->data->port_id);
232 	ret = mlx5_txq_ibv_verify(dev);
233 	if (ret)
234 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
235 			dev->data->port_id);
236 	ret = mlx5_txq_verify(dev);
237 	if (ret)
238 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
239 			dev->data->port_id);
240 	ret = mlx5_flow_verify(dev);
241 	if (ret)
242 		DRV_LOG(WARNING, "port %u some flows still remain",
243 			dev->data->port_id);
244 	ret = mlx5_mr_verify(dev);
245 	if (ret)
246 		DRV_LOG(WARNING, "port %u some memory region still remain",
247 			dev->data->port_id);
248 	memset(priv, 0, sizeof(*priv));
249 }
250 
251 const struct eth_dev_ops mlx5_dev_ops = {
252 	.dev_configure = mlx5_dev_configure,
253 	.dev_start = mlx5_dev_start,
254 	.dev_stop = mlx5_dev_stop,
255 	.dev_set_link_down = mlx5_set_link_down,
256 	.dev_set_link_up = mlx5_set_link_up,
257 	.dev_close = mlx5_dev_close,
258 	.promiscuous_enable = mlx5_promiscuous_enable,
259 	.promiscuous_disable = mlx5_promiscuous_disable,
260 	.allmulticast_enable = mlx5_allmulticast_enable,
261 	.allmulticast_disable = mlx5_allmulticast_disable,
262 	.link_update = mlx5_link_update,
263 	.stats_get = mlx5_stats_get,
264 	.stats_reset = mlx5_stats_reset,
265 	.xstats_get = mlx5_xstats_get,
266 	.xstats_reset = mlx5_xstats_reset,
267 	.xstats_get_names = mlx5_xstats_get_names,
268 	.dev_infos_get = mlx5_dev_infos_get,
269 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
270 	.vlan_filter_set = mlx5_vlan_filter_set,
271 	.rx_queue_setup = mlx5_rx_queue_setup,
272 	.tx_queue_setup = mlx5_tx_queue_setup,
273 	.rx_queue_release = mlx5_rx_queue_release,
274 	.tx_queue_release = mlx5_tx_queue_release,
275 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
276 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
277 	.mac_addr_remove = mlx5_mac_addr_remove,
278 	.mac_addr_add = mlx5_mac_addr_add,
279 	.mac_addr_set = mlx5_mac_addr_set,
280 	.mtu_set = mlx5_dev_set_mtu,
281 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
282 	.vlan_offload_set = mlx5_vlan_offload_set,
283 	.reta_update = mlx5_dev_rss_reta_update,
284 	.reta_query = mlx5_dev_rss_reta_query,
285 	.rss_hash_update = mlx5_rss_hash_update,
286 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
287 	.filter_ctrl = mlx5_dev_filter_ctrl,
288 	.rx_descriptor_status = mlx5_rx_descriptor_status,
289 	.tx_descriptor_status = mlx5_tx_descriptor_status,
290 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
291 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
292 	.is_removed = mlx5_is_removed,
293 };
294 
295 static const struct eth_dev_ops mlx5_dev_sec_ops = {
296 	.stats_get = mlx5_stats_get,
297 	.stats_reset = mlx5_stats_reset,
298 	.xstats_get = mlx5_xstats_get,
299 	.xstats_reset = mlx5_xstats_reset,
300 	.xstats_get_names = mlx5_xstats_get_names,
301 	.dev_infos_get = mlx5_dev_infos_get,
302 	.rx_descriptor_status = mlx5_rx_descriptor_status,
303 	.tx_descriptor_status = mlx5_tx_descriptor_status,
304 };
305 
306 /* Available operators in flow isolated mode. */
307 const struct eth_dev_ops mlx5_dev_ops_isolate = {
308 	.dev_configure = mlx5_dev_configure,
309 	.dev_start = mlx5_dev_start,
310 	.dev_stop = mlx5_dev_stop,
311 	.dev_set_link_down = mlx5_set_link_down,
312 	.dev_set_link_up = mlx5_set_link_up,
313 	.dev_close = mlx5_dev_close,
314 	.link_update = mlx5_link_update,
315 	.stats_get = mlx5_stats_get,
316 	.stats_reset = mlx5_stats_reset,
317 	.xstats_get = mlx5_xstats_get,
318 	.xstats_reset = mlx5_xstats_reset,
319 	.xstats_get_names = mlx5_xstats_get_names,
320 	.dev_infos_get = mlx5_dev_infos_get,
321 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
322 	.vlan_filter_set = mlx5_vlan_filter_set,
323 	.rx_queue_setup = mlx5_rx_queue_setup,
324 	.tx_queue_setup = mlx5_tx_queue_setup,
325 	.rx_queue_release = mlx5_rx_queue_release,
326 	.tx_queue_release = mlx5_tx_queue_release,
327 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
328 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
329 	.mac_addr_remove = mlx5_mac_addr_remove,
330 	.mac_addr_add = mlx5_mac_addr_add,
331 	.mac_addr_set = mlx5_mac_addr_set,
332 	.mtu_set = mlx5_dev_set_mtu,
333 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
334 	.vlan_offload_set = mlx5_vlan_offload_set,
335 	.filter_ctrl = mlx5_dev_filter_ctrl,
336 	.rx_descriptor_status = mlx5_rx_descriptor_status,
337 	.tx_descriptor_status = mlx5_tx_descriptor_status,
338 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
339 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
340 	.is_removed = mlx5_is_removed,
341 };
342 
343 static struct {
344 	struct rte_pci_addr pci_addr; /* associated PCI address */
345 	uint32_t ports; /* physical ports bitfield. */
346 } mlx5_dev[32];
347 
348 /**
349  * Get device index in mlx5_dev[] from PCI bus address.
350  *
351  * @param[in] pci_addr
352  *   PCI bus address to look for.
353  *
354  * @return
355  *   mlx5_dev[] index on success, -1 on failure.
356  */
357 static int
358 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
359 {
360 	unsigned int i;
361 	int ret = -1;
362 
363 	assert(pci_addr != NULL);
364 	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
365 		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
366 		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
367 		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
368 		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
369 			return i;
370 		if ((mlx5_dev[i].ports == 0) && (ret == -1))
371 			ret = i;
372 	}
373 	return ret;
374 }
375 
376 /**
377  * Verify and store value for device argument.
378  *
379  * @param[in] key
380  *   Key argument to verify.
381  * @param[in] val
382  *   Value associated with key.
383  * @param opaque
384  *   User data.
385  *
386  * @return
387  *   0 on success, a negative errno value otherwise and rte_errno is set.
388  */
389 static int
390 mlx5_args_check(const char *key, const char *val, void *opaque)
391 {
392 	struct mlx5_dev_config *config = opaque;
393 	unsigned long tmp;
394 
395 	errno = 0;
396 	tmp = strtoul(val, NULL, 0);
397 	if (errno) {
398 		rte_errno = errno;
399 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
400 		return -rte_errno;
401 	}
402 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
403 		config->cqe_comp = !!tmp;
404 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
405 		config->txq_inline = tmp;
406 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
407 		config->txqs_inline = tmp;
408 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
409 		config->mps = !!tmp ? config->mps : 0;
410 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
411 		config->mpw_hdr_dseg = !!tmp;
412 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
413 		config->inline_max_packet_sz = tmp;
414 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
415 		config->tx_vec_en = !!tmp;
416 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
417 		config->rx_vec_en = !!tmp;
418 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
419 		config->vf_nl_en = !!tmp;
420 	} else {
421 		DRV_LOG(WARNING, "%s: unknown parameter", key);
422 		rte_errno = EINVAL;
423 		return -rte_errno;
424 	}
425 	return 0;
426 }
427 
428 /**
429  * Parse device parameters.
430  *
431  * @param config
432  *   Pointer to device configuration structure.
433  * @param devargs
434  *   Device arguments structure.
435  *
436  * @return
437  *   0 on success, a negative errno value otherwise and rte_errno is set.
438  */
439 static int
440 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
441 {
442 	const char **params = (const char *[]){
443 		MLX5_RXQ_CQE_COMP_EN,
444 		MLX5_TXQ_INLINE,
445 		MLX5_TXQS_MIN_INLINE,
446 		MLX5_TXQ_MPW_EN,
447 		MLX5_TXQ_MPW_HDR_DSEG_EN,
448 		MLX5_TXQ_MAX_INLINE_LEN,
449 		MLX5_TX_VEC_EN,
450 		MLX5_RX_VEC_EN,
451 		MLX5_VF_NL_EN,
452 		NULL,
453 	};
454 	struct rte_kvargs *kvlist;
455 	int ret = 0;
456 	int i;
457 
458 	if (devargs == NULL)
459 		return 0;
460 	/* Following UGLY cast is done to pass checkpatch. */
461 	kvlist = rte_kvargs_parse(devargs->args, params);
462 	if (kvlist == NULL)
463 		return 0;
464 	/* Process parameters. */
465 	for (i = 0; (params[i] != NULL); ++i) {
466 		if (rte_kvargs_count(kvlist, params[i])) {
467 			ret = rte_kvargs_process(kvlist, params[i],
468 						 mlx5_args_check, config);
469 			if (ret) {
470 				rte_errno = EINVAL;
471 				rte_kvargs_free(kvlist);
472 				return -rte_errno;
473 			}
474 		}
475 	}
476 	rte_kvargs_free(kvlist);
477 	return 0;
478 }
479 
480 static struct rte_pci_driver mlx5_driver;
481 
482 /*
483  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
484  * local resource used by both primary and secondary to avoid duplicate
485  * reservation.
486  * The space has to be available on both primary and secondary process,
487  * TXQ UAR maps to this area using fixed mmap w/o double check.
488  */
489 static void *uar_base;
490 
491 static int
492 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
493 		const struct rte_memseg *ms, void *arg)
494 {
495 	void **addr = arg;
496 
497 	if (*addr == NULL)
498 		*addr = ms->addr;
499 	else
500 		*addr = RTE_MIN(*addr, ms->addr);
501 
502 	return 0;
503 }
504 
505 /**
506  * Reserve UAR address space for primary process.
507  *
508  * @param[in] dev
509  *   Pointer to Ethernet device.
510  *
511  * @return
512  *   0 on success, a negative errno value otherwise and rte_errno is set.
513  */
514 static int
515 mlx5_uar_init_primary(struct rte_eth_dev *dev)
516 {
517 	struct priv *priv = dev->data->dev_private;
518 	void *addr = (void *)0;
519 
520 	if (uar_base) { /* UAR address space mapped. */
521 		priv->uar_base = uar_base;
522 		return 0;
523 	}
524 	/* find out lower bound of hugepage segments */
525 	rte_memseg_walk(find_lower_va_bound, &addr);
526 
527 	/* keep distance to hugepages to minimize potential conflicts. */
528 	addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
529 	/* anonymous mmap, no real memory consumption. */
530 	addr = mmap(addr, MLX5_UAR_SIZE,
531 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
532 	if (addr == MAP_FAILED) {
533 		DRV_LOG(ERR,
534 			"port %u failed to reserve UAR address space, please"
535 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
536 			dev->data->port_id);
537 		rte_errno = ENOMEM;
538 		return -rte_errno;
539 	}
540 	/* Accept either same addr or a new addr returned from mmap if target
541 	 * range occupied.
542 	 */
543 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
544 		dev->data->port_id, addr);
545 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
546 	uar_base = addr; /* process local, don't reserve again. */
547 	return 0;
548 }
549 
550 /**
551  * Reserve UAR address space for secondary process, align with
552  * primary process.
553  *
554  * @param[in] dev
555  *   Pointer to Ethernet device.
556  *
557  * @return
558  *   0 on success, a negative errno value otherwise and rte_errno is set.
559  */
560 static int
561 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
562 {
563 	struct priv *priv = dev->data->dev_private;
564 	void *addr;
565 
566 	assert(priv->uar_base);
567 	if (uar_base) { /* already reserved. */
568 		assert(uar_base == priv->uar_base);
569 		return 0;
570 	}
571 	/* anonymous mmap, no real memory consumption. */
572 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
573 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
574 	if (addr == MAP_FAILED) {
575 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
576 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
577 		rte_errno = ENXIO;
578 		return -rte_errno;
579 	}
580 	if (priv->uar_base != addr) {
581 		DRV_LOG(ERR,
582 			"port %u UAR address %p size %llu occupied, please"
583 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
584 			" --base-virtaddr",
585 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
586 		rte_errno = ENXIO;
587 		return -rte_errno;
588 	}
589 	uar_base = addr; /* process local, don't reserve again */
590 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
591 		dev->data->port_id, addr);
592 	return 0;
593 }
594 
595 /**
596  * DPDK callback to register a PCI device.
597  *
598  * This function creates an Ethernet device for each port of a given
599  * PCI device.
600  *
601  * @param[in] pci_drv
602  *   PCI driver structure (mlx5_driver).
603  * @param[in] pci_dev
604  *   PCI device information.
605  *
606  * @return
607  *   0 on success, a negative errno value otherwise and rte_errno is set.
608  */
609 static int
610 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
611 	       struct rte_pci_device *pci_dev)
612 {
613 	struct ibv_device **list = NULL;
614 	struct ibv_device *ibv_dev;
615 	int err = 0;
616 	struct ibv_context *attr_ctx = NULL;
617 	struct ibv_device_attr_ex device_attr;
618 	unsigned int vf;
619 	unsigned int mps;
620 	unsigned int cqe_comp;
621 	unsigned int tunnel_en = 0;
622 	int idx;
623 	int i;
624 	struct mlx5dv_context attrs_out = {0};
625 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
626 	struct ibv_counter_set_description cs_desc;
627 #endif
628 
629 	assert(pci_drv == &mlx5_driver);
630 	/* Get mlx5_dev[] index. */
631 	idx = mlx5_dev_idx(&pci_dev->addr);
632 	if (idx == -1) {
633 		DRV_LOG(ERR, "this driver cannot support any more adapters");
634 		err = ENOMEM;
635 		goto error;
636 	}
637 	DRV_LOG(DEBUG, "using driver device index %d", idx);
638 	/* Save PCI address. */
639 	mlx5_dev[idx].pci_addr = pci_dev->addr;
640 	list = mlx5_glue->get_device_list(&i);
641 	if (list == NULL) {
642 		assert(errno);
643 		err = errno;
644 		if (errno == ENOSYS)
645 			DRV_LOG(ERR,
646 				"cannot list devices, is ib_uverbs loaded?");
647 		goto error;
648 	}
649 	assert(i >= 0);
650 	/*
651 	 * For each listed device, check related sysfs entry against
652 	 * the provided PCI ID.
653 	 */
654 	while (i != 0) {
655 		struct rte_pci_addr pci_addr;
656 
657 		--i;
658 		DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
659 		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
660 			continue;
661 		if ((pci_dev->addr.domain != pci_addr.domain) ||
662 		    (pci_dev->addr.bus != pci_addr.bus) ||
663 		    (pci_dev->addr.devid != pci_addr.devid) ||
664 		    (pci_dev->addr.function != pci_addr.function))
665 			continue;
666 		DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
667 			list[i]->name);
668 		vf = ((pci_dev->id.device_id ==
669 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
670 		      (pci_dev->id.device_id ==
671 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
672 		      (pci_dev->id.device_id ==
673 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
674 		      (pci_dev->id.device_id ==
675 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
676 		attr_ctx = mlx5_glue->open_device(list[i]);
677 		rte_errno = errno;
678 		err = rte_errno;
679 		break;
680 	}
681 	if (attr_ctx == NULL) {
682 		mlx5_glue->free_device_list(list);
683 		switch (err) {
684 		case 0:
685 			DRV_LOG(ERR,
686 				"cannot access device, is mlx5_ib loaded?");
687 			err = ENODEV;
688 			goto error;
689 		case EINVAL:
690 			DRV_LOG(ERR,
691 				"cannot use device, are drivers up to date?");
692 			goto error;
693 		}
694 	}
695 	ibv_dev = list[i];
696 	DRV_LOG(DEBUG, "device opened");
697 	/*
698 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
699 	 * as all ConnectX-5 devices.
700 	 */
701 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
702 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
703 #endif
704 	mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
705 	if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
706 		if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
707 			DRV_LOG(DEBUG, "enhanced MPW is supported");
708 			mps = MLX5_MPW_ENHANCED;
709 		} else {
710 			DRV_LOG(DEBUG, "MPW is supported");
711 			mps = MLX5_MPW;
712 		}
713 	} else {
714 		DRV_LOG(DEBUG, "MPW isn't supported");
715 		mps = MLX5_MPW_DISABLED;
716 	}
717 	if (RTE_CACHE_LINE_SIZE == 128 &&
718 	    !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
719 		cqe_comp = 0;
720 	else
721 		cqe_comp = 1;
722 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
723 	if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
724 		tunnel_en = ((attrs_out.tunnel_offloads_caps &
725 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
726 			     (attrs_out.tunnel_offloads_caps &
727 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
728 	}
729 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
730 		tunnel_en ? "" : "not ");
731 #else
732 	DRV_LOG(WARNING,
733 		"tunnel offloading disabled due to old OFED/rdma-core version");
734 #endif
735 	if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr)) {
736 		err = errno;
737 		goto error;
738 	}
739 	DRV_LOG(INFO, "%u port(s) detected",
740 		device_attr.orig_attr.phys_port_cnt);
741 	for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
742 		char name[RTE_ETH_NAME_MAX_LEN];
743 		int len;
744 		uint32_t port = i + 1; /* ports are indexed from one */
745 		uint32_t test = (1 << i);
746 		struct ibv_context *ctx = NULL;
747 		struct ibv_port_attr port_attr;
748 		struct ibv_pd *pd = NULL;
749 		struct priv *priv = NULL;
750 		struct rte_eth_dev *eth_dev = NULL;
751 		struct ibv_device_attr_ex device_attr_ex;
752 		struct ether_addr mac;
753 		struct mlx5_dev_config config = {
754 			.cqe_comp = cqe_comp,
755 			.mps = mps,
756 			.tunnel_en = tunnel_en,
757 			.tx_vec_en = 1,
758 			.rx_vec_en = 1,
759 			.mpw_hdr_dseg = 0,
760 			.txq_inline = MLX5_ARG_UNSET,
761 			.txqs_inline = MLX5_ARG_UNSET,
762 			.inline_max_packet_sz = MLX5_ARG_UNSET,
763 			.vf_nl_en = 1,
764 		};
765 
766 		len = snprintf(name, sizeof(name), PCI_PRI_FMT,
767 			 pci_dev->addr.domain, pci_dev->addr.bus,
768 			 pci_dev->addr.devid, pci_dev->addr.function);
769 		if (device_attr.orig_attr.phys_port_cnt > 1)
770 			snprintf(name + len, sizeof(name), " port %u", i);
771 		mlx5_dev[idx].ports |= test;
772 		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
773 			eth_dev = rte_eth_dev_attach_secondary(name);
774 			if (eth_dev == NULL) {
775 				DRV_LOG(ERR, "can not attach rte ethdev");
776 				rte_errno = ENOMEM;
777 				err = rte_errno;
778 				goto error;
779 			}
780 			eth_dev->device = &pci_dev->device;
781 			eth_dev->dev_ops = &mlx5_dev_sec_ops;
782 			err = mlx5_uar_init_secondary(eth_dev);
783 			if (err)
784 				goto error;
785 			/* Receive command fd from primary process */
786 			err = mlx5_socket_connect(eth_dev);
787 			if (err)
788 				goto error;
789 			/* Remap UAR for Tx queues. */
790 			err = mlx5_tx_uar_remap(eth_dev, err);
791 			if (err)
792 				goto error;
793 			/*
794 			 * Ethdev pointer is still required as input since
795 			 * the primary device is not accessible from the
796 			 * secondary process.
797 			 */
798 			eth_dev->rx_pkt_burst =
799 				mlx5_select_rx_function(eth_dev);
800 			eth_dev->tx_pkt_burst =
801 				mlx5_select_tx_function(eth_dev);
802 			continue;
803 		}
804 		DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
805 		ctx = mlx5_glue->open_device(ibv_dev);
806 		if (ctx == NULL) {
807 			err = ENODEV;
808 			goto port_error;
809 		}
810 		/* Check port status. */
811 		err = mlx5_glue->query_port(ctx, port, &port_attr);
812 		if (err) {
813 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
814 			goto port_error;
815 		}
816 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
817 			DRV_LOG(ERR,
818 				"port %d is not configured in Ethernet mode",
819 				port);
820 			err = EINVAL;
821 			goto port_error;
822 		}
823 		if (port_attr.state != IBV_PORT_ACTIVE)
824 			DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
825 				port,
826 				mlx5_glue->port_state_str(port_attr.state),
827 				port_attr.state);
828 		/* Allocate protection domain. */
829 		pd = mlx5_glue->alloc_pd(ctx);
830 		if (pd == NULL) {
831 			DRV_LOG(ERR, "PD allocation failure");
832 			err = ENOMEM;
833 			goto port_error;
834 		}
835 		mlx5_dev[idx].ports |= test;
836 		/* from rte_ethdev.c */
837 		priv = rte_zmalloc("ethdev private structure",
838 				   sizeof(*priv),
839 				   RTE_CACHE_LINE_SIZE);
840 		if (priv == NULL) {
841 			DRV_LOG(ERR, "priv allocation failure");
842 			err = ENOMEM;
843 			goto port_error;
844 		}
845 		priv->ctx = ctx;
846 		strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
847 			sizeof(priv->ibdev_path));
848 		priv->device_attr = device_attr;
849 		priv->port = port;
850 		priv->pd = pd;
851 		priv->mtu = ETHER_MTU;
852 		err = mlx5_args(&config, pci_dev->device.devargs);
853 		if (err) {
854 			DRV_LOG(ERR, "failed to process device arguments: %s",
855 				strerror(err));
856 			goto port_error;
857 		}
858 		if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
859 			DRV_LOG(ERR, "ibv_query_device_ex() failed");
860 			err = errno;
861 			goto port_error;
862 		}
863 		config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
864 				    IBV_DEVICE_RAW_IP_CSUM);
865 		DRV_LOG(DEBUG, "checksum offloading is %ssupported",
866 			(config.hw_csum ? "" : "not "));
867 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
868 		config.flow_counter_en = !!(device_attr.max_counter_sets);
869 		mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
870 		DRV_LOG(DEBUG,
871 			"counter type = %d, num of cs = %ld, attributes = %d",
872 			cs_desc.counter_type, cs_desc.num_of_cs,
873 			cs_desc.attributes);
874 #endif
875 		config.ind_table_max_size =
876 			device_attr_ex.rss_caps.max_rwq_indirection_table_size;
877 		/* Remove this check once DPDK supports larger/variable
878 		 * indirection tables. */
879 		if (config.ind_table_max_size >
880 				(unsigned int)ETH_RSS_RETA_SIZE_512)
881 			config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
882 		DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
883 			config.ind_table_max_size);
884 		config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
885 					 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
886 		DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
887 			(config.hw_vlan_strip ? "" : "not "));
888 
889 		config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
890 					 IBV_RAW_PACKET_CAP_SCATTER_FCS);
891 		DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
892 			(config.hw_fcs_strip ? "" : "not "));
893 
894 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
895 		config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
896 #endif
897 		DRV_LOG(DEBUG,
898 			"hardware Rx end alignment padding is %ssupported",
899 			(config.hw_padding ? "" : "not "));
900 		config.vf = vf;
901 		config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
902 			      (device_attr_ex.tso_caps.supported_qpts &
903 			      (1 << IBV_QPT_RAW_PACKET)));
904 		if (config.tso)
905 			config.tso_max_payload_sz =
906 					device_attr_ex.tso_caps.max_tso;
907 		if (config.mps && !mps) {
908 			DRV_LOG(ERR,
909 				"multi-packet send not supported on this device"
910 				" (" MLX5_TXQ_MPW_EN ")");
911 			err = ENOTSUP;
912 			goto port_error;
913 		}
914 		DRV_LOG(INFO, "%s MPS is %s",
915 			config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
916 			config.mps != MLX5_MPW_DISABLED ? "enabled" :
917 			"disabled");
918 		if (config.cqe_comp && !cqe_comp) {
919 			DRV_LOG(WARNING, "Rx CQE compression isn't supported");
920 			config.cqe_comp = 0;
921 		}
922 		eth_dev = rte_eth_dev_allocate(name);
923 		if (eth_dev == NULL) {
924 			DRV_LOG(ERR, "can not allocate rte ethdev");
925 			err = ENOMEM;
926 			goto port_error;
927 		}
928 		eth_dev->data->dev_private = priv;
929 		priv->dev = eth_dev;
930 		eth_dev->data->mac_addrs = priv->mac;
931 		eth_dev->device = &pci_dev->device;
932 		rte_eth_copy_pci_info(eth_dev, pci_dev);
933 		eth_dev->device->driver = &mlx5_driver.driver;
934 		err = mlx5_uar_init_primary(eth_dev);
935 		if (err)
936 			goto port_error;
937 		/* Configure the first MAC address by default. */
938 		if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
939 			DRV_LOG(ERR,
940 				"port %u cannot get MAC address, is mlx5_en"
941 				" loaded? (errno: %s)",
942 				eth_dev->data->port_id, strerror(errno));
943 			err = ENODEV;
944 			goto port_error;
945 		}
946 		DRV_LOG(INFO,
947 			"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
948 			eth_dev->data->port_id,
949 			mac.addr_bytes[0], mac.addr_bytes[1],
950 			mac.addr_bytes[2], mac.addr_bytes[3],
951 			mac.addr_bytes[4], mac.addr_bytes[5]);
952 #ifndef NDEBUG
953 		{
954 			char ifname[IF_NAMESIZE];
955 
956 			if (mlx5_get_ifname(eth_dev, &ifname) == 0)
957 				DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
958 					eth_dev->data->port_id, ifname);
959 			else
960 				DRV_LOG(DEBUG, "port %u ifname is unknown",
961 					eth_dev->data->port_id);
962 		}
963 #endif
964 		/* Get actual MTU if possible. */
965 		err = mlx5_get_mtu(eth_dev, &priv->mtu);
966 		if (err)
967 			goto port_error;
968 		DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
969 			priv->mtu);
970 		/*
971 		 * Initialize burst functions to prevent crashes before link-up.
972 		 */
973 		eth_dev->rx_pkt_burst = removed_rx_burst;
974 		eth_dev->tx_pkt_burst = removed_tx_burst;
975 		eth_dev->dev_ops = &mlx5_dev_ops;
976 		/* Register MAC address. */
977 		claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
978 		priv->nl_socket = -1;
979 		priv->nl_sn = 0;
980 		if (vf && config.vf_nl_en) {
981 			priv->nl_socket = mlx5_nl_init(RTMGRP_LINK);
982 			if (priv->nl_socket < 0)
983 				priv->nl_socket = -1;
984 			mlx5_nl_mac_addr_sync(eth_dev);
985 		}
986 		TAILQ_INIT(&priv->flows);
987 		TAILQ_INIT(&priv->ctrl_flows);
988 		/* Hint libmlx5 to use PMD allocator for data plane resources */
989 		struct mlx5dv_ctx_allocators alctr = {
990 			.alloc = &mlx5_alloc_verbs_buf,
991 			.free = &mlx5_free_verbs_buf,
992 			.data = priv,
993 		};
994 		mlx5_glue->dv_set_context_attr(ctx,
995 					       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
996 					       (void *)((uintptr_t)&alctr));
997 		/* Bring Ethernet device up. */
998 		DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
999 			eth_dev->data->port_id);
1000 		mlx5_set_link_up(eth_dev);
1001 		/*
1002 		 * Even though the interrupt handler is not installed yet,
1003 		 * interrupts will still trigger on the asyn_fd from
1004 		 * Verbs context returned by ibv_open_device().
1005 		 */
1006 		mlx5_link_update(eth_dev, 0);
1007 		/* Store device configuration on private structure. */
1008 		priv->config = config;
1009 		continue;
1010 port_error:
1011 		if (priv)
1012 			rte_free(priv);
1013 		if (pd)
1014 			claim_zero(mlx5_glue->dealloc_pd(pd));
1015 		if (ctx)
1016 			claim_zero(mlx5_glue->close_device(ctx));
1017 		break;
1018 	}
1019 	/*
1020 	 * XXX if something went wrong in the loop above, there is a resource
1021 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1022 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
1023 	 * way to enumerate the registered ethdevs to free the previous ones.
1024 	 */
1025 	/* no port found, complain */
1026 	if (!mlx5_dev[idx].ports) {
1027 		rte_errno = ENODEV;
1028 		err = rte_errno;
1029 	}
1030 error:
1031 	if (attr_ctx)
1032 		claim_zero(mlx5_glue->close_device(attr_ctx));
1033 	if (list)
1034 		mlx5_glue->free_device_list(list);
1035 	if (err) {
1036 		rte_errno = err;
1037 		return -rte_errno;
1038 	}
1039 	return 0;
1040 }
1041 
1042 static const struct rte_pci_id mlx5_pci_id_map[] = {
1043 	{
1044 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1045 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1046 	},
1047 	{
1048 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1049 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1050 	},
1051 	{
1052 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1053 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1054 	},
1055 	{
1056 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1057 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1058 	},
1059 	{
1060 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1061 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1062 	},
1063 	{
1064 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1065 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1066 	},
1067 	{
1068 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1069 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1070 	},
1071 	{
1072 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1073 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1074 	},
1075 	{
1076 		.vendor_id = 0
1077 	}
1078 };
1079 
1080 static struct rte_pci_driver mlx5_driver = {
1081 	.driver = {
1082 		.name = MLX5_DRIVER_NAME
1083 	},
1084 	.id_table = mlx5_pci_id_map,
1085 	.probe = mlx5_pci_probe,
1086 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1087 };
1088 
1089 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1090 
1091 /**
1092  * Suffix RTE_EAL_PMD_PATH with "-glue".
1093  *
1094  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1095  * suffixing its last component.
1096  *
1097  * @param buf[out]
1098  *   Output buffer, should be large enough otherwise NULL is returned.
1099  * @param size
1100  *   Size of @p out.
1101  *
1102  * @return
1103  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1104  */
1105 static char *
1106 mlx5_glue_path(char *buf, size_t size)
1107 {
1108 	static const char *const bad[] = { "/", ".", "..", NULL };
1109 	const char *path = RTE_EAL_PMD_PATH;
1110 	size_t len = strlen(path);
1111 	size_t off;
1112 	int i;
1113 
1114 	while (len && path[len - 1] == '/')
1115 		--len;
1116 	for (off = len; off && path[off - 1] != '/'; --off)
1117 		;
1118 	for (i = 0; bad[i]; ++i)
1119 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1120 			goto error;
1121 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1122 	if (i == -1 || (size_t)i >= size)
1123 		goto error;
1124 	return buf;
1125 error:
1126 	DRV_LOG(ERR,
1127 		"unable to append \"-glue\" to last component of"
1128 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1129 		" please re-configure DPDK");
1130 	return NULL;
1131 }
1132 
1133 /**
1134  * Initialization routine for run-time dependency on rdma-core.
1135  */
1136 static int
1137 mlx5_glue_init(void)
1138 {
1139 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1140 	const char *path[] = {
1141 		/*
1142 		 * A basic security check is necessary before trusting
1143 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1144 		 */
1145 		(geteuid() == getuid() && getegid() == getgid() ?
1146 		 getenv("MLX5_GLUE_PATH") : NULL),
1147 		/*
1148 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1149 		 * variant, otherwise let dlopen() look up libraries on its
1150 		 * own.
1151 		 */
1152 		(*RTE_EAL_PMD_PATH ?
1153 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1154 	};
1155 	unsigned int i = 0;
1156 	void *handle = NULL;
1157 	void **sym;
1158 	const char *dlmsg;
1159 
1160 	while (!handle && i != RTE_DIM(path)) {
1161 		const char *end;
1162 		size_t len;
1163 		int ret;
1164 
1165 		if (!path[i]) {
1166 			++i;
1167 			continue;
1168 		}
1169 		end = strpbrk(path[i], ":;");
1170 		if (!end)
1171 			end = path[i] + strlen(path[i]);
1172 		len = end - path[i];
1173 		ret = 0;
1174 		do {
1175 			char name[ret + 1];
1176 
1177 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1178 				       (int)len, path[i],
1179 				       (!len || *(end - 1) == '/') ? "" : "/");
1180 			if (ret == -1)
1181 				break;
1182 			if (sizeof(name) != (size_t)ret + 1)
1183 				continue;
1184 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1185 				name);
1186 			handle = dlopen(name, RTLD_LAZY);
1187 			break;
1188 		} while (1);
1189 		path[i] = end + 1;
1190 		if (!*end)
1191 			++i;
1192 	}
1193 	if (!handle) {
1194 		rte_errno = EINVAL;
1195 		dlmsg = dlerror();
1196 		if (dlmsg)
1197 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1198 		goto glue_error;
1199 	}
1200 	sym = dlsym(handle, "mlx5_glue");
1201 	if (!sym || !*sym) {
1202 		rte_errno = EINVAL;
1203 		dlmsg = dlerror();
1204 		if (dlmsg)
1205 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1206 		goto glue_error;
1207 	}
1208 	mlx5_glue = *sym;
1209 	return 0;
1210 glue_error:
1211 	if (handle)
1212 		dlclose(handle);
1213 	DRV_LOG(WARNING,
1214 		"cannot initialize PMD due to missing run-time dependency on"
1215 		" rdma-core libraries (libibverbs, libmlx5)");
1216 	return -rte_errno;
1217 }
1218 
1219 #endif
1220 
1221 /**
1222  * Driver initialization routine.
1223  */
1224 RTE_INIT(rte_mlx5_pmd_init);
1225 static void
1226 rte_mlx5_pmd_init(void)
1227 {
1228 	/* Build the static table for ptype conversion. */
1229 	mlx5_set_ptype_table();
1230 	/*
1231 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1232 	 * huge pages. Calling ibv_fork_init() during init allows
1233 	 * applications to use fork() safely for purposes other than
1234 	 * using this PMD, which is not supported in forked processes.
1235 	 */
1236 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1237 	/* Match the size of Rx completion entry to the size of a cacheline. */
1238 	if (RTE_CACHE_LINE_SIZE == 128)
1239 		setenv("MLX5_CQE_SIZE", "128", 0);
1240 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1241 	if (mlx5_glue_init())
1242 		return;
1243 	assert(mlx5_glue);
1244 #endif
1245 #ifndef NDEBUG
1246 	/* Glue structure must not contain any NULL pointers. */
1247 	{
1248 		unsigned int i;
1249 
1250 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1251 			assert(((const void *const *)mlx5_glue)[i]);
1252 	}
1253 #endif
1254 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1255 		DRV_LOG(ERR,
1256 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1257 			mlx5_glue->version, MLX5_GLUE_VERSION);
1258 		return;
1259 	}
1260 	mlx5_glue->fork_init();
1261 	rte_pci_register(&mlx5_driver);
1262 }
1263 
1264 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1265 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1266 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1267 
1268 /** Initialize driver log type. */
1269 RTE_INIT(vdev_netvsc_init_log)
1270 {
1271 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1272 	if (mlx5_logtype >= 0)
1273 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1274 }
1275