xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 2f82d143fb318042f47a50694baa4507b51b7381)
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 /* Allow L3 VXLAN flow creation. */
73 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
74 
75 /* Activate Netlink support in VF mode. */
76 #define MLX5_VF_NL_EN "vf_nl_en"
77 
78 #ifndef HAVE_IBV_MLX5_MOD_MPW
79 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
80 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
81 #endif
82 
83 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
84 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
85 #endif
86 
87 /** Driver-specific log messages type. */
88 int mlx5_logtype;
89 
90 /**
91  * Retrieve integer value from environment variable.
92  *
93  * @param[in] name
94  *   Environment variable name.
95  *
96  * @return
97  *   Integer value, 0 if the variable is not set.
98  */
99 int
100 mlx5_getenv_int(const char *name)
101 {
102 	const char *val = getenv(name);
103 
104 	if (val == NULL)
105 		return 0;
106 	return atoi(val);
107 }
108 
109 /**
110  * Verbs callback to allocate a memory. This function should allocate the space
111  * according to the size provided residing inside a huge page.
112  * Please note that all allocation must respect the alignment from libmlx5
113  * (i.e. currently sysconf(_SC_PAGESIZE)).
114  *
115  * @param[in] size
116  *   The size in bytes of the memory to allocate.
117  * @param[in] data
118  *   A pointer to the callback data.
119  *
120  * @return
121  *   Allocated buffer, NULL otherwise and rte_errno is set.
122  */
123 static void *
124 mlx5_alloc_verbs_buf(size_t size, void *data)
125 {
126 	struct priv *priv = data;
127 	void *ret;
128 	size_t alignment = sysconf(_SC_PAGESIZE);
129 	unsigned int socket = SOCKET_ID_ANY;
130 
131 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
132 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
133 
134 		socket = ctrl->socket;
135 	} else if (priv->verbs_alloc_ctx.type ==
136 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
137 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
138 
139 		socket = ctrl->socket;
140 	}
141 	assert(data != NULL);
142 	ret = rte_malloc_socket(__func__, size, alignment, socket);
143 	if (!ret && size)
144 		rte_errno = ENOMEM;
145 	return ret;
146 }
147 
148 /**
149  * Verbs callback to free a memory.
150  *
151  * @param[in] ptr
152  *   A pointer to the memory to free.
153  * @param[in] data
154  *   A pointer to the callback data.
155  */
156 static void
157 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
158 {
159 	assert(data != NULL);
160 	rte_free(ptr);
161 }
162 
163 /**
164  * DPDK callback to close the device.
165  *
166  * Destroy all queues and objects, free memory.
167  *
168  * @param dev
169  *   Pointer to Ethernet device structure.
170  */
171 static void
172 mlx5_dev_close(struct rte_eth_dev *dev)
173 {
174 	struct priv *priv = dev->data->dev_private;
175 	unsigned int i;
176 	int ret;
177 
178 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
179 		dev->data->port_id,
180 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
181 	/* In case mlx5_dev_stop() has not been called. */
182 	mlx5_dev_interrupt_handler_uninstall(dev);
183 	mlx5_traffic_disable(dev);
184 	/* Prevent crashes when queues are still in use. */
185 	dev->rx_pkt_burst = removed_rx_burst;
186 	dev->tx_pkt_burst = removed_tx_burst;
187 	if (priv->rxqs != NULL) {
188 		/* XXX race condition if mlx5_rx_burst() is still running. */
189 		usleep(1000);
190 		for (i = 0; (i != priv->rxqs_n); ++i)
191 			mlx5_rxq_release(dev, i);
192 		priv->rxqs_n = 0;
193 		priv->rxqs = NULL;
194 	}
195 	if (priv->txqs != NULL) {
196 		/* XXX race condition if mlx5_tx_burst() is still running. */
197 		usleep(1000);
198 		for (i = 0; (i != priv->txqs_n); ++i)
199 			mlx5_txq_release(dev, i);
200 		priv->txqs_n = 0;
201 		priv->txqs = NULL;
202 	}
203 	mlx5_flow_delete_drop_queue(dev);
204 	if (priv->pd != NULL) {
205 		assert(priv->ctx != NULL);
206 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
207 		claim_zero(mlx5_glue->close_device(priv->ctx));
208 	} else
209 		assert(priv->ctx == NULL);
210 	if (priv->rss_conf.rss_key != NULL)
211 		rte_free(priv->rss_conf.rss_key);
212 	if (priv->reta_idx != NULL)
213 		rte_free(priv->reta_idx);
214 	if (priv->primary_socket)
215 		mlx5_socket_uninit(dev);
216 	if (priv->config.vf)
217 		mlx5_nl_mac_addr_flush(dev);
218 	if (priv->nl_socket >= 0)
219 		close(priv->nl_socket);
220 	ret = mlx5_hrxq_ibv_verify(dev);
221 	if (ret)
222 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
223 			dev->data->port_id);
224 	ret = mlx5_ind_table_ibv_verify(dev);
225 	if (ret)
226 		DRV_LOG(WARNING, "port %u some indirection table still remain",
227 			dev->data->port_id);
228 	ret = mlx5_rxq_ibv_verify(dev);
229 	if (ret)
230 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
231 			dev->data->port_id);
232 	ret = mlx5_rxq_verify(dev);
233 	if (ret)
234 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
235 			dev->data->port_id);
236 	ret = mlx5_txq_ibv_verify(dev);
237 	if (ret)
238 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
239 			dev->data->port_id);
240 	ret = mlx5_txq_verify(dev);
241 	if (ret)
242 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
243 			dev->data->port_id);
244 	ret = mlx5_flow_verify(dev);
245 	if (ret)
246 		DRV_LOG(WARNING, "port %u some flows still remain",
247 			dev->data->port_id);
248 	ret = mlx5_mr_verify(dev);
249 	if (ret)
250 		DRV_LOG(WARNING, "port %u some memory region still remain",
251 			dev->data->port_id);
252 	memset(priv, 0, sizeof(*priv));
253 }
254 
255 const struct eth_dev_ops mlx5_dev_ops = {
256 	.dev_configure = mlx5_dev_configure,
257 	.dev_start = mlx5_dev_start,
258 	.dev_stop = mlx5_dev_stop,
259 	.dev_set_link_down = mlx5_set_link_down,
260 	.dev_set_link_up = mlx5_set_link_up,
261 	.dev_close = mlx5_dev_close,
262 	.promiscuous_enable = mlx5_promiscuous_enable,
263 	.promiscuous_disable = mlx5_promiscuous_disable,
264 	.allmulticast_enable = mlx5_allmulticast_enable,
265 	.allmulticast_disable = mlx5_allmulticast_disable,
266 	.link_update = mlx5_link_update,
267 	.stats_get = mlx5_stats_get,
268 	.stats_reset = mlx5_stats_reset,
269 	.xstats_get = mlx5_xstats_get,
270 	.xstats_reset = mlx5_xstats_reset,
271 	.xstats_get_names = mlx5_xstats_get_names,
272 	.dev_infos_get = mlx5_dev_infos_get,
273 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
274 	.vlan_filter_set = mlx5_vlan_filter_set,
275 	.rx_queue_setup = mlx5_rx_queue_setup,
276 	.tx_queue_setup = mlx5_tx_queue_setup,
277 	.rx_queue_release = mlx5_rx_queue_release,
278 	.tx_queue_release = mlx5_tx_queue_release,
279 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
280 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
281 	.mac_addr_remove = mlx5_mac_addr_remove,
282 	.mac_addr_add = mlx5_mac_addr_add,
283 	.mac_addr_set = mlx5_mac_addr_set,
284 	.set_mc_addr_list = mlx5_set_mc_addr_list,
285 	.mtu_set = mlx5_dev_set_mtu,
286 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
287 	.vlan_offload_set = mlx5_vlan_offload_set,
288 	.reta_update = mlx5_dev_rss_reta_update,
289 	.reta_query = mlx5_dev_rss_reta_query,
290 	.rss_hash_update = mlx5_rss_hash_update,
291 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
292 	.filter_ctrl = mlx5_dev_filter_ctrl,
293 	.rx_descriptor_status = mlx5_rx_descriptor_status,
294 	.tx_descriptor_status = mlx5_tx_descriptor_status,
295 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
296 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
297 	.is_removed = mlx5_is_removed,
298 };
299 
300 static const struct eth_dev_ops mlx5_dev_sec_ops = {
301 	.stats_get = mlx5_stats_get,
302 	.stats_reset = mlx5_stats_reset,
303 	.xstats_get = mlx5_xstats_get,
304 	.xstats_reset = mlx5_xstats_reset,
305 	.xstats_get_names = mlx5_xstats_get_names,
306 	.dev_infos_get = mlx5_dev_infos_get,
307 	.rx_descriptor_status = mlx5_rx_descriptor_status,
308 	.tx_descriptor_status = mlx5_tx_descriptor_status,
309 };
310 
311 /* Available operators in flow isolated mode. */
312 const struct eth_dev_ops mlx5_dev_ops_isolate = {
313 	.dev_configure = mlx5_dev_configure,
314 	.dev_start = mlx5_dev_start,
315 	.dev_stop = mlx5_dev_stop,
316 	.dev_set_link_down = mlx5_set_link_down,
317 	.dev_set_link_up = mlx5_set_link_up,
318 	.dev_close = mlx5_dev_close,
319 	.link_update = mlx5_link_update,
320 	.stats_get = mlx5_stats_get,
321 	.stats_reset = mlx5_stats_reset,
322 	.xstats_get = mlx5_xstats_get,
323 	.xstats_reset = mlx5_xstats_reset,
324 	.xstats_get_names = mlx5_xstats_get_names,
325 	.dev_infos_get = mlx5_dev_infos_get,
326 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
327 	.vlan_filter_set = mlx5_vlan_filter_set,
328 	.rx_queue_setup = mlx5_rx_queue_setup,
329 	.tx_queue_setup = mlx5_tx_queue_setup,
330 	.rx_queue_release = mlx5_rx_queue_release,
331 	.tx_queue_release = mlx5_tx_queue_release,
332 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
333 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
334 	.mac_addr_remove = mlx5_mac_addr_remove,
335 	.mac_addr_add = mlx5_mac_addr_add,
336 	.mac_addr_set = mlx5_mac_addr_set,
337 	.set_mc_addr_list = mlx5_set_mc_addr_list,
338 	.mtu_set = mlx5_dev_set_mtu,
339 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
340 	.vlan_offload_set = mlx5_vlan_offload_set,
341 	.filter_ctrl = mlx5_dev_filter_ctrl,
342 	.rx_descriptor_status = mlx5_rx_descriptor_status,
343 	.tx_descriptor_status = mlx5_tx_descriptor_status,
344 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
345 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
346 	.is_removed = mlx5_is_removed,
347 };
348 
349 static struct {
350 	struct rte_pci_addr pci_addr; /* associated PCI address */
351 	uint32_t ports; /* physical ports bitfield. */
352 } mlx5_dev[32];
353 
354 /**
355  * Get device index in mlx5_dev[] from PCI bus address.
356  *
357  * @param[in] pci_addr
358  *   PCI bus address to look for.
359  *
360  * @return
361  *   mlx5_dev[] index on success, -1 on failure.
362  */
363 static int
364 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
365 {
366 	unsigned int i;
367 	int ret = -1;
368 
369 	assert(pci_addr != NULL);
370 	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
371 		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
372 		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
373 		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
374 		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
375 			return i;
376 		if ((mlx5_dev[i].ports == 0) && (ret == -1))
377 			ret = i;
378 	}
379 	return ret;
380 }
381 
382 /**
383  * Verify and store value for device argument.
384  *
385  * @param[in] key
386  *   Key argument to verify.
387  * @param[in] val
388  *   Value associated with key.
389  * @param opaque
390  *   User data.
391  *
392  * @return
393  *   0 on success, a negative errno value otherwise and rte_errno is set.
394  */
395 static int
396 mlx5_args_check(const char *key, const char *val, void *opaque)
397 {
398 	struct mlx5_dev_config *config = opaque;
399 	unsigned long tmp;
400 
401 	errno = 0;
402 	tmp = strtoul(val, NULL, 0);
403 	if (errno) {
404 		rte_errno = errno;
405 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
406 		return -rte_errno;
407 	}
408 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
409 		config->cqe_comp = !!tmp;
410 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
411 		config->txq_inline = tmp;
412 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
413 		config->txqs_inline = tmp;
414 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
415 		config->mps = !!tmp ? config->mps : 0;
416 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
417 		config->mpw_hdr_dseg = !!tmp;
418 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
419 		config->inline_max_packet_sz = tmp;
420 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
421 		config->tx_vec_en = !!tmp;
422 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
423 		config->rx_vec_en = !!tmp;
424 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
425 		config->l3_vxlan_en = !!tmp;
426 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
427 		config->vf_nl_en = !!tmp;
428 	} else {
429 		DRV_LOG(WARNING, "%s: unknown parameter", key);
430 		rte_errno = EINVAL;
431 		return -rte_errno;
432 	}
433 	return 0;
434 }
435 
436 /**
437  * Parse device parameters.
438  *
439  * @param config
440  *   Pointer to device configuration structure.
441  * @param devargs
442  *   Device arguments structure.
443  *
444  * @return
445  *   0 on success, a negative errno value otherwise and rte_errno is set.
446  */
447 static int
448 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
449 {
450 	const char **params = (const char *[]){
451 		MLX5_RXQ_CQE_COMP_EN,
452 		MLX5_TXQ_INLINE,
453 		MLX5_TXQS_MIN_INLINE,
454 		MLX5_TXQ_MPW_EN,
455 		MLX5_TXQ_MPW_HDR_DSEG_EN,
456 		MLX5_TXQ_MAX_INLINE_LEN,
457 		MLX5_TX_VEC_EN,
458 		MLX5_RX_VEC_EN,
459 		MLX5_L3_VXLAN_EN,
460 		MLX5_VF_NL_EN,
461 		NULL,
462 	};
463 	struct rte_kvargs *kvlist;
464 	int ret = 0;
465 	int i;
466 
467 	if (devargs == NULL)
468 		return 0;
469 	/* Following UGLY cast is done to pass checkpatch. */
470 	kvlist = rte_kvargs_parse(devargs->args, params);
471 	if (kvlist == NULL)
472 		return 0;
473 	/* Process parameters. */
474 	for (i = 0; (params[i] != NULL); ++i) {
475 		if (rte_kvargs_count(kvlist, params[i])) {
476 			ret = rte_kvargs_process(kvlist, params[i],
477 						 mlx5_args_check, config);
478 			if (ret) {
479 				rte_errno = EINVAL;
480 				rte_kvargs_free(kvlist);
481 				return -rte_errno;
482 			}
483 		}
484 	}
485 	rte_kvargs_free(kvlist);
486 	return 0;
487 }
488 
489 static struct rte_pci_driver mlx5_driver;
490 
491 /*
492  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
493  * local resource used by both primary and secondary to avoid duplicate
494  * reservation.
495  * The space has to be available on both primary and secondary process,
496  * TXQ UAR maps to this area using fixed mmap w/o double check.
497  */
498 static void *uar_base;
499 
500 static int
501 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
502 		const struct rte_memseg *ms, void *arg)
503 {
504 	void **addr = arg;
505 
506 	if (*addr == NULL)
507 		*addr = ms->addr;
508 	else
509 		*addr = RTE_MIN(*addr, ms->addr);
510 
511 	return 0;
512 }
513 
514 /**
515  * Reserve UAR address space for primary process.
516  *
517  * @param[in] dev
518  *   Pointer to Ethernet device.
519  *
520  * @return
521  *   0 on success, a negative errno value otherwise and rte_errno is set.
522  */
523 static int
524 mlx5_uar_init_primary(struct rte_eth_dev *dev)
525 {
526 	struct priv *priv = dev->data->dev_private;
527 	void *addr = (void *)0;
528 
529 	if (uar_base) { /* UAR address space mapped. */
530 		priv->uar_base = uar_base;
531 		return 0;
532 	}
533 	/* find out lower bound of hugepage segments */
534 	rte_memseg_walk(find_lower_va_bound, &addr);
535 
536 	/* keep distance to hugepages to minimize potential conflicts. */
537 	addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
538 	/* anonymous mmap, no real memory consumption. */
539 	addr = mmap(addr, MLX5_UAR_SIZE,
540 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
541 	if (addr == MAP_FAILED) {
542 		DRV_LOG(ERR,
543 			"port %u failed to reserve UAR address space, please"
544 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
545 			dev->data->port_id);
546 		rte_errno = ENOMEM;
547 		return -rte_errno;
548 	}
549 	/* Accept either same addr or a new addr returned from mmap if target
550 	 * range occupied.
551 	 */
552 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
553 		dev->data->port_id, addr);
554 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
555 	uar_base = addr; /* process local, don't reserve again. */
556 	return 0;
557 }
558 
559 /**
560  * Reserve UAR address space for secondary process, align with
561  * primary process.
562  *
563  * @param[in] dev
564  *   Pointer to Ethernet device.
565  *
566  * @return
567  *   0 on success, a negative errno value otherwise and rte_errno is set.
568  */
569 static int
570 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
571 {
572 	struct priv *priv = dev->data->dev_private;
573 	void *addr;
574 
575 	assert(priv->uar_base);
576 	if (uar_base) { /* already reserved. */
577 		assert(uar_base == priv->uar_base);
578 		return 0;
579 	}
580 	/* anonymous mmap, no real memory consumption. */
581 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
582 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
583 	if (addr == MAP_FAILED) {
584 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
585 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
586 		rte_errno = ENXIO;
587 		return -rte_errno;
588 	}
589 	if (priv->uar_base != addr) {
590 		DRV_LOG(ERR,
591 			"port %u UAR address %p size %llu occupied, please"
592 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
593 			" --base-virtaddr",
594 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
595 		rte_errno = ENXIO;
596 		return -rte_errno;
597 	}
598 	uar_base = addr; /* process local, don't reserve again */
599 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
600 		dev->data->port_id, addr);
601 	return 0;
602 }
603 
604 /**
605  * DPDK callback to register a PCI device.
606  *
607  * This function creates an Ethernet device for each port of a given
608  * PCI device.
609  *
610  * @param[in] pci_drv
611  *   PCI driver structure (mlx5_driver).
612  * @param[in] pci_dev
613  *   PCI device information.
614  *
615  * @return
616  *   0 on success, a negative errno value otherwise and rte_errno is set.
617  */
618 static int
619 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
620 	       struct rte_pci_device *pci_dev)
621 {
622 	struct ibv_device **list = NULL;
623 	struct ibv_device *ibv_dev;
624 	int err = 0;
625 	struct ibv_context *attr_ctx = NULL;
626 	struct ibv_device_attr_ex device_attr;
627 	unsigned int vf;
628 	unsigned int mps;
629 	unsigned int cqe_comp;
630 	unsigned int tunnel_en = 0;
631 	unsigned int swp = 0;
632 	unsigned int verb_priorities = 0;
633 	int idx;
634 	int i;
635 	struct mlx5dv_context attrs_out = {0};
636 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
637 	struct ibv_counter_set_description cs_desc;
638 #endif
639 
640 	assert(pci_drv == &mlx5_driver);
641 	/* Get mlx5_dev[] index. */
642 	idx = mlx5_dev_idx(&pci_dev->addr);
643 	if (idx == -1) {
644 		DRV_LOG(ERR, "this driver cannot support any more adapters");
645 		err = ENOMEM;
646 		goto error;
647 	}
648 	DRV_LOG(DEBUG, "using driver device index %d", idx);
649 	/* Save PCI address. */
650 	mlx5_dev[idx].pci_addr = pci_dev->addr;
651 	list = mlx5_glue->get_device_list(&i);
652 	if (list == NULL) {
653 		assert(errno);
654 		err = errno;
655 		if (errno == ENOSYS)
656 			DRV_LOG(ERR,
657 				"cannot list devices, is ib_uverbs loaded?");
658 		goto error;
659 	}
660 	assert(i >= 0);
661 	/*
662 	 * For each listed device, check related sysfs entry against
663 	 * the provided PCI ID.
664 	 */
665 	while (i != 0) {
666 		struct rte_pci_addr pci_addr;
667 
668 		--i;
669 		DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
670 		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
671 			continue;
672 		if ((pci_dev->addr.domain != pci_addr.domain) ||
673 		    (pci_dev->addr.bus != pci_addr.bus) ||
674 		    (pci_dev->addr.devid != pci_addr.devid) ||
675 		    (pci_dev->addr.function != pci_addr.function))
676 			continue;
677 		DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
678 			list[i]->name);
679 		vf = ((pci_dev->id.device_id ==
680 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
681 		      (pci_dev->id.device_id ==
682 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
683 		      (pci_dev->id.device_id ==
684 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
685 		      (pci_dev->id.device_id ==
686 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
687 		attr_ctx = mlx5_glue->open_device(list[i]);
688 		rte_errno = errno;
689 		err = rte_errno;
690 		break;
691 	}
692 	if (attr_ctx == NULL) {
693 		mlx5_glue->free_device_list(list);
694 		switch (err) {
695 		case 0:
696 			DRV_LOG(ERR,
697 				"cannot access device, is mlx5_ib loaded?");
698 			err = ENODEV;
699 			goto error;
700 		case EINVAL:
701 			DRV_LOG(ERR,
702 				"cannot use device, are drivers up to date?");
703 			goto error;
704 		}
705 	}
706 	ibv_dev = list[i];
707 	DRV_LOG(DEBUG, "device opened");
708 #ifdef HAVE_IBV_MLX5_MOD_SWP
709 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
710 #endif
711 	/*
712 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
713 	 * as all ConnectX-5 devices.
714 	 */
715 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
716 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
717 #endif
718 	mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
719 	if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
720 		if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
721 			DRV_LOG(DEBUG, "enhanced MPW is supported");
722 			mps = MLX5_MPW_ENHANCED;
723 		} else {
724 			DRV_LOG(DEBUG, "MPW is supported");
725 			mps = MLX5_MPW;
726 		}
727 	} else {
728 		DRV_LOG(DEBUG, "MPW isn't supported");
729 		mps = MLX5_MPW_DISABLED;
730 	}
731 #ifdef HAVE_IBV_MLX5_MOD_SWP
732 	if (attrs_out.comp_mask | MLX5DV_CONTEXT_MASK_SWP)
733 		swp = attrs_out.sw_parsing_caps.sw_parsing_offloads;
734 	DRV_LOG(DEBUG, "SWP support: %u", swp);
735 #endif
736 	if (RTE_CACHE_LINE_SIZE == 128 &&
737 	    !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
738 		cqe_comp = 0;
739 	else
740 		cqe_comp = 1;
741 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
742 	if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
743 		tunnel_en = ((attrs_out.tunnel_offloads_caps &
744 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
745 			     (attrs_out.tunnel_offloads_caps &
746 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
747 	}
748 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
749 		tunnel_en ? "" : "not ");
750 #else
751 	DRV_LOG(WARNING,
752 		"tunnel offloading disabled due to old OFED/rdma-core version");
753 #endif
754 	if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr)) {
755 		err = errno;
756 		goto error;
757 	}
758 	DRV_LOG(INFO, "%u port(s) detected",
759 		device_attr.orig_attr.phys_port_cnt);
760 	for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
761 		char name[RTE_ETH_NAME_MAX_LEN];
762 		int len;
763 		uint32_t port = i + 1; /* ports are indexed from one */
764 		uint32_t test = (1 << i);
765 		struct ibv_context *ctx = NULL;
766 		struct ibv_port_attr port_attr;
767 		struct ibv_pd *pd = NULL;
768 		struct priv *priv = NULL;
769 		struct rte_eth_dev *eth_dev = NULL;
770 		struct ibv_device_attr_ex device_attr_ex;
771 		struct ether_addr mac;
772 		struct mlx5_dev_config config = {
773 			.cqe_comp = cqe_comp,
774 			.mps = mps,
775 			.tunnel_en = tunnel_en,
776 			.tx_vec_en = 1,
777 			.rx_vec_en = 1,
778 			.mpw_hdr_dseg = 0,
779 			.txq_inline = MLX5_ARG_UNSET,
780 			.txqs_inline = MLX5_ARG_UNSET,
781 			.inline_max_packet_sz = MLX5_ARG_UNSET,
782 			.vf_nl_en = 1,
783 			.swp = !!swp,
784 		};
785 
786 		len = snprintf(name, sizeof(name), PCI_PRI_FMT,
787 			 pci_dev->addr.domain, pci_dev->addr.bus,
788 			 pci_dev->addr.devid, pci_dev->addr.function);
789 		if (device_attr.orig_attr.phys_port_cnt > 1)
790 			snprintf(name + len, sizeof(name), " port %u", i);
791 		mlx5_dev[idx].ports |= test;
792 		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
793 			eth_dev = rte_eth_dev_attach_secondary(name);
794 			if (eth_dev == NULL) {
795 				DRV_LOG(ERR, "can not attach rte ethdev");
796 				rte_errno = ENOMEM;
797 				err = rte_errno;
798 				goto error;
799 			}
800 			eth_dev->device = &pci_dev->device;
801 			eth_dev->dev_ops = &mlx5_dev_sec_ops;
802 			err = mlx5_uar_init_secondary(eth_dev);
803 			if (err)
804 				goto error;
805 			/* Receive command fd from primary process */
806 			err = mlx5_socket_connect(eth_dev);
807 			if (err)
808 				goto error;
809 			/* Remap UAR for Tx queues. */
810 			err = mlx5_tx_uar_remap(eth_dev, err);
811 			if (err)
812 				goto error;
813 			/*
814 			 * Ethdev pointer is still required as input since
815 			 * the primary device is not accessible from the
816 			 * secondary process.
817 			 */
818 			eth_dev->rx_pkt_burst =
819 				mlx5_select_rx_function(eth_dev);
820 			eth_dev->tx_pkt_burst =
821 				mlx5_select_tx_function(eth_dev);
822 			continue;
823 		}
824 		DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
825 		ctx = mlx5_glue->open_device(ibv_dev);
826 		if (ctx == NULL) {
827 			err = ENODEV;
828 			goto port_error;
829 		}
830 		/* Check port status. */
831 		err = mlx5_glue->query_port(ctx, port, &port_attr);
832 		if (err) {
833 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
834 			goto port_error;
835 		}
836 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
837 			DRV_LOG(ERR,
838 				"port %d is not configured in Ethernet mode",
839 				port);
840 			err = EINVAL;
841 			goto port_error;
842 		}
843 		if (port_attr.state != IBV_PORT_ACTIVE)
844 			DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
845 				port,
846 				mlx5_glue->port_state_str(port_attr.state),
847 				port_attr.state);
848 		/* Allocate protection domain. */
849 		pd = mlx5_glue->alloc_pd(ctx);
850 		if (pd == NULL) {
851 			DRV_LOG(ERR, "PD allocation failure");
852 			err = ENOMEM;
853 			goto port_error;
854 		}
855 		mlx5_dev[idx].ports |= test;
856 		/* from rte_ethdev.c */
857 		priv = rte_zmalloc("ethdev private structure",
858 				   sizeof(*priv),
859 				   RTE_CACHE_LINE_SIZE);
860 		if (priv == NULL) {
861 			DRV_LOG(ERR, "priv allocation failure");
862 			err = ENOMEM;
863 			goto port_error;
864 		}
865 		priv->ctx = ctx;
866 		strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
867 			sizeof(priv->ibdev_path));
868 		priv->device_attr = device_attr;
869 		priv->port = port;
870 		priv->pd = pd;
871 		priv->mtu = ETHER_MTU;
872 		err = mlx5_args(&config, pci_dev->device.devargs);
873 		if (err) {
874 			DRV_LOG(ERR, "failed to process device arguments: %s",
875 				strerror(err));
876 			goto port_error;
877 		}
878 		if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
879 			DRV_LOG(ERR, "ibv_query_device_ex() failed");
880 			err = errno;
881 			goto port_error;
882 		}
883 		config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
884 				    IBV_DEVICE_RAW_IP_CSUM);
885 		DRV_LOG(DEBUG, "checksum offloading is %ssupported",
886 			(config.hw_csum ? "" : "not "));
887 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
888 		config.flow_counter_en = !!(device_attr.max_counter_sets);
889 		mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
890 		DRV_LOG(DEBUG,
891 			"counter type = %d, num of cs = %ld, attributes = %d",
892 			cs_desc.counter_type, cs_desc.num_of_cs,
893 			cs_desc.attributes);
894 #endif
895 		config.ind_table_max_size =
896 			device_attr_ex.rss_caps.max_rwq_indirection_table_size;
897 		/* Remove this check once DPDK supports larger/variable
898 		 * indirection tables. */
899 		if (config.ind_table_max_size >
900 				(unsigned int)ETH_RSS_RETA_SIZE_512)
901 			config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
902 		DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
903 			config.ind_table_max_size);
904 		config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
905 					 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
906 		DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
907 			(config.hw_vlan_strip ? "" : "not "));
908 
909 		config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
910 					 IBV_RAW_PACKET_CAP_SCATTER_FCS);
911 		DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
912 			(config.hw_fcs_strip ? "" : "not "));
913 
914 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
915 		config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
916 #endif
917 		DRV_LOG(DEBUG,
918 			"hardware Rx end alignment padding is %ssupported",
919 			(config.hw_padding ? "" : "not "));
920 		config.vf = vf;
921 		config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
922 			      (device_attr_ex.tso_caps.supported_qpts &
923 			      (1 << IBV_QPT_RAW_PACKET)));
924 		if (config.tso)
925 			config.tso_max_payload_sz =
926 					device_attr_ex.tso_caps.max_tso;
927 		if (config.mps && !mps) {
928 			DRV_LOG(ERR,
929 				"multi-packet send not supported on this device"
930 				" (" MLX5_TXQ_MPW_EN ")");
931 			err = ENOTSUP;
932 			goto port_error;
933 		}
934 		DRV_LOG(INFO, "%s MPS is %s",
935 			config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
936 			config.mps != MLX5_MPW_DISABLED ? "enabled" :
937 			"disabled");
938 		if (config.cqe_comp && !cqe_comp) {
939 			DRV_LOG(WARNING, "Rx CQE compression isn't supported");
940 			config.cqe_comp = 0;
941 		}
942 		eth_dev = rte_eth_dev_allocate(name);
943 		if (eth_dev == NULL) {
944 			DRV_LOG(ERR, "can not allocate rte ethdev");
945 			err = ENOMEM;
946 			goto port_error;
947 		}
948 		eth_dev->data->dev_private = priv;
949 		priv->dev = eth_dev;
950 		eth_dev->data->mac_addrs = priv->mac;
951 		eth_dev->device = &pci_dev->device;
952 		rte_eth_copy_pci_info(eth_dev, pci_dev);
953 		eth_dev->device->driver = &mlx5_driver.driver;
954 		err = mlx5_uar_init_primary(eth_dev);
955 		if (err)
956 			goto port_error;
957 		/* Configure the first MAC address by default. */
958 		if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
959 			DRV_LOG(ERR,
960 				"port %u cannot get MAC address, is mlx5_en"
961 				" loaded? (errno: %s)",
962 				eth_dev->data->port_id, strerror(errno));
963 			err = ENODEV;
964 			goto port_error;
965 		}
966 		DRV_LOG(INFO,
967 			"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
968 			eth_dev->data->port_id,
969 			mac.addr_bytes[0], mac.addr_bytes[1],
970 			mac.addr_bytes[2], mac.addr_bytes[3],
971 			mac.addr_bytes[4], mac.addr_bytes[5]);
972 #ifndef NDEBUG
973 		{
974 			char ifname[IF_NAMESIZE];
975 
976 			if (mlx5_get_ifname(eth_dev, &ifname) == 0)
977 				DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
978 					eth_dev->data->port_id, ifname);
979 			else
980 				DRV_LOG(DEBUG, "port %u ifname is unknown",
981 					eth_dev->data->port_id);
982 		}
983 #endif
984 		/* Get actual MTU if possible. */
985 		err = mlx5_get_mtu(eth_dev, &priv->mtu);
986 		if (err)
987 			goto port_error;
988 		DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
989 			priv->mtu);
990 		/*
991 		 * Initialize burst functions to prevent crashes before link-up.
992 		 */
993 		eth_dev->rx_pkt_burst = removed_rx_burst;
994 		eth_dev->tx_pkt_burst = removed_tx_burst;
995 		eth_dev->dev_ops = &mlx5_dev_ops;
996 		/* Register MAC address. */
997 		claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
998 		priv->nl_socket = -1;
999 		priv->nl_sn = 0;
1000 		if (vf && config.vf_nl_en) {
1001 			priv->nl_socket = mlx5_nl_init(RTMGRP_LINK);
1002 			if (priv->nl_socket < 0)
1003 				priv->nl_socket = -1;
1004 			mlx5_nl_mac_addr_sync(eth_dev);
1005 		}
1006 		TAILQ_INIT(&priv->flows);
1007 		TAILQ_INIT(&priv->ctrl_flows);
1008 		/* Hint libmlx5 to use PMD allocator for data plane resources */
1009 		struct mlx5dv_ctx_allocators alctr = {
1010 			.alloc = &mlx5_alloc_verbs_buf,
1011 			.free = &mlx5_free_verbs_buf,
1012 			.data = priv,
1013 		};
1014 		mlx5_glue->dv_set_context_attr(ctx,
1015 					       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1016 					       (void *)((uintptr_t)&alctr));
1017 		/* Bring Ethernet device up. */
1018 		DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1019 			eth_dev->data->port_id);
1020 		mlx5_set_link_up(eth_dev);
1021 		/*
1022 		 * Even though the interrupt handler is not installed yet,
1023 		 * interrupts will still trigger on the asyn_fd from
1024 		 * Verbs context returned by ibv_open_device().
1025 		 */
1026 		mlx5_link_update(eth_dev, 0);
1027 		/* Store device configuration on private structure. */
1028 		priv->config = config;
1029 		/* Create drop queue. */
1030 		err = mlx5_flow_create_drop_queue(eth_dev);
1031 		if (err) {
1032 			DRV_LOG(ERR, "port %u drop queue allocation failed: %s",
1033 				eth_dev->data->port_id, strerror(rte_errno));
1034 			goto port_error;
1035 		}
1036 		/* Supported Verbs flow priority number detection. */
1037 		if (verb_priorities == 0)
1038 			verb_priorities = mlx5_get_max_verbs_prio(eth_dev);
1039 		if (verb_priorities < MLX5_VERBS_FLOW_PRIO_8) {
1040 			DRV_LOG(ERR, "port %u wrong Verbs flow priorities: %u",
1041 				eth_dev->data->port_id, verb_priorities);
1042 			goto port_error;
1043 		}
1044 		priv->config.max_verbs_prio = verb_priorities;
1045 		continue;
1046 port_error:
1047 		if (priv)
1048 			rte_free(priv);
1049 		if (pd)
1050 			claim_zero(mlx5_glue->dealloc_pd(pd));
1051 		if (ctx)
1052 			claim_zero(mlx5_glue->close_device(ctx));
1053 		break;
1054 	}
1055 	/*
1056 	 * XXX if something went wrong in the loop above, there is a resource
1057 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1058 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
1059 	 * way to enumerate the registered ethdevs to free the previous ones.
1060 	 */
1061 	/* no port found, complain */
1062 	if (!mlx5_dev[idx].ports) {
1063 		rte_errno = ENODEV;
1064 		err = rte_errno;
1065 	}
1066 error:
1067 	if (attr_ctx)
1068 		claim_zero(mlx5_glue->close_device(attr_ctx));
1069 	if (list)
1070 		mlx5_glue->free_device_list(list);
1071 	if (err) {
1072 		rte_errno = err;
1073 		return -rte_errno;
1074 	}
1075 	return 0;
1076 }
1077 
1078 static const struct rte_pci_id mlx5_pci_id_map[] = {
1079 	{
1080 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1081 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1082 	},
1083 	{
1084 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1085 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1086 	},
1087 	{
1088 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1089 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1090 	},
1091 	{
1092 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1093 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1094 	},
1095 	{
1096 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1097 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1098 	},
1099 	{
1100 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1101 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1102 	},
1103 	{
1104 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1105 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1106 	},
1107 	{
1108 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1109 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1110 	},
1111 	{
1112 		.vendor_id = 0
1113 	}
1114 };
1115 
1116 static struct rte_pci_driver mlx5_driver = {
1117 	.driver = {
1118 		.name = MLX5_DRIVER_NAME
1119 	},
1120 	.id_table = mlx5_pci_id_map,
1121 	.probe = mlx5_pci_probe,
1122 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1123 };
1124 
1125 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1126 
1127 /**
1128  * Suffix RTE_EAL_PMD_PATH with "-glue".
1129  *
1130  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1131  * suffixing its last component.
1132  *
1133  * @param buf[out]
1134  *   Output buffer, should be large enough otherwise NULL is returned.
1135  * @param size
1136  *   Size of @p out.
1137  *
1138  * @return
1139  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1140  */
1141 static char *
1142 mlx5_glue_path(char *buf, size_t size)
1143 {
1144 	static const char *const bad[] = { "/", ".", "..", NULL };
1145 	const char *path = RTE_EAL_PMD_PATH;
1146 	size_t len = strlen(path);
1147 	size_t off;
1148 	int i;
1149 
1150 	while (len && path[len - 1] == '/')
1151 		--len;
1152 	for (off = len; off && path[off - 1] != '/'; --off)
1153 		;
1154 	for (i = 0; bad[i]; ++i)
1155 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1156 			goto error;
1157 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1158 	if (i == -1 || (size_t)i >= size)
1159 		goto error;
1160 	return buf;
1161 error:
1162 	DRV_LOG(ERR,
1163 		"unable to append \"-glue\" to last component of"
1164 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1165 		" please re-configure DPDK");
1166 	return NULL;
1167 }
1168 
1169 /**
1170  * Initialization routine for run-time dependency on rdma-core.
1171  */
1172 static int
1173 mlx5_glue_init(void)
1174 {
1175 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1176 	const char *path[] = {
1177 		/*
1178 		 * A basic security check is necessary before trusting
1179 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1180 		 */
1181 		(geteuid() == getuid() && getegid() == getgid() ?
1182 		 getenv("MLX5_GLUE_PATH") : NULL),
1183 		/*
1184 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1185 		 * variant, otherwise let dlopen() look up libraries on its
1186 		 * own.
1187 		 */
1188 		(*RTE_EAL_PMD_PATH ?
1189 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1190 	};
1191 	unsigned int i = 0;
1192 	void *handle = NULL;
1193 	void **sym;
1194 	const char *dlmsg;
1195 
1196 	while (!handle && i != RTE_DIM(path)) {
1197 		const char *end;
1198 		size_t len;
1199 		int ret;
1200 
1201 		if (!path[i]) {
1202 			++i;
1203 			continue;
1204 		}
1205 		end = strpbrk(path[i], ":;");
1206 		if (!end)
1207 			end = path[i] + strlen(path[i]);
1208 		len = end - path[i];
1209 		ret = 0;
1210 		do {
1211 			char name[ret + 1];
1212 
1213 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1214 				       (int)len, path[i],
1215 				       (!len || *(end - 1) == '/') ? "" : "/");
1216 			if (ret == -1)
1217 				break;
1218 			if (sizeof(name) != (size_t)ret + 1)
1219 				continue;
1220 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1221 				name);
1222 			handle = dlopen(name, RTLD_LAZY);
1223 			break;
1224 		} while (1);
1225 		path[i] = end + 1;
1226 		if (!*end)
1227 			++i;
1228 	}
1229 	if (!handle) {
1230 		rte_errno = EINVAL;
1231 		dlmsg = dlerror();
1232 		if (dlmsg)
1233 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1234 		goto glue_error;
1235 	}
1236 	sym = dlsym(handle, "mlx5_glue");
1237 	if (!sym || !*sym) {
1238 		rte_errno = EINVAL;
1239 		dlmsg = dlerror();
1240 		if (dlmsg)
1241 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1242 		goto glue_error;
1243 	}
1244 	mlx5_glue = *sym;
1245 	return 0;
1246 glue_error:
1247 	if (handle)
1248 		dlclose(handle);
1249 	DRV_LOG(WARNING,
1250 		"cannot initialize PMD due to missing run-time dependency on"
1251 		" rdma-core libraries (libibverbs, libmlx5)");
1252 	return -rte_errno;
1253 }
1254 
1255 #endif
1256 
1257 /**
1258  * Driver initialization routine.
1259  */
1260 RTE_INIT(rte_mlx5_pmd_init);
1261 static void
1262 rte_mlx5_pmd_init(void)
1263 {
1264 	/* Build the static tables for Verbs conversion. */
1265 	mlx5_set_ptype_table();
1266 	mlx5_set_cksum_table();
1267 	mlx5_set_swp_types_table();
1268 	/*
1269 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1270 	 * huge pages. Calling ibv_fork_init() during init allows
1271 	 * applications to use fork() safely for purposes other than
1272 	 * using this PMD, which is not supported in forked processes.
1273 	 */
1274 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1275 	/* Match the size of Rx completion entry to the size of a cacheline. */
1276 	if (RTE_CACHE_LINE_SIZE == 128)
1277 		setenv("MLX5_CQE_SIZE", "128", 0);
1278 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1279 	if (mlx5_glue_init())
1280 		return;
1281 	assert(mlx5_glue);
1282 #endif
1283 #ifndef NDEBUG
1284 	/* Glue structure must not contain any NULL pointers. */
1285 	{
1286 		unsigned int i;
1287 
1288 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1289 			assert(((const void *const *)mlx5_glue)[i]);
1290 	}
1291 #endif
1292 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1293 		DRV_LOG(ERR,
1294 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1295 			mlx5_glue->version, MLX5_GLUE_VERSION);
1296 		return;
1297 	}
1298 	mlx5_glue->fork_init();
1299 	rte_pci_register(&mlx5_driver);
1300 }
1301 
1302 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1303 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1304 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1305 
1306 /** Initialize driver log type. */
1307 RTE_INIT(vdev_netvsc_init_log)
1308 {
1309 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1310 	if (mlx5_logtype >= 0)
1311 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1312 }
1313