xref: /dpdk/drivers/net/mlx5/mlx5.c (revision b733c60f68f12e064359b27e630305c541a3fbdf)
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/netlink.h>
17 #include <linux/rtnetlink.h>
18 
19 /* Verbs header. */
20 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
21 #ifdef PEDANTIC
22 #pragma GCC diagnostic ignored "-Wpedantic"
23 #endif
24 #include <infiniband/verbs.h>
25 #ifdef PEDANTIC
26 #pragma GCC diagnostic error "-Wpedantic"
27 #endif
28 
29 #include <rte_malloc.h>
30 #include <rte_ethdev_driver.h>
31 #include <rte_ethdev_pci.h>
32 #include <rte_pci.h>
33 #include <rte_bus_pci.h>
34 #include <rte_common.h>
35 #include <rte_config.h>
36 #include <rte_eal_memconfig.h>
37 #include <rte_kvargs.h>
38 #include <rte_rwlock.h>
39 #include <rte_spinlock.h>
40 #include <rte_string_fns.h>
41 
42 #include "mlx5.h"
43 #include "mlx5_utils.h"
44 #include "mlx5_rxtx.h"
45 #include "mlx5_autoconf.h"
46 #include "mlx5_defs.h"
47 #include "mlx5_glue.h"
48 #include "mlx5_mr.h"
49 
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52 
53 /* Device parameter to enable Multi-Packet Rx queue. */
54 #define MLX5_RX_MPRQ_EN "mprq_en"
55 
56 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
57 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
58 
59 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
60 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
61 
62 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
63 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
64 
65 /* Device parameter to configure inline send. */
66 #define MLX5_TXQ_INLINE "txq_inline"
67 
68 /*
69  * Device parameter to configure the number of TX queues threshold for
70  * enabling inline send.
71  */
72 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
73 
74 /* Device parameter to enable multi-packet send WQEs. */
75 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
76 
77 /* Device parameter to include 2 dsegs in the title WQEBB. */
78 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
79 
80 /* Device parameter to limit the size of inlining packet. */
81 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
82 
83 /* Device parameter to enable hardware Tx vector. */
84 #define MLX5_TX_VEC_EN "tx_vec_en"
85 
86 /* Device parameter to enable hardware Rx vector. */
87 #define MLX5_RX_VEC_EN "rx_vec_en"
88 
89 /* Allow L3 VXLAN flow creation. */
90 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
91 
92 /* Activate Netlink support in VF mode. */
93 #define MLX5_VF_NL_EN "vf_nl_en"
94 
95 /* Select port representors to instantiate. */
96 #define MLX5_REPRESENTOR "representor"
97 
98 #ifndef HAVE_IBV_MLX5_MOD_MPW
99 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
100 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
101 #endif
102 
103 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
104 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
105 #endif
106 
107 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
108 
109 /* Shared memory between primary and secondary processes. */
110 struct mlx5_shared_data *mlx5_shared_data;
111 
112 /* Spinlock for mlx5_shared_data allocation. */
113 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
114 
115 /** Driver-specific log messages type. */
116 int mlx5_logtype;
117 
118 /**
119  * Prepare shared data between primary and secondary process.
120  */
121 static void
122 mlx5_prepare_shared_data(void)
123 {
124 	const struct rte_memzone *mz;
125 
126 	rte_spinlock_lock(&mlx5_shared_data_lock);
127 	if (mlx5_shared_data == NULL) {
128 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
129 			/* Allocate shared memory. */
130 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
131 						 sizeof(*mlx5_shared_data),
132 						 SOCKET_ID_ANY, 0);
133 		} else {
134 			/* Lookup allocated shared memory. */
135 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
136 		}
137 		if (mz == NULL)
138 			rte_panic("Cannot allocate mlx5 shared data\n");
139 		mlx5_shared_data = mz->addr;
140 		/* Initialize shared data. */
141 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
142 			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
143 			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
144 		}
145 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
146 						mlx5_mr_mem_event_cb, NULL);
147 	}
148 	rte_spinlock_unlock(&mlx5_shared_data_lock);
149 }
150 
151 /**
152  * Retrieve integer value from environment variable.
153  *
154  * @param[in] name
155  *   Environment variable name.
156  *
157  * @return
158  *   Integer value, 0 if the variable is not set.
159  */
160 int
161 mlx5_getenv_int(const char *name)
162 {
163 	const char *val = getenv(name);
164 
165 	if (val == NULL)
166 		return 0;
167 	return atoi(val);
168 }
169 
170 /**
171  * Verbs callback to allocate a memory. This function should allocate the space
172  * according to the size provided residing inside a huge page.
173  * Please note that all allocation must respect the alignment from libmlx5
174  * (i.e. currently sysconf(_SC_PAGESIZE)).
175  *
176  * @param[in] size
177  *   The size in bytes of the memory to allocate.
178  * @param[in] data
179  *   A pointer to the callback data.
180  *
181  * @return
182  *   Allocated buffer, NULL otherwise and rte_errno is set.
183  */
184 static void *
185 mlx5_alloc_verbs_buf(size_t size, void *data)
186 {
187 	struct priv *priv = data;
188 	void *ret;
189 	size_t alignment = sysconf(_SC_PAGESIZE);
190 	unsigned int socket = SOCKET_ID_ANY;
191 
192 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
193 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
194 
195 		socket = ctrl->socket;
196 	} else if (priv->verbs_alloc_ctx.type ==
197 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
198 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
199 
200 		socket = ctrl->socket;
201 	}
202 	assert(data != NULL);
203 	ret = rte_malloc_socket(__func__, size, alignment, socket);
204 	if (!ret && size)
205 		rte_errno = ENOMEM;
206 	return ret;
207 }
208 
209 /**
210  * Verbs callback to free a memory.
211  *
212  * @param[in] ptr
213  *   A pointer to the memory to free.
214  * @param[in] data
215  *   A pointer to the callback data.
216  */
217 static void
218 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
219 {
220 	assert(data != NULL);
221 	rte_free(ptr);
222 }
223 
224 /**
225  * DPDK callback to close the device.
226  *
227  * Destroy all queues and objects, free memory.
228  *
229  * @param dev
230  *   Pointer to Ethernet device structure.
231  */
232 static void
233 mlx5_dev_close(struct rte_eth_dev *dev)
234 {
235 	struct priv *priv = dev->data->dev_private;
236 	unsigned int i;
237 	int ret;
238 
239 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
240 		dev->data->port_id,
241 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
242 	/* In case mlx5_dev_stop() has not been called. */
243 	mlx5_dev_interrupt_handler_uninstall(dev);
244 	mlx5_traffic_disable(dev);
245 	mlx5_flow_flush(dev, NULL);
246 	/* Prevent crashes when queues are still in use. */
247 	dev->rx_pkt_burst = removed_rx_burst;
248 	dev->tx_pkt_burst = removed_tx_burst;
249 	if (priv->rxqs != NULL) {
250 		/* XXX race condition if mlx5_rx_burst() is still running. */
251 		usleep(1000);
252 		for (i = 0; (i != priv->rxqs_n); ++i)
253 			mlx5_rxq_release(dev, i);
254 		priv->rxqs_n = 0;
255 		priv->rxqs = NULL;
256 	}
257 	if (priv->txqs != NULL) {
258 		/* XXX race condition if mlx5_tx_burst() is still running. */
259 		usleep(1000);
260 		for (i = 0; (i != priv->txqs_n); ++i)
261 			mlx5_txq_release(dev, i);
262 		priv->txqs_n = 0;
263 		priv->txqs = NULL;
264 	}
265 	mlx5_mprq_free_mp(dev);
266 	mlx5_mr_release(dev);
267 	if (priv->pd != NULL) {
268 		assert(priv->ctx != NULL);
269 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
270 		claim_zero(mlx5_glue->close_device(priv->ctx));
271 	} else
272 		assert(priv->ctx == NULL);
273 	if (priv->rss_conf.rss_key != NULL)
274 		rte_free(priv->rss_conf.rss_key);
275 	if (priv->reta_idx != NULL)
276 		rte_free(priv->reta_idx);
277 	if (priv->primary_socket)
278 		mlx5_socket_uninit(dev);
279 	if (priv->config.vf)
280 		mlx5_nl_mac_addr_flush(dev);
281 	if (priv->nl_socket_route >= 0)
282 		close(priv->nl_socket_route);
283 	if (priv->nl_socket_rdma >= 0)
284 		close(priv->nl_socket_rdma);
285 	if (priv->mnl_socket)
286 		mlx5_nl_flow_socket_destroy(priv->mnl_socket);
287 	ret = mlx5_hrxq_ibv_verify(dev);
288 	if (ret)
289 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
290 			dev->data->port_id);
291 	ret = mlx5_ind_table_ibv_verify(dev);
292 	if (ret)
293 		DRV_LOG(WARNING, "port %u some indirection table still remain",
294 			dev->data->port_id);
295 	ret = mlx5_rxq_ibv_verify(dev);
296 	if (ret)
297 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
298 			dev->data->port_id);
299 	ret = mlx5_rxq_verify(dev);
300 	if (ret)
301 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
302 			dev->data->port_id);
303 	ret = mlx5_txq_ibv_verify(dev);
304 	if (ret)
305 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
306 			dev->data->port_id);
307 	ret = mlx5_txq_verify(dev);
308 	if (ret)
309 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
310 			dev->data->port_id);
311 	ret = mlx5_flow_verify(dev);
312 	if (ret)
313 		DRV_LOG(WARNING, "port %u some flows still remain",
314 			dev->data->port_id);
315 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
316 		unsigned int c = 0;
317 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
318 		uint16_t port_id[i];
319 
320 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
321 		while (i--) {
322 			struct priv *opriv =
323 				rte_eth_devices[port_id[i]].data->dev_private;
324 
325 			if (!opriv ||
326 			    opriv->domain_id != priv->domain_id ||
327 			    &rte_eth_devices[port_id[i]] == dev)
328 				continue;
329 			++c;
330 		}
331 		if (!c)
332 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
333 	}
334 	memset(priv, 0, sizeof(*priv));
335 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
336 }
337 
338 const struct eth_dev_ops mlx5_dev_ops = {
339 	.dev_configure = mlx5_dev_configure,
340 	.dev_start = mlx5_dev_start,
341 	.dev_stop = mlx5_dev_stop,
342 	.dev_set_link_down = mlx5_set_link_down,
343 	.dev_set_link_up = mlx5_set_link_up,
344 	.dev_close = mlx5_dev_close,
345 	.promiscuous_enable = mlx5_promiscuous_enable,
346 	.promiscuous_disable = mlx5_promiscuous_disable,
347 	.allmulticast_enable = mlx5_allmulticast_enable,
348 	.allmulticast_disable = mlx5_allmulticast_disable,
349 	.link_update = mlx5_link_update,
350 	.stats_get = mlx5_stats_get,
351 	.stats_reset = mlx5_stats_reset,
352 	.xstats_get = mlx5_xstats_get,
353 	.xstats_reset = mlx5_xstats_reset,
354 	.xstats_get_names = mlx5_xstats_get_names,
355 	.dev_infos_get = mlx5_dev_infos_get,
356 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
357 	.vlan_filter_set = mlx5_vlan_filter_set,
358 	.rx_queue_setup = mlx5_rx_queue_setup,
359 	.tx_queue_setup = mlx5_tx_queue_setup,
360 	.rx_queue_release = mlx5_rx_queue_release,
361 	.tx_queue_release = mlx5_tx_queue_release,
362 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
363 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
364 	.mac_addr_remove = mlx5_mac_addr_remove,
365 	.mac_addr_add = mlx5_mac_addr_add,
366 	.mac_addr_set = mlx5_mac_addr_set,
367 	.set_mc_addr_list = mlx5_set_mc_addr_list,
368 	.mtu_set = mlx5_dev_set_mtu,
369 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
370 	.vlan_offload_set = mlx5_vlan_offload_set,
371 	.reta_update = mlx5_dev_rss_reta_update,
372 	.reta_query = mlx5_dev_rss_reta_query,
373 	.rss_hash_update = mlx5_rss_hash_update,
374 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
375 	.filter_ctrl = mlx5_dev_filter_ctrl,
376 	.rx_descriptor_status = mlx5_rx_descriptor_status,
377 	.tx_descriptor_status = mlx5_tx_descriptor_status,
378 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
379 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
380 	.is_removed = mlx5_is_removed,
381 };
382 
383 static const struct eth_dev_ops mlx5_dev_sec_ops = {
384 	.stats_get = mlx5_stats_get,
385 	.stats_reset = mlx5_stats_reset,
386 	.xstats_get = mlx5_xstats_get,
387 	.xstats_reset = mlx5_xstats_reset,
388 	.xstats_get_names = mlx5_xstats_get_names,
389 	.dev_infos_get = mlx5_dev_infos_get,
390 	.rx_descriptor_status = mlx5_rx_descriptor_status,
391 	.tx_descriptor_status = mlx5_tx_descriptor_status,
392 };
393 
394 /* Available operators in flow isolated mode. */
395 const struct eth_dev_ops mlx5_dev_ops_isolate = {
396 	.dev_configure = mlx5_dev_configure,
397 	.dev_start = mlx5_dev_start,
398 	.dev_stop = mlx5_dev_stop,
399 	.dev_set_link_down = mlx5_set_link_down,
400 	.dev_set_link_up = mlx5_set_link_up,
401 	.dev_close = mlx5_dev_close,
402 	.promiscuous_enable = mlx5_promiscuous_enable,
403 	.promiscuous_disable = mlx5_promiscuous_disable,
404 	.allmulticast_enable = mlx5_allmulticast_enable,
405 	.allmulticast_disable = mlx5_allmulticast_disable,
406 	.link_update = mlx5_link_update,
407 	.stats_get = mlx5_stats_get,
408 	.stats_reset = mlx5_stats_reset,
409 	.xstats_get = mlx5_xstats_get,
410 	.xstats_reset = mlx5_xstats_reset,
411 	.xstats_get_names = mlx5_xstats_get_names,
412 	.dev_infos_get = mlx5_dev_infos_get,
413 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
414 	.vlan_filter_set = mlx5_vlan_filter_set,
415 	.rx_queue_setup = mlx5_rx_queue_setup,
416 	.tx_queue_setup = mlx5_tx_queue_setup,
417 	.rx_queue_release = mlx5_rx_queue_release,
418 	.tx_queue_release = mlx5_tx_queue_release,
419 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
420 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
421 	.mac_addr_remove = mlx5_mac_addr_remove,
422 	.mac_addr_add = mlx5_mac_addr_add,
423 	.mac_addr_set = mlx5_mac_addr_set,
424 	.set_mc_addr_list = mlx5_set_mc_addr_list,
425 	.mtu_set = mlx5_dev_set_mtu,
426 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
427 	.vlan_offload_set = mlx5_vlan_offload_set,
428 	.filter_ctrl = mlx5_dev_filter_ctrl,
429 	.rx_descriptor_status = mlx5_rx_descriptor_status,
430 	.tx_descriptor_status = mlx5_tx_descriptor_status,
431 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
432 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
433 	.is_removed = mlx5_is_removed,
434 };
435 
436 /**
437  * Verify and store value for device argument.
438  *
439  * @param[in] key
440  *   Key argument to verify.
441  * @param[in] val
442  *   Value associated with key.
443  * @param opaque
444  *   User data.
445  *
446  * @return
447  *   0 on success, a negative errno value otherwise and rte_errno is set.
448  */
449 static int
450 mlx5_args_check(const char *key, const char *val, void *opaque)
451 {
452 	struct mlx5_dev_config *config = opaque;
453 	unsigned long tmp;
454 
455 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
456 	if (!strcmp(MLX5_REPRESENTOR, key))
457 		return 0;
458 	errno = 0;
459 	tmp = strtoul(val, NULL, 0);
460 	if (errno) {
461 		rte_errno = errno;
462 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
463 		return -rte_errno;
464 	}
465 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
466 		config->cqe_comp = !!tmp;
467 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
468 		config->mprq.enabled = !!tmp;
469 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
470 		config->mprq.stride_num_n = tmp;
471 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
472 		config->mprq.max_memcpy_len = tmp;
473 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
474 		config->mprq.min_rxqs_num = tmp;
475 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
476 		config->txq_inline = tmp;
477 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
478 		config->txqs_inline = tmp;
479 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
480 		config->mps = !!tmp;
481 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
482 		config->mpw_hdr_dseg = !!tmp;
483 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
484 		config->inline_max_packet_sz = tmp;
485 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
486 		config->tx_vec_en = !!tmp;
487 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
488 		config->rx_vec_en = !!tmp;
489 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
490 		config->l3_vxlan_en = !!tmp;
491 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
492 		config->vf_nl_en = !!tmp;
493 	} else {
494 		DRV_LOG(WARNING, "%s: unknown parameter", key);
495 		rte_errno = EINVAL;
496 		return -rte_errno;
497 	}
498 	return 0;
499 }
500 
501 /**
502  * Parse device parameters.
503  *
504  * @param config
505  *   Pointer to device configuration structure.
506  * @param devargs
507  *   Device arguments structure.
508  *
509  * @return
510  *   0 on success, a negative errno value otherwise and rte_errno is set.
511  */
512 static int
513 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
514 {
515 	const char **params = (const char *[]){
516 		MLX5_RXQ_CQE_COMP_EN,
517 		MLX5_RX_MPRQ_EN,
518 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
519 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
520 		MLX5_RXQS_MIN_MPRQ,
521 		MLX5_TXQ_INLINE,
522 		MLX5_TXQS_MIN_INLINE,
523 		MLX5_TXQ_MPW_EN,
524 		MLX5_TXQ_MPW_HDR_DSEG_EN,
525 		MLX5_TXQ_MAX_INLINE_LEN,
526 		MLX5_TX_VEC_EN,
527 		MLX5_RX_VEC_EN,
528 		MLX5_L3_VXLAN_EN,
529 		MLX5_VF_NL_EN,
530 		MLX5_REPRESENTOR,
531 		NULL,
532 	};
533 	struct rte_kvargs *kvlist;
534 	int ret = 0;
535 	int i;
536 
537 	if (devargs == NULL)
538 		return 0;
539 	/* Following UGLY cast is done to pass checkpatch. */
540 	kvlist = rte_kvargs_parse(devargs->args, params);
541 	if (kvlist == NULL)
542 		return 0;
543 	/* Process parameters. */
544 	for (i = 0; (params[i] != NULL); ++i) {
545 		if (rte_kvargs_count(kvlist, params[i])) {
546 			ret = rte_kvargs_process(kvlist, params[i],
547 						 mlx5_args_check, config);
548 			if (ret) {
549 				rte_errno = EINVAL;
550 				rte_kvargs_free(kvlist);
551 				return -rte_errno;
552 			}
553 		}
554 	}
555 	rte_kvargs_free(kvlist);
556 	return 0;
557 }
558 
559 static struct rte_pci_driver mlx5_driver;
560 
561 /*
562  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
563  * local resource used by both primary and secondary to avoid duplicate
564  * reservation.
565  * The space has to be available on both primary and secondary process,
566  * TXQ UAR maps to this area using fixed mmap w/o double check.
567  */
568 static void *uar_base;
569 
570 static int
571 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
572 		const struct rte_memseg *ms, void *arg)
573 {
574 	void **addr = arg;
575 
576 	if (*addr == NULL)
577 		*addr = ms->addr;
578 	else
579 		*addr = RTE_MIN(*addr, ms->addr);
580 
581 	return 0;
582 }
583 
584 /**
585  * Reserve UAR address space for primary process.
586  *
587  * @param[in] dev
588  *   Pointer to Ethernet device.
589  *
590  * @return
591  *   0 on success, a negative errno value otherwise and rte_errno is set.
592  */
593 static int
594 mlx5_uar_init_primary(struct rte_eth_dev *dev)
595 {
596 	struct priv *priv = dev->data->dev_private;
597 	void *addr = (void *)0;
598 
599 	if (uar_base) { /* UAR address space mapped. */
600 		priv->uar_base = uar_base;
601 		return 0;
602 	}
603 	/* find out lower bound of hugepage segments */
604 	rte_memseg_walk(find_lower_va_bound, &addr);
605 
606 	/* keep distance to hugepages to minimize potential conflicts. */
607 	addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
608 	/* anonymous mmap, no real memory consumption. */
609 	addr = mmap(addr, MLX5_UAR_SIZE,
610 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
611 	if (addr == MAP_FAILED) {
612 		DRV_LOG(ERR,
613 			"port %u failed to reserve UAR address space, please"
614 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
615 			dev->data->port_id);
616 		rte_errno = ENOMEM;
617 		return -rte_errno;
618 	}
619 	/* Accept either same addr or a new addr returned from mmap if target
620 	 * range occupied.
621 	 */
622 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
623 		dev->data->port_id, addr);
624 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
625 	uar_base = addr; /* process local, don't reserve again. */
626 	return 0;
627 }
628 
629 /**
630  * Reserve UAR address space for secondary process, align with
631  * primary process.
632  *
633  * @param[in] dev
634  *   Pointer to Ethernet device.
635  *
636  * @return
637  *   0 on success, a negative errno value otherwise and rte_errno is set.
638  */
639 static int
640 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
641 {
642 	struct priv *priv = dev->data->dev_private;
643 	void *addr;
644 
645 	assert(priv->uar_base);
646 	if (uar_base) { /* already reserved. */
647 		assert(uar_base == priv->uar_base);
648 		return 0;
649 	}
650 	/* anonymous mmap, no real memory consumption. */
651 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
652 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
653 	if (addr == MAP_FAILED) {
654 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
655 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
656 		rte_errno = ENXIO;
657 		return -rte_errno;
658 	}
659 	if (priv->uar_base != addr) {
660 		DRV_LOG(ERR,
661 			"port %u UAR address %p size %llu occupied, please"
662 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
663 			" --base-virtaddr",
664 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
665 		rte_errno = ENXIO;
666 		return -rte_errno;
667 	}
668 	uar_base = addr; /* process local, don't reserve again */
669 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
670 		dev->data->port_id, addr);
671 	return 0;
672 }
673 
674 /**
675  * Spawn an Ethernet device from Verbs information.
676  *
677  * @param dpdk_dev
678  *   Backing DPDK device.
679  * @param ibv_dev
680  *   Verbs device.
681  * @param vf
682  *   If nonzero, enable VF-specific features.
683  * @param[in] switch_info
684  *   Switch properties of Ethernet device.
685  *
686  * @return
687  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
688  *   is set. The following error is defined:
689  *
690  *   EBUSY: device is not supposed to be spawned.
691  */
692 static struct rte_eth_dev *
693 mlx5_dev_spawn(struct rte_device *dpdk_dev,
694 	       struct ibv_device *ibv_dev,
695 	       int vf,
696 	       const struct mlx5_switch_info *switch_info)
697 {
698 	struct ibv_context *ctx;
699 	struct ibv_device_attr_ex attr;
700 	struct ibv_port_attr port_attr;
701 	struct ibv_pd *pd = NULL;
702 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
703 	struct mlx5_dev_config config = {
704 		.vf = !!vf,
705 		.mps = MLX5_ARG_UNSET,
706 		.tx_vec_en = 1,
707 		.rx_vec_en = 1,
708 		.mpw_hdr_dseg = 0,
709 		.txq_inline = MLX5_ARG_UNSET,
710 		.txqs_inline = MLX5_ARG_UNSET,
711 		.inline_max_packet_sz = MLX5_ARG_UNSET,
712 		.vf_nl_en = 1,
713 		.mprq = {
714 			.enabled = 0,
715 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
716 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
717 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
718 		},
719 	};
720 	struct rte_eth_dev *eth_dev = NULL;
721 	struct priv *priv = NULL;
722 	int err = 0;
723 	unsigned int mps;
724 	unsigned int cqe_comp;
725 	unsigned int tunnel_en = 0;
726 	unsigned int mpls_en = 0;
727 	unsigned int swp = 0;
728 	unsigned int mprq = 0;
729 	unsigned int mprq_min_stride_size_n = 0;
730 	unsigned int mprq_max_stride_size_n = 0;
731 	unsigned int mprq_min_stride_num_n = 0;
732 	unsigned int mprq_max_stride_num_n = 0;
733 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
734 	struct ibv_counter_set_description cs_desc = { .counter_type = 0 };
735 #endif
736 	struct ether_addr mac;
737 	char name[RTE_ETH_NAME_MAX_LEN];
738 	int own_domain_id = 0;
739 	unsigned int i;
740 
741 	/* Determine if this port representor is supposed to be spawned. */
742 	if (switch_info->representor && dpdk_dev->devargs) {
743 		struct rte_eth_devargs eth_da;
744 
745 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
746 		if (err) {
747 			rte_errno = -err;
748 			DRV_LOG(ERR, "failed to process device arguments: %s",
749 				strerror(rte_errno));
750 			return NULL;
751 		}
752 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
753 			if (eth_da.representor_ports[i] ==
754 			    (uint16_t)switch_info->port_name)
755 				break;
756 		if (i == eth_da.nb_representor_ports) {
757 			rte_errno = EBUSY;
758 			return NULL;
759 		}
760 	}
761 	/* Prepare shared data between primary and secondary process. */
762 	mlx5_prepare_shared_data();
763 	errno = 0;
764 	ctx = mlx5_glue->open_device(ibv_dev);
765 	if (!ctx) {
766 		rte_errno = errno ? errno : ENODEV;
767 		return NULL;
768 	}
769 #ifdef HAVE_IBV_MLX5_MOD_SWP
770 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
771 #endif
772 	/*
773 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
774 	 * as all ConnectX-5 devices.
775 	 */
776 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
777 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
778 #endif
779 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
780 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
781 #endif
782 	mlx5_glue->dv_query_device(ctx, &dv_attr);
783 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
784 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
785 			DRV_LOG(DEBUG, "enhanced MPW is supported");
786 			mps = MLX5_MPW_ENHANCED;
787 		} else {
788 			DRV_LOG(DEBUG, "MPW is supported");
789 			mps = MLX5_MPW;
790 		}
791 	} else {
792 		DRV_LOG(DEBUG, "MPW isn't supported");
793 		mps = MLX5_MPW_DISABLED;
794 	}
795 #ifdef HAVE_IBV_MLX5_MOD_SWP
796 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
797 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
798 	DRV_LOG(DEBUG, "SWP support: %u", swp);
799 #endif
800 	config.swp = !!swp;
801 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
802 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
803 		struct mlx5dv_striding_rq_caps mprq_caps =
804 			dv_attr.striding_rq_caps;
805 
806 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
807 			mprq_caps.min_single_stride_log_num_of_bytes);
808 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
809 			mprq_caps.max_single_stride_log_num_of_bytes);
810 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
811 			mprq_caps.min_single_wqe_log_num_of_strides);
812 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
813 			mprq_caps.max_single_wqe_log_num_of_strides);
814 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
815 			mprq_caps.supported_qpts);
816 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
817 		mprq = 1;
818 		mprq_min_stride_size_n =
819 			mprq_caps.min_single_stride_log_num_of_bytes;
820 		mprq_max_stride_size_n =
821 			mprq_caps.max_single_stride_log_num_of_bytes;
822 		mprq_min_stride_num_n =
823 			mprq_caps.min_single_wqe_log_num_of_strides;
824 		mprq_max_stride_num_n =
825 			mprq_caps.max_single_wqe_log_num_of_strides;
826 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
827 						   mprq_min_stride_num_n);
828 	}
829 #endif
830 	if (RTE_CACHE_LINE_SIZE == 128 &&
831 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
832 		cqe_comp = 0;
833 	else
834 		cqe_comp = 1;
835 	config.cqe_comp = cqe_comp;
836 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
837 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
838 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
839 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
840 			     (dv_attr.tunnel_offloads_caps &
841 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
842 	}
843 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
844 		tunnel_en ? "" : "not ");
845 #else
846 	DRV_LOG(WARNING,
847 		"tunnel offloading disabled due to old OFED/rdma-core version");
848 #endif
849 	config.tunnel_en = tunnel_en;
850 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
851 	mpls_en = ((dv_attr.tunnel_offloads_caps &
852 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
853 		   (dv_attr.tunnel_offloads_caps &
854 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
855 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
856 		mpls_en ? "" : "not ");
857 #else
858 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
859 		" old OFED/rdma-core version or firmware configuration");
860 #endif
861 	config.mpls_en = mpls_en;
862 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr);
863 	if (err) {
864 		DEBUG("ibv_query_device_ex() failed");
865 		goto error;
866 	}
867 	if (!switch_info->representor)
868 		rte_strlcpy(name, dpdk_dev->name, sizeof(name));
869 	else
870 		snprintf(name, sizeof(name), "%s_representor_%u",
871 			 dpdk_dev->name, switch_info->port_name);
872 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
873 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
874 		eth_dev = rte_eth_dev_attach_secondary(name);
875 		if (eth_dev == NULL) {
876 			DRV_LOG(ERR, "can not attach rte ethdev");
877 			rte_errno = ENOMEM;
878 			err = rte_errno;
879 			goto error;
880 		}
881 		eth_dev->device = dpdk_dev;
882 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
883 		err = mlx5_uar_init_secondary(eth_dev);
884 		if (err) {
885 			err = rte_errno;
886 			goto error;
887 		}
888 		/* Receive command fd from primary process */
889 		err = mlx5_socket_connect(eth_dev);
890 		if (err < 0) {
891 			err = rte_errno;
892 			goto error;
893 		}
894 		/* Remap UAR for Tx queues. */
895 		err = mlx5_tx_uar_remap(eth_dev, err);
896 		if (err) {
897 			err = rte_errno;
898 			goto error;
899 		}
900 		/*
901 		 * Ethdev pointer is still required as input since
902 		 * the primary device is not accessible from the
903 		 * secondary process.
904 		 */
905 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
906 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
907 		claim_zero(mlx5_glue->close_device(ctx));
908 		return eth_dev;
909 	}
910 	/* Check port status. */
911 	err = mlx5_glue->query_port(ctx, 1, &port_attr);
912 	if (err) {
913 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
914 		goto error;
915 	}
916 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
917 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
918 		err = EINVAL;
919 		goto error;
920 	}
921 	if (port_attr.state != IBV_PORT_ACTIVE)
922 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
923 			mlx5_glue->port_state_str(port_attr.state),
924 			port_attr.state);
925 	/* Allocate protection domain. */
926 	pd = mlx5_glue->alloc_pd(ctx);
927 	if (pd == NULL) {
928 		DRV_LOG(ERR, "PD allocation failure");
929 		err = ENOMEM;
930 		goto error;
931 	}
932 	priv = rte_zmalloc("ethdev private structure",
933 			   sizeof(*priv),
934 			   RTE_CACHE_LINE_SIZE);
935 	if (priv == NULL) {
936 		DRV_LOG(ERR, "priv allocation failure");
937 		err = ENOMEM;
938 		goto error;
939 	}
940 	priv->ctx = ctx;
941 	strncpy(priv->ibdev_name, priv->ctx->device->name,
942 		sizeof(priv->ibdev_name));
943 	strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
944 		sizeof(priv->ibdev_path));
945 	priv->device_attr = attr;
946 	priv->pd = pd;
947 	priv->mtu = ETHER_MTU;
948 #ifndef RTE_ARCH_64
949 	/* Initialize UAR access locks for 32bit implementations. */
950 	rte_spinlock_init(&priv->uar_lock_cq);
951 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
952 		rte_spinlock_init(&priv->uar_lock[i]);
953 #endif
954 	/* Some internal functions rely on Netlink sockets, open them now. */
955 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
956 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
957 	priv->nl_sn = 0;
958 	priv->representor = !!switch_info->representor;
959 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
960 	priv->representor_id =
961 		switch_info->representor ? switch_info->port_name : -1;
962 	/*
963 	 * Look for sibling devices in order to reuse their switch domain
964 	 * if any, otherwise allocate one.
965 	 */
966 	i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
967 	if (i > 0) {
968 		uint16_t port_id[i];
969 
970 		i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
971 		while (i--) {
972 			const struct priv *opriv =
973 				rte_eth_devices[port_id[i]].data->dev_private;
974 
975 			if (!opriv ||
976 			    opriv->domain_id ==
977 			    RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
978 				continue;
979 			priv->domain_id = opriv->domain_id;
980 			break;
981 		}
982 	}
983 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
984 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
985 		if (err) {
986 			err = rte_errno;
987 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
988 				strerror(rte_errno));
989 			goto error;
990 		}
991 		own_domain_id = 1;
992 	}
993 	err = mlx5_args(&config, dpdk_dev->devargs);
994 	if (err) {
995 		err = rte_errno;
996 		DRV_LOG(ERR, "failed to process device arguments: %s",
997 			strerror(rte_errno));
998 		goto error;
999 	}
1000 	config.hw_csum = !!(attr.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
1001 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1002 		(config.hw_csum ? "" : "not "));
1003 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1004 	config.flow_counter_en = !!attr.max_counter_sets;
1005 	mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
1006 	DRV_LOG(DEBUG, "counter type = %d, num of cs = %ld, attributes = %d",
1007 		cs_desc.counter_type, cs_desc.num_of_cs,
1008 		cs_desc.attributes);
1009 #endif
1010 	config.ind_table_max_size =
1011 		attr.rss_caps.max_rwq_indirection_table_size;
1012 	/*
1013 	 * Remove this check once DPDK supports larger/variable
1014 	 * indirection tables.
1015 	 */
1016 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1017 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1018 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1019 		config.ind_table_max_size);
1020 	config.hw_vlan_strip = !!(attr.raw_packet_caps &
1021 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1022 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1023 		(config.hw_vlan_strip ? "" : "not "));
1024 	config.hw_fcs_strip = !!(attr.raw_packet_caps &
1025 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1026 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1027 		(config.hw_fcs_strip ? "" : "not "));
1028 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
1029 	config.hw_padding = !!attr.rx_pad_end_addr_align;
1030 #endif
1031 	DRV_LOG(DEBUG, "hardware Rx end alignment padding is %ssupported",
1032 		(config.hw_padding ? "" : "not "));
1033 	config.tso = (attr.tso_caps.max_tso > 0 &&
1034 		      (attr.tso_caps.supported_qpts &
1035 		       (1 << IBV_QPT_RAW_PACKET)));
1036 	if (config.tso)
1037 		config.tso_max_payload_sz = attr.tso_caps.max_tso;
1038 	/*
1039 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1040 	 * by default.
1041 	 */
1042 	if (config.mps == MLX5_ARG_UNSET)
1043 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1044 							  MLX5_MPW_DISABLED;
1045 	else
1046 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1047 	DRV_LOG(INFO, "%sMPS is %s",
1048 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1049 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1050 	if (config.cqe_comp && !cqe_comp) {
1051 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1052 		config.cqe_comp = 0;
1053 	}
1054 	if (config.mprq.enabled && mprq) {
1055 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1056 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1057 			config.mprq.stride_num_n =
1058 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1059 					mprq_min_stride_num_n);
1060 			DRV_LOG(WARNING,
1061 				"the number of strides"
1062 				" for Multi-Packet RQ is out of range,"
1063 				" setting default value (%u)",
1064 				1 << config.mprq.stride_num_n);
1065 		}
1066 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1067 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1068 	} else if (config.mprq.enabled && !mprq) {
1069 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1070 		config.mprq.enabled = 0;
1071 	}
1072 	eth_dev = rte_eth_dev_allocate(name);
1073 	if (eth_dev == NULL) {
1074 		DRV_LOG(ERR, "can not allocate rte ethdev");
1075 		err = ENOMEM;
1076 		goto error;
1077 	}
1078 	if (priv->representor)
1079 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1080 	eth_dev->data->dev_private = priv;
1081 	priv->dev_data = eth_dev->data;
1082 	eth_dev->data->mac_addrs = priv->mac;
1083 	eth_dev->device = dpdk_dev;
1084 	eth_dev->device->driver = &mlx5_driver.driver;
1085 	err = mlx5_uar_init_primary(eth_dev);
1086 	if (err) {
1087 		err = rte_errno;
1088 		goto error;
1089 	}
1090 	/* Configure the first MAC address by default. */
1091 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1092 		DRV_LOG(ERR,
1093 			"port %u cannot get MAC address, is mlx5_en"
1094 			" loaded? (errno: %s)",
1095 			eth_dev->data->port_id, strerror(rte_errno));
1096 		err = ENODEV;
1097 		goto error;
1098 	}
1099 	DRV_LOG(INFO,
1100 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1101 		eth_dev->data->port_id,
1102 		mac.addr_bytes[0], mac.addr_bytes[1],
1103 		mac.addr_bytes[2], mac.addr_bytes[3],
1104 		mac.addr_bytes[4], mac.addr_bytes[5]);
1105 #ifndef NDEBUG
1106 	{
1107 		char ifname[IF_NAMESIZE];
1108 
1109 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1110 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1111 				eth_dev->data->port_id, ifname);
1112 		else
1113 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1114 				eth_dev->data->port_id);
1115 	}
1116 #endif
1117 	/* Get actual MTU if possible. */
1118 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1119 	if (err) {
1120 		err = rte_errno;
1121 		goto error;
1122 	}
1123 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1124 		priv->mtu);
1125 	/* Initialize burst functions to prevent crashes before link-up. */
1126 	eth_dev->rx_pkt_burst = removed_rx_burst;
1127 	eth_dev->tx_pkt_burst = removed_tx_burst;
1128 	eth_dev->dev_ops = &mlx5_dev_ops;
1129 	/* Register MAC address. */
1130 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1131 	if (vf && config.vf_nl_en)
1132 		mlx5_nl_mac_addr_sync(eth_dev);
1133 	priv->mnl_socket = mlx5_nl_flow_socket_create();
1134 	if (!priv->mnl_socket) {
1135 		err = -rte_errno;
1136 		DRV_LOG(WARNING,
1137 			"flow rules relying on switch offloads will not be"
1138 			" supported: cannot open libmnl socket: %s",
1139 			strerror(rte_errno));
1140 	} else {
1141 		struct rte_flow_error error;
1142 		unsigned int ifindex = mlx5_ifindex(eth_dev);
1143 
1144 		if (!ifindex) {
1145 			err = -rte_errno;
1146 			error.message =
1147 				"cannot retrieve network interface index";
1148 		} else {
1149 			err = mlx5_nl_flow_init(priv->mnl_socket, ifindex,
1150 						&error);
1151 		}
1152 		if (err) {
1153 			DRV_LOG(WARNING,
1154 				"flow rules relying on switch offloads will"
1155 				" not be supported: %s: %s",
1156 				error.message, strerror(rte_errno));
1157 			mlx5_nl_flow_socket_destroy(priv->mnl_socket);
1158 			priv->mnl_socket = NULL;
1159 		}
1160 	}
1161 	TAILQ_INIT(&priv->flows);
1162 	TAILQ_INIT(&priv->ctrl_flows);
1163 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1164 	struct mlx5dv_ctx_allocators alctr = {
1165 		.alloc = &mlx5_alloc_verbs_buf,
1166 		.free = &mlx5_free_verbs_buf,
1167 		.data = priv,
1168 	};
1169 	mlx5_glue->dv_set_context_attr(ctx, MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1170 				       (void *)((uintptr_t)&alctr));
1171 	/* Bring Ethernet device up. */
1172 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1173 		eth_dev->data->port_id);
1174 	mlx5_set_link_up(eth_dev);
1175 	/*
1176 	 * Even though the interrupt handler is not installed yet,
1177 	 * interrupts will still trigger on the asyn_fd from
1178 	 * Verbs context returned by ibv_open_device().
1179 	 */
1180 	mlx5_link_update(eth_dev, 0);
1181 	/* Store device configuration on private structure. */
1182 	priv->config = config;
1183 	/* Supported Verbs flow priority number detection. */
1184 	err = mlx5_flow_discover_priorities(eth_dev);
1185 	if (err < 0)
1186 		goto error;
1187 	priv->config.flow_prio = err;
1188 	/*
1189 	 * Once the device is added to the list of memory event
1190 	 * callback, its global MR cache table cannot be expanded
1191 	 * on the fly because of deadlock. If it overflows, lookup
1192 	 * should be done by searching MR list linearly, which is slow.
1193 	 */
1194 	err = mlx5_mr_btree_init(&priv->mr.cache,
1195 				 MLX5_MR_BTREE_CACHE_N * 2,
1196 				 eth_dev->device->numa_node);
1197 	if (err) {
1198 		err = rte_errno;
1199 		goto error;
1200 	}
1201 	/* Add device to memory callback list. */
1202 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1203 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1204 			 priv, mem_event_cb);
1205 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1206 	return eth_dev;
1207 error:
1208 	if (priv) {
1209 		if (priv->nl_socket_route >= 0)
1210 			close(priv->nl_socket_route);
1211 		if (priv->nl_socket_rdma >= 0)
1212 			close(priv->nl_socket_rdma);
1213 		if (priv->mnl_socket)
1214 			mlx5_nl_flow_socket_destroy(priv->mnl_socket);
1215 		if (own_domain_id)
1216 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1217 		rte_free(priv);
1218 	}
1219 	if (pd)
1220 		claim_zero(mlx5_glue->dealloc_pd(pd));
1221 	if (eth_dev)
1222 		rte_eth_dev_release_port(eth_dev);
1223 	if (ctx)
1224 		claim_zero(mlx5_glue->close_device(ctx));
1225 	assert(err > 0);
1226 	rte_errno = err;
1227 	return NULL;
1228 }
1229 
1230 /** Data associated with devices to spawn. */
1231 struct mlx5_dev_spawn_data {
1232 	unsigned int ifindex; /**< Network interface index. */
1233 	struct mlx5_switch_info info; /**< Switch information. */
1234 	struct ibv_device *ibv_dev; /**< Associated IB device. */
1235 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
1236 };
1237 
1238 /**
1239  * Comparison callback to sort device data.
1240  *
1241  * This is meant to be used with qsort().
1242  *
1243  * @param a[in]
1244  *   Pointer to pointer to first data object.
1245  * @param b[in]
1246  *   Pointer to pointer to second data object.
1247  *
1248  * @return
1249  *   0 if both objects are equal, less than 0 if the first argument is less
1250  *   than the second, greater than 0 otherwise.
1251  */
1252 static int
1253 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1254 {
1255 	const struct mlx5_switch_info *si_a =
1256 		&((const struct mlx5_dev_spawn_data *)a)->info;
1257 	const struct mlx5_switch_info *si_b =
1258 		&((const struct mlx5_dev_spawn_data *)b)->info;
1259 	int ret;
1260 
1261 	/* Master device first. */
1262 	ret = si_b->master - si_a->master;
1263 	if (ret)
1264 		return ret;
1265 	/* Then representor devices. */
1266 	ret = si_b->representor - si_a->representor;
1267 	if (ret)
1268 		return ret;
1269 	/* Unidentified devices come last in no specific order. */
1270 	if (!si_a->representor)
1271 		return 0;
1272 	/* Order representors by name. */
1273 	return si_a->port_name - si_b->port_name;
1274 }
1275 
1276 /**
1277  * DPDK callback to register a PCI device.
1278  *
1279  * This function spawns Ethernet devices out of a given PCI device.
1280  *
1281  * @param[in] pci_drv
1282  *   PCI driver structure (mlx5_driver).
1283  * @param[in] pci_dev
1284  *   PCI device information.
1285  *
1286  * @return
1287  *   0 on success, a negative errno value otherwise and rte_errno is set.
1288  */
1289 static int
1290 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1291 	       struct rte_pci_device *pci_dev)
1292 {
1293 	struct ibv_device **ibv_list;
1294 	unsigned int n = 0;
1295 	int vf;
1296 	int ret;
1297 
1298 	assert(pci_drv == &mlx5_driver);
1299 	errno = 0;
1300 	ibv_list = mlx5_glue->get_device_list(&ret);
1301 	if (!ibv_list) {
1302 		rte_errno = errno ? errno : ENOSYS;
1303 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1304 		return -rte_errno;
1305 	}
1306 
1307 	struct ibv_device *ibv_match[ret + 1];
1308 
1309 	while (ret-- > 0) {
1310 		struct rte_pci_addr pci_addr;
1311 
1312 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1313 		if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1314 			continue;
1315 		if (pci_dev->addr.domain != pci_addr.domain ||
1316 		    pci_dev->addr.bus != pci_addr.bus ||
1317 		    pci_dev->addr.devid != pci_addr.devid ||
1318 		    pci_dev->addr.function != pci_addr.function)
1319 			continue;
1320 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1321 			ibv_list[ret]->name);
1322 		ibv_match[n++] = ibv_list[ret];
1323 	}
1324 	ibv_match[n] = NULL;
1325 
1326 	struct mlx5_dev_spawn_data list[n];
1327 	int nl_route = n ? mlx5_nl_init(NETLINK_ROUTE) : -1;
1328 	int nl_rdma = n ? mlx5_nl_init(NETLINK_RDMA) : -1;
1329 	unsigned int i;
1330 	unsigned int u;
1331 
1332 	/*
1333 	 * The existence of several matching entries (n > 1) means port
1334 	 * representors have been instantiated. No existing Verbs call nor
1335 	 * /sys entries can tell them apart, this can only be done through
1336 	 * Netlink calls assuming kernel drivers are recent enough to
1337 	 * support them.
1338 	 *
1339 	 * In the event of identification failure through Netlink, try again
1340 	 * through sysfs, then either:
1341 	 *
1342 	 * 1. No device matches (n == 0), complain and bail out.
1343 	 * 2. A single IB device matches (n == 1) and is not a representor,
1344 	 *    assume no switch support.
1345 	 * 3. Otherwise no safe assumptions can be made; complain louder and
1346 	 *    bail out.
1347 	 */
1348 	for (i = 0; i != n; ++i) {
1349 		list[i].ibv_dev = ibv_match[i];
1350 		list[i].eth_dev = NULL;
1351 		if (nl_rdma < 0)
1352 			list[i].ifindex = 0;
1353 		else
1354 			list[i].ifindex = mlx5_nl_ifindex
1355 				(nl_rdma, list[i].ibv_dev->name);
1356 		if (nl_route < 0 ||
1357 		    !list[i].ifindex ||
1358 		    mlx5_nl_switch_info(nl_route, list[i].ifindex,
1359 					&list[i].info) ||
1360 		    ((!list[i].info.representor && !list[i].info.master) &&
1361 		     mlx5_sysfs_switch_info(list[i].ifindex, &list[i].info))) {
1362 			list[i].ifindex = 0;
1363 			memset(&list[i].info, 0, sizeof(list[i].info));
1364 			continue;
1365 		}
1366 	}
1367 	if (nl_rdma >= 0)
1368 		close(nl_rdma);
1369 	if (nl_route >= 0)
1370 		close(nl_route);
1371 	/* Count unidentified devices. */
1372 	for (u = 0, i = 0; i != n; ++i)
1373 		if (!list[i].info.master && !list[i].info.representor)
1374 			++u;
1375 	if (u) {
1376 		if (n == 1 && u == 1) {
1377 			/* Case #2. */
1378 			DRV_LOG(INFO, "no switch support detected");
1379 		} else {
1380 			/* Case #3. */
1381 			DRV_LOG(ERR,
1382 				"unable to tell which of the matching devices"
1383 				" is the master (lack of kernel support?)");
1384 			n = 0;
1385 		}
1386 	}
1387 	/*
1388 	 * Sort list to probe devices in natural order for users convenience
1389 	 * (i.e. master first, then representors from lowest to highest ID).
1390 	 */
1391 	if (n)
1392 		qsort(list, n, sizeof(*list), mlx5_dev_spawn_data_cmp);
1393 	switch (pci_dev->id.device_id) {
1394 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1395 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1396 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1397 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1398 		vf = 1;
1399 		break;
1400 	default:
1401 		vf = 0;
1402 	}
1403 	for (i = 0; i != n; ++i) {
1404 		uint32_t restore;
1405 
1406 		list[i].eth_dev = mlx5_dev_spawn
1407 			(&pci_dev->device, list[i].ibv_dev, vf, &list[i].info);
1408 		if (!list[i].eth_dev) {
1409 			if (rte_errno != EBUSY)
1410 				break;
1411 			/* Device is disabled, ignore it. */
1412 			continue;
1413 		}
1414 		restore = list[i].eth_dev->data->dev_flags;
1415 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1416 		/* Restore non-PCI flags cleared by the above call. */
1417 		list[i].eth_dev->data->dev_flags |= restore;
1418 		rte_eth_dev_probing_finish(list[i].eth_dev);
1419 	}
1420 	mlx5_glue->free_device_list(ibv_list);
1421 	if (!n) {
1422 		DRV_LOG(WARNING,
1423 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1424 			" are kernel drivers loaded?",
1425 			pci_dev->addr.domain, pci_dev->addr.bus,
1426 			pci_dev->addr.devid, pci_dev->addr.function);
1427 		rte_errno = ENOENT;
1428 		ret = -rte_errno;
1429 	} else if (i != n) {
1430 		DRV_LOG(ERR,
1431 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1432 			" encountering an error: %s",
1433 			pci_dev->addr.domain, pci_dev->addr.bus,
1434 			pci_dev->addr.devid, pci_dev->addr.function,
1435 			strerror(rte_errno));
1436 		ret = -rte_errno;
1437 		/* Roll back. */
1438 		while (i--) {
1439 			if (!list[i].eth_dev)
1440 				continue;
1441 			mlx5_dev_close(list[i].eth_dev);
1442 			if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1443 				rte_free(list[i].eth_dev->data->dev_private);
1444 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1445 		}
1446 		/* Restore original error. */
1447 		rte_errno = -ret;
1448 	} else {
1449 		ret = 0;
1450 	}
1451 	return ret;
1452 }
1453 
1454 static const struct rte_pci_id mlx5_pci_id_map[] = {
1455 	{
1456 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1457 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1458 	},
1459 	{
1460 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1461 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1462 	},
1463 	{
1464 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1465 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1466 	},
1467 	{
1468 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1469 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1470 	},
1471 	{
1472 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1473 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1474 	},
1475 	{
1476 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1477 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1478 	},
1479 	{
1480 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1481 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1482 	},
1483 	{
1484 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1485 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1486 	},
1487 	{
1488 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1489 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1490 	},
1491 	{
1492 		.vendor_id = 0
1493 	}
1494 };
1495 
1496 static struct rte_pci_driver mlx5_driver = {
1497 	.driver = {
1498 		.name = MLX5_DRIVER_NAME
1499 	},
1500 	.id_table = mlx5_pci_id_map,
1501 	.probe = mlx5_pci_probe,
1502 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1503 };
1504 
1505 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1506 
1507 /**
1508  * Suffix RTE_EAL_PMD_PATH with "-glue".
1509  *
1510  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1511  * suffixing its last component.
1512  *
1513  * @param buf[out]
1514  *   Output buffer, should be large enough otherwise NULL is returned.
1515  * @param size
1516  *   Size of @p out.
1517  *
1518  * @return
1519  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1520  */
1521 static char *
1522 mlx5_glue_path(char *buf, size_t size)
1523 {
1524 	static const char *const bad[] = { "/", ".", "..", NULL };
1525 	const char *path = RTE_EAL_PMD_PATH;
1526 	size_t len = strlen(path);
1527 	size_t off;
1528 	int i;
1529 
1530 	while (len && path[len - 1] == '/')
1531 		--len;
1532 	for (off = len; off && path[off - 1] != '/'; --off)
1533 		;
1534 	for (i = 0; bad[i]; ++i)
1535 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1536 			goto error;
1537 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1538 	if (i == -1 || (size_t)i >= size)
1539 		goto error;
1540 	return buf;
1541 error:
1542 	DRV_LOG(ERR,
1543 		"unable to append \"-glue\" to last component of"
1544 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1545 		" please re-configure DPDK");
1546 	return NULL;
1547 }
1548 
1549 /**
1550  * Initialization routine for run-time dependency on rdma-core.
1551  */
1552 static int
1553 mlx5_glue_init(void)
1554 {
1555 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1556 	const char *path[] = {
1557 		/*
1558 		 * A basic security check is necessary before trusting
1559 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1560 		 */
1561 		(geteuid() == getuid() && getegid() == getgid() ?
1562 		 getenv("MLX5_GLUE_PATH") : NULL),
1563 		/*
1564 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1565 		 * variant, otherwise let dlopen() look up libraries on its
1566 		 * own.
1567 		 */
1568 		(*RTE_EAL_PMD_PATH ?
1569 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1570 	};
1571 	unsigned int i = 0;
1572 	void *handle = NULL;
1573 	void **sym;
1574 	const char *dlmsg;
1575 
1576 	while (!handle && i != RTE_DIM(path)) {
1577 		const char *end;
1578 		size_t len;
1579 		int ret;
1580 
1581 		if (!path[i]) {
1582 			++i;
1583 			continue;
1584 		}
1585 		end = strpbrk(path[i], ":;");
1586 		if (!end)
1587 			end = path[i] + strlen(path[i]);
1588 		len = end - path[i];
1589 		ret = 0;
1590 		do {
1591 			char name[ret + 1];
1592 
1593 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1594 				       (int)len, path[i],
1595 				       (!len || *(end - 1) == '/') ? "" : "/");
1596 			if (ret == -1)
1597 				break;
1598 			if (sizeof(name) != (size_t)ret + 1)
1599 				continue;
1600 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1601 				name);
1602 			handle = dlopen(name, RTLD_LAZY);
1603 			break;
1604 		} while (1);
1605 		path[i] = end + 1;
1606 		if (!*end)
1607 			++i;
1608 	}
1609 	if (!handle) {
1610 		rte_errno = EINVAL;
1611 		dlmsg = dlerror();
1612 		if (dlmsg)
1613 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1614 		goto glue_error;
1615 	}
1616 	sym = dlsym(handle, "mlx5_glue");
1617 	if (!sym || !*sym) {
1618 		rte_errno = EINVAL;
1619 		dlmsg = dlerror();
1620 		if (dlmsg)
1621 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1622 		goto glue_error;
1623 	}
1624 	mlx5_glue = *sym;
1625 	return 0;
1626 glue_error:
1627 	if (handle)
1628 		dlclose(handle);
1629 	DRV_LOG(WARNING,
1630 		"cannot initialize PMD due to missing run-time dependency on"
1631 		" rdma-core libraries (libibverbs, libmlx5)");
1632 	return -rte_errno;
1633 }
1634 
1635 #endif
1636 
1637 /**
1638  * Driver initialization routine.
1639  */
1640 RTE_INIT(rte_mlx5_pmd_init)
1641 {
1642 	/* Initialize driver log type. */
1643 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1644 	if (mlx5_logtype >= 0)
1645 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1646 
1647 	/* Build the static tables for Verbs conversion. */
1648 	mlx5_set_ptype_table();
1649 	mlx5_set_cksum_table();
1650 	mlx5_set_swp_types_table();
1651 	/*
1652 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1653 	 * huge pages. Calling ibv_fork_init() during init allows
1654 	 * applications to use fork() safely for purposes other than
1655 	 * using this PMD, which is not supported in forked processes.
1656 	 */
1657 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1658 	/* Match the size of Rx completion entry to the size of a cacheline. */
1659 	if (RTE_CACHE_LINE_SIZE == 128)
1660 		setenv("MLX5_CQE_SIZE", "128", 0);
1661 	/*
1662 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
1663 	 * cleanup all the Verbs resources even when the device was removed.
1664 	 */
1665 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
1666 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1667 	if (mlx5_glue_init())
1668 		return;
1669 	assert(mlx5_glue);
1670 #endif
1671 #ifndef NDEBUG
1672 	/* Glue structure must not contain any NULL pointers. */
1673 	{
1674 		unsigned int i;
1675 
1676 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1677 			assert(((const void *const *)mlx5_glue)[i]);
1678 	}
1679 #endif
1680 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1681 		DRV_LOG(ERR,
1682 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1683 			mlx5_glue->version, MLX5_GLUE_VERSION);
1684 		return;
1685 	}
1686 	mlx5_glue->fork_init();
1687 	rte_pci_register(&mlx5_driver);
1688 }
1689 
1690 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1691 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1692 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1693