xref: /dpdk/drivers/net/mlx5/mlx5.c (revision c18feafa193c0d816eae3a4861b1f9016cf236d7)
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 #include "mlx5_flow.h"
50 
51 /* Device parameter to enable RX completion queue compression. */
52 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
53 
54 /* Device parameter to enable Multi-Packet Rx queue. */
55 #define MLX5_RX_MPRQ_EN "mprq_en"
56 
57 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
58 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
59 
60 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
61 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
62 
63 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
64 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
65 
66 /* Device parameter to configure inline send. */
67 #define MLX5_TXQ_INLINE "txq_inline"
68 
69 /*
70  * Device parameter to configure the number of TX queues threshold for
71  * enabling inline send.
72  */
73 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
74 
75 /* Device parameter to enable multi-packet send WQEs. */
76 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
77 
78 /* Device parameter to include 2 dsegs in the title WQEBB. */
79 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
80 
81 /* Device parameter to limit the size of inlining packet. */
82 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
83 
84 /* Device parameter to enable hardware Tx vector. */
85 #define MLX5_TX_VEC_EN "tx_vec_en"
86 
87 /* Device parameter to enable hardware Rx vector. */
88 #define MLX5_RX_VEC_EN "rx_vec_en"
89 
90 /* Allow L3 VXLAN flow creation. */
91 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
92 
93 /* Activate DV flow steering. */
94 #define MLX5_DV_FLOW_EN "dv_flow_en"
95 
96 /* Activate Netlink support in VF mode. */
97 #define MLX5_VF_NL_EN "vf_nl_en"
98 
99 /* Select port representors to instantiate. */
100 #define MLX5_REPRESENTOR "representor"
101 
102 #ifndef HAVE_IBV_MLX5_MOD_MPW
103 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
104 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
105 #endif
106 
107 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
108 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
109 #endif
110 
111 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
112 
113 /* Shared memory between primary and secondary processes. */
114 struct mlx5_shared_data *mlx5_shared_data;
115 
116 /* Spinlock for mlx5_shared_data allocation. */
117 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
118 
119 /** Driver-specific log messages type. */
120 int mlx5_logtype;
121 
122 /**
123  * Prepare shared data between primary and secondary process.
124  */
125 static void
126 mlx5_prepare_shared_data(void)
127 {
128 	const struct rte_memzone *mz;
129 
130 	rte_spinlock_lock(&mlx5_shared_data_lock);
131 	if (mlx5_shared_data == NULL) {
132 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
133 			/* Allocate shared memory. */
134 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
135 						 sizeof(*mlx5_shared_data),
136 						 SOCKET_ID_ANY, 0);
137 		} else {
138 			/* Lookup allocated shared memory. */
139 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
140 		}
141 		if (mz == NULL)
142 			rte_panic("Cannot allocate mlx5 shared data\n");
143 		mlx5_shared_data = mz->addr;
144 		/* Initialize shared data. */
145 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
146 			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
147 			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
148 		}
149 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
150 						mlx5_mr_mem_event_cb, NULL);
151 	}
152 	rte_spinlock_unlock(&mlx5_shared_data_lock);
153 }
154 
155 /**
156  * Retrieve integer value from environment variable.
157  *
158  * @param[in] name
159  *   Environment variable name.
160  *
161  * @return
162  *   Integer value, 0 if the variable is not set.
163  */
164 int
165 mlx5_getenv_int(const char *name)
166 {
167 	const char *val = getenv(name);
168 
169 	if (val == NULL)
170 		return 0;
171 	return atoi(val);
172 }
173 
174 /**
175  * Verbs callback to allocate a memory. This function should allocate the space
176  * according to the size provided residing inside a huge page.
177  * Please note that all allocation must respect the alignment from libmlx5
178  * (i.e. currently sysconf(_SC_PAGESIZE)).
179  *
180  * @param[in] size
181  *   The size in bytes of the memory to allocate.
182  * @param[in] data
183  *   A pointer to the callback data.
184  *
185  * @return
186  *   Allocated buffer, NULL otherwise and rte_errno is set.
187  */
188 static void *
189 mlx5_alloc_verbs_buf(size_t size, void *data)
190 {
191 	struct priv *priv = data;
192 	void *ret;
193 	size_t alignment = sysconf(_SC_PAGESIZE);
194 	unsigned int socket = SOCKET_ID_ANY;
195 
196 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
197 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
198 
199 		socket = ctrl->socket;
200 	} else if (priv->verbs_alloc_ctx.type ==
201 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
202 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
203 
204 		socket = ctrl->socket;
205 	}
206 	assert(data != NULL);
207 	ret = rte_malloc_socket(__func__, size, alignment, socket);
208 	if (!ret && size)
209 		rte_errno = ENOMEM;
210 	return ret;
211 }
212 
213 /**
214  * Verbs callback to free a memory.
215  *
216  * @param[in] ptr
217  *   A pointer to the memory to free.
218  * @param[in] data
219  *   A pointer to the callback data.
220  */
221 static void
222 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
223 {
224 	assert(data != NULL);
225 	rte_free(ptr);
226 }
227 
228 /**
229  * DPDK callback to close the device.
230  *
231  * Destroy all queues and objects, free memory.
232  *
233  * @param dev
234  *   Pointer to Ethernet device structure.
235  */
236 static void
237 mlx5_dev_close(struct rte_eth_dev *dev)
238 {
239 	struct priv *priv = dev->data->dev_private;
240 	unsigned int i;
241 	int ret;
242 
243 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
244 		dev->data->port_id,
245 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
246 	/* In case mlx5_dev_stop() has not been called. */
247 	mlx5_dev_interrupt_handler_uninstall(dev);
248 	mlx5_traffic_disable(dev);
249 	mlx5_flow_flush(dev, NULL);
250 	/* Prevent crashes when queues are still in use. */
251 	dev->rx_pkt_burst = removed_rx_burst;
252 	dev->tx_pkt_burst = removed_tx_burst;
253 	if (priv->rxqs != NULL) {
254 		/* XXX race condition if mlx5_rx_burst() is still running. */
255 		usleep(1000);
256 		for (i = 0; (i != priv->rxqs_n); ++i)
257 			mlx5_rxq_release(dev, i);
258 		priv->rxqs_n = 0;
259 		priv->rxqs = NULL;
260 	}
261 	if (priv->txqs != NULL) {
262 		/* XXX race condition if mlx5_tx_burst() is still running. */
263 		usleep(1000);
264 		for (i = 0; (i != priv->txqs_n); ++i)
265 			mlx5_txq_release(dev, i);
266 		priv->txqs_n = 0;
267 		priv->txqs = NULL;
268 	}
269 	mlx5_mprq_free_mp(dev);
270 	mlx5_mr_release(dev);
271 	if (priv->pd != NULL) {
272 		assert(priv->ctx != NULL);
273 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
274 		claim_zero(mlx5_glue->close_device(priv->ctx));
275 	} else
276 		assert(priv->ctx == NULL);
277 	if (priv->rss_conf.rss_key != NULL)
278 		rte_free(priv->rss_conf.rss_key);
279 	if (priv->reta_idx != NULL)
280 		rte_free(priv->reta_idx);
281 	if (priv->primary_socket)
282 		mlx5_socket_uninit(dev);
283 	if (priv->config.vf)
284 		mlx5_nl_mac_addr_flush(dev);
285 	if (priv->nl_socket_route >= 0)
286 		close(priv->nl_socket_route);
287 	if (priv->nl_socket_rdma >= 0)
288 		close(priv->nl_socket_rdma);
289 	if (priv->mnl_socket)
290 		mlx5_flow_tcf_socket_destroy(priv->mnl_socket);
291 	ret = mlx5_hrxq_ibv_verify(dev);
292 	if (ret)
293 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
294 			dev->data->port_id);
295 	ret = mlx5_ind_table_ibv_verify(dev);
296 	if (ret)
297 		DRV_LOG(WARNING, "port %u some indirection table still remain",
298 			dev->data->port_id);
299 	ret = mlx5_rxq_ibv_verify(dev);
300 	if (ret)
301 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
302 			dev->data->port_id);
303 	ret = mlx5_rxq_verify(dev);
304 	if (ret)
305 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
306 			dev->data->port_id);
307 	ret = mlx5_txq_ibv_verify(dev);
308 	if (ret)
309 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
310 			dev->data->port_id);
311 	ret = mlx5_txq_verify(dev);
312 	if (ret)
313 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
314 			dev->data->port_id);
315 	ret = mlx5_flow_verify(dev);
316 	if (ret)
317 		DRV_LOG(WARNING, "port %u some flows still remain",
318 			dev->data->port_id);
319 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
320 		unsigned int c = 0;
321 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
322 		uint16_t port_id[i];
323 
324 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
325 		while (i--) {
326 			struct priv *opriv =
327 				rte_eth_devices[port_id[i]].data->dev_private;
328 
329 			if (!opriv ||
330 			    opriv->domain_id != priv->domain_id ||
331 			    &rte_eth_devices[port_id[i]] == dev)
332 				continue;
333 			++c;
334 		}
335 		if (!c)
336 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
337 	}
338 	memset(priv, 0, sizeof(*priv));
339 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
340 }
341 
342 const struct eth_dev_ops mlx5_dev_ops = {
343 	.dev_configure = mlx5_dev_configure,
344 	.dev_start = mlx5_dev_start,
345 	.dev_stop = mlx5_dev_stop,
346 	.dev_set_link_down = mlx5_set_link_down,
347 	.dev_set_link_up = mlx5_set_link_up,
348 	.dev_close = mlx5_dev_close,
349 	.promiscuous_enable = mlx5_promiscuous_enable,
350 	.promiscuous_disable = mlx5_promiscuous_disable,
351 	.allmulticast_enable = mlx5_allmulticast_enable,
352 	.allmulticast_disable = mlx5_allmulticast_disable,
353 	.link_update = mlx5_link_update,
354 	.stats_get = mlx5_stats_get,
355 	.stats_reset = mlx5_stats_reset,
356 	.xstats_get = mlx5_xstats_get,
357 	.xstats_reset = mlx5_xstats_reset,
358 	.xstats_get_names = mlx5_xstats_get_names,
359 	.dev_infos_get = mlx5_dev_infos_get,
360 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
361 	.vlan_filter_set = mlx5_vlan_filter_set,
362 	.rx_queue_setup = mlx5_rx_queue_setup,
363 	.tx_queue_setup = mlx5_tx_queue_setup,
364 	.rx_queue_release = mlx5_rx_queue_release,
365 	.tx_queue_release = mlx5_tx_queue_release,
366 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
367 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
368 	.mac_addr_remove = mlx5_mac_addr_remove,
369 	.mac_addr_add = mlx5_mac_addr_add,
370 	.mac_addr_set = mlx5_mac_addr_set,
371 	.set_mc_addr_list = mlx5_set_mc_addr_list,
372 	.mtu_set = mlx5_dev_set_mtu,
373 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
374 	.vlan_offload_set = mlx5_vlan_offload_set,
375 	.reta_update = mlx5_dev_rss_reta_update,
376 	.reta_query = mlx5_dev_rss_reta_query,
377 	.rss_hash_update = mlx5_rss_hash_update,
378 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
379 	.filter_ctrl = mlx5_dev_filter_ctrl,
380 	.rx_descriptor_status = mlx5_rx_descriptor_status,
381 	.tx_descriptor_status = mlx5_tx_descriptor_status,
382 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
383 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
384 	.is_removed = mlx5_is_removed,
385 };
386 
387 static const struct eth_dev_ops mlx5_dev_sec_ops = {
388 	.stats_get = mlx5_stats_get,
389 	.stats_reset = mlx5_stats_reset,
390 	.xstats_get = mlx5_xstats_get,
391 	.xstats_reset = mlx5_xstats_reset,
392 	.xstats_get_names = mlx5_xstats_get_names,
393 	.dev_infos_get = mlx5_dev_infos_get,
394 	.rx_descriptor_status = mlx5_rx_descriptor_status,
395 	.tx_descriptor_status = mlx5_tx_descriptor_status,
396 };
397 
398 /* Available operators in flow isolated mode. */
399 const struct eth_dev_ops mlx5_dev_ops_isolate = {
400 	.dev_configure = mlx5_dev_configure,
401 	.dev_start = mlx5_dev_start,
402 	.dev_stop = mlx5_dev_stop,
403 	.dev_set_link_down = mlx5_set_link_down,
404 	.dev_set_link_up = mlx5_set_link_up,
405 	.dev_close = mlx5_dev_close,
406 	.promiscuous_enable = mlx5_promiscuous_enable,
407 	.promiscuous_disable = mlx5_promiscuous_disable,
408 	.allmulticast_enable = mlx5_allmulticast_enable,
409 	.allmulticast_disable = mlx5_allmulticast_disable,
410 	.link_update = mlx5_link_update,
411 	.stats_get = mlx5_stats_get,
412 	.stats_reset = mlx5_stats_reset,
413 	.xstats_get = mlx5_xstats_get,
414 	.xstats_reset = mlx5_xstats_reset,
415 	.xstats_get_names = mlx5_xstats_get_names,
416 	.dev_infos_get = mlx5_dev_infos_get,
417 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
418 	.vlan_filter_set = mlx5_vlan_filter_set,
419 	.rx_queue_setup = mlx5_rx_queue_setup,
420 	.tx_queue_setup = mlx5_tx_queue_setup,
421 	.rx_queue_release = mlx5_rx_queue_release,
422 	.tx_queue_release = mlx5_tx_queue_release,
423 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
424 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
425 	.mac_addr_remove = mlx5_mac_addr_remove,
426 	.mac_addr_add = mlx5_mac_addr_add,
427 	.mac_addr_set = mlx5_mac_addr_set,
428 	.set_mc_addr_list = mlx5_set_mc_addr_list,
429 	.mtu_set = mlx5_dev_set_mtu,
430 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
431 	.vlan_offload_set = mlx5_vlan_offload_set,
432 	.filter_ctrl = mlx5_dev_filter_ctrl,
433 	.rx_descriptor_status = mlx5_rx_descriptor_status,
434 	.tx_descriptor_status = mlx5_tx_descriptor_status,
435 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
436 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
437 	.is_removed = mlx5_is_removed,
438 };
439 
440 /**
441  * Verify and store value for device argument.
442  *
443  * @param[in] key
444  *   Key argument to verify.
445  * @param[in] val
446  *   Value associated with key.
447  * @param opaque
448  *   User data.
449  *
450  * @return
451  *   0 on success, a negative errno value otherwise and rte_errno is set.
452  */
453 static int
454 mlx5_args_check(const char *key, const char *val, void *opaque)
455 {
456 	struct mlx5_dev_config *config = opaque;
457 	unsigned long tmp;
458 
459 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
460 	if (!strcmp(MLX5_REPRESENTOR, key))
461 		return 0;
462 	errno = 0;
463 	tmp = strtoul(val, NULL, 0);
464 	if (errno) {
465 		rte_errno = errno;
466 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
467 		return -rte_errno;
468 	}
469 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
470 		config->cqe_comp = !!tmp;
471 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
472 		config->mprq.enabled = !!tmp;
473 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
474 		config->mprq.stride_num_n = tmp;
475 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
476 		config->mprq.max_memcpy_len = tmp;
477 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
478 		config->mprq.min_rxqs_num = tmp;
479 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
480 		config->txq_inline = tmp;
481 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
482 		config->txqs_inline = tmp;
483 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
484 		config->mps = !!tmp;
485 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
486 		config->mpw_hdr_dseg = !!tmp;
487 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
488 		config->inline_max_packet_sz = tmp;
489 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
490 		config->tx_vec_en = !!tmp;
491 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
492 		config->rx_vec_en = !!tmp;
493 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
494 		config->l3_vxlan_en = !!tmp;
495 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
496 		config->vf_nl_en = !!tmp;
497 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
498 		config->dv_flow_en = !!tmp;
499 	} else {
500 		DRV_LOG(WARNING, "%s: unknown parameter", key);
501 		rte_errno = EINVAL;
502 		return -rte_errno;
503 	}
504 	return 0;
505 }
506 
507 /**
508  * Parse device parameters.
509  *
510  * @param config
511  *   Pointer to device configuration structure.
512  * @param devargs
513  *   Device arguments structure.
514  *
515  * @return
516  *   0 on success, a negative errno value otherwise and rte_errno is set.
517  */
518 static int
519 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
520 {
521 	const char **params = (const char *[]){
522 		MLX5_RXQ_CQE_COMP_EN,
523 		MLX5_RX_MPRQ_EN,
524 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
525 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
526 		MLX5_RXQS_MIN_MPRQ,
527 		MLX5_TXQ_INLINE,
528 		MLX5_TXQS_MIN_INLINE,
529 		MLX5_TXQ_MPW_EN,
530 		MLX5_TXQ_MPW_HDR_DSEG_EN,
531 		MLX5_TXQ_MAX_INLINE_LEN,
532 		MLX5_TX_VEC_EN,
533 		MLX5_RX_VEC_EN,
534 		MLX5_L3_VXLAN_EN,
535 		MLX5_VF_NL_EN,
536 		MLX5_DV_FLOW_EN,
537 		MLX5_REPRESENTOR,
538 		NULL,
539 	};
540 	struct rte_kvargs *kvlist;
541 	int ret = 0;
542 	int i;
543 
544 	if (devargs == NULL)
545 		return 0;
546 	/* Following UGLY cast is done to pass checkpatch. */
547 	kvlist = rte_kvargs_parse(devargs->args, params);
548 	if (kvlist == NULL)
549 		return 0;
550 	/* Process parameters. */
551 	for (i = 0; (params[i] != NULL); ++i) {
552 		if (rte_kvargs_count(kvlist, params[i])) {
553 			ret = rte_kvargs_process(kvlist, params[i],
554 						 mlx5_args_check, config);
555 			if (ret) {
556 				rte_errno = EINVAL;
557 				rte_kvargs_free(kvlist);
558 				return -rte_errno;
559 			}
560 		}
561 	}
562 	rte_kvargs_free(kvlist);
563 	return 0;
564 }
565 
566 static struct rte_pci_driver mlx5_driver;
567 
568 /*
569  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
570  * local resource used by both primary and secondary to avoid duplicate
571  * reservation.
572  * The space has to be available on both primary and secondary process,
573  * TXQ UAR maps to this area using fixed mmap w/o double check.
574  */
575 static void *uar_base;
576 
577 static int
578 find_lower_va_bound(const struct rte_memseg_list *msl,
579 		const struct rte_memseg *ms, void *arg)
580 {
581 	void **addr = arg;
582 
583 	if (msl->external)
584 		return 0;
585 	if (*addr == NULL)
586 		*addr = ms->addr;
587 	else
588 		*addr = RTE_MIN(*addr, ms->addr);
589 
590 	return 0;
591 }
592 
593 /**
594  * Reserve UAR address space for primary process.
595  *
596  * @param[in] dev
597  *   Pointer to Ethernet device.
598  *
599  * @return
600  *   0 on success, a negative errno value otherwise and rte_errno is set.
601  */
602 static int
603 mlx5_uar_init_primary(struct rte_eth_dev *dev)
604 {
605 	struct priv *priv = dev->data->dev_private;
606 	void *addr = (void *)0;
607 
608 	if (uar_base) { /* UAR address space mapped. */
609 		priv->uar_base = uar_base;
610 		return 0;
611 	}
612 	/* find out lower bound of hugepage segments */
613 	rte_memseg_walk(find_lower_va_bound, &addr);
614 
615 	/* keep distance to hugepages to minimize potential conflicts. */
616 	addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
617 	/* anonymous mmap, no real memory consumption. */
618 	addr = mmap(addr, MLX5_UAR_SIZE,
619 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
620 	if (addr == MAP_FAILED) {
621 		DRV_LOG(ERR,
622 			"port %u failed to reserve UAR address space, please"
623 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
624 			dev->data->port_id);
625 		rte_errno = ENOMEM;
626 		return -rte_errno;
627 	}
628 	/* Accept either same addr or a new addr returned from mmap if target
629 	 * range occupied.
630 	 */
631 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
632 		dev->data->port_id, addr);
633 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
634 	uar_base = addr; /* process local, don't reserve again. */
635 	return 0;
636 }
637 
638 /**
639  * Reserve UAR address space for secondary process, align with
640  * primary process.
641  *
642  * @param[in] dev
643  *   Pointer to Ethernet device.
644  *
645  * @return
646  *   0 on success, a negative errno value otherwise and rte_errno is set.
647  */
648 static int
649 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
650 {
651 	struct priv *priv = dev->data->dev_private;
652 	void *addr;
653 
654 	assert(priv->uar_base);
655 	if (uar_base) { /* already reserved. */
656 		assert(uar_base == priv->uar_base);
657 		return 0;
658 	}
659 	/* anonymous mmap, no real memory consumption. */
660 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
661 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
662 	if (addr == MAP_FAILED) {
663 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
664 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
665 		rte_errno = ENXIO;
666 		return -rte_errno;
667 	}
668 	if (priv->uar_base != addr) {
669 		DRV_LOG(ERR,
670 			"port %u UAR address %p size %llu occupied, please"
671 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
672 			" --base-virtaddr",
673 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
674 		rte_errno = ENXIO;
675 		return -rte_errno;
676 	}
677 	uar_base = addr; /* process local, don't reserve again */
678 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
679 		dev->data->port_id, addr);
680 	return 0;
681 }
682 
683 /**
684  * Spawn an Ethernet device from Verbs information.
685  *
686  * @param dpdk_dev
687  *   Backing DPDK device.
688  * @param ibv_dev
689  *   Verbs device.
690  * @param vf
691  *   If nonzero, enable VF-specific features.
692  * @param[in] switch_info
693  *   Switch properties of Ethernet device.
694  *
695  * @return
696  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
697  *   is set. The following error is defined:
698  *
699  *   EBUSY: device is not supposed to be spawned.
700  */
701 static struct rte_eth_dev *
702 mlx5_dev_spawn(struct rte_device *dpdk_dev,
703 	       struct ibv_device *ibv_dev,
704 	       int vf,
705 	       const struct mlx5_switch_info *switch_info)
706 {
707 	struct ibv_context *ctx;
708 	struct ibv_device_attr_ex attr;
709 	struct ibv_port_attr port_attr;
710 	struct ibv_pd *pd = NULL;
711 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
712 	struct mlx5_dev_config config = {
713 		.vf = !!vf,
714 		.mps = MLX5_ARG_UNSET,
715 		.tx_vec_en = 1,
716 		.rx_vec_en = 1,
717 		.mpw_hdr_dseg = 0,
718 		.txq_inline = MLX5_ARG_UNSET,
719 		.txqs_inline = MLX5_ARG_UNSET,
720 		.inline_max_packet_sz = MLX5_ARG_UNSET,
721 		.vf_nl_en = 1,
722 		.mprq = {
723 			.enabled = 0,
724 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
725 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
726 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
727 		},
728 	};
729 	struct rte_eth_dev *eth_dev = NULL;
730 	struct priv *priv = NULL;
731 	int err = 0;
732 	unsigned int mps;
733 	unsigned int cqe_comp;
734 	unsigned int tunnel_en = 0;
735 	unsigned int mpls_en = 0;
736 	unsigned int swp = 0;
737 	unsigned int mprq = 0;
738 	unsigned int mprq_min_stride_size_n = 0;
739 	unsigned int mprq_max_stride_size_n = 0;
740 	unsigned int mprq_min_stride_num_n = 0;
741 	unsigned int mprq_max_stride_num_n = 0;
742 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
743 	struct ibv_counter_set_description cs_desc = { .counter_type = 0 };
744 #endif
745 	struct ether_addr mac;
746 	char name[RTE_ETH_NAME_MAX_LEN];
747 	int own_domain_id = 0;
748 	unsigned int i;
749 
750 	/* Determine if this port representor is supposed to be spawned. */
751 	if (switch_info->representor && dpdk_dev->devargs) {
752 		struct rte_eth_devargs eth_da;
753 
754 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
755 		if (err) {
756 			rte_errno = -err;
757 			DRV_LOG(ERR, "failed to process device arguments: %s",
758 				strerror(rte_errno));
759 			return NULL;
760 		}
761 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
762 			if (eth_da.representor_ports[i] ==
763 			    (uint16_t)switch_info->port_name)
764 				break;
765 		if (i == eth_da.nb_representor_ports) {
766 			rte_errno = EBUSY;
767 			return NULL;
768 		}
769 	}
770 	/* Prepare shared data between primary and secondary process. */
771 	mlx5_prepare_shared_data();
772 	errno = 0;
773 	ctx = mlx5_glue->open_device(ibv_dev);
774 	if (!ctx) {
775 		rte_errno = errno ? errno : ENODEV;
776 		return NULL;
777 	}
778 #ifdef HAVE_IBV_MLX5_MOD_SWP
779 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
780 #endif
781 	/*
782 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
783 	 * as all ConnectX-5 devices.
784 	 */
785 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
786 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
787 #endif
788 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
789 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
790 #endif
791 	mlx5_glue->dv_query_device(ctx, &dv_attr);
792 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
793 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
794 			DRV_LOG(DEBUG, "enhanced MPW is supported");
795 			mps = MLX5_MPW_ENHANCED;
796 		} else {
797 			DRV_LOG(DEBUG, "MPW is supported");
798 			mps = MLX5_MPW;
799 		}
800 	} else {
801 		DRV_LOG(DEBUG, "MPW isn't supported");
802 		mps = MLX5_MPW_DISABLED;
803 	}
804 #ifdef HAVE_IBV_MLX5_MOD_SWP
805 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
806 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
807 	DRV_LOG(DEBUG, "SWP support: %u", swp);
808 #endif
809 	config.swp = !!swp;
810 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
811 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
812 		struct mlx5dv_striding_rq_caps mprq_caps =
813 			dv_attr.striding_rq_caps;
814 
815 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
816 			mprq_caps.min_single_stride_log_num_of_bytes);
817 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
818 			mprq_caps.max_single_stride_log_num_of_bytes);
819 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
820 			mprq_caps.min_single_wqe_log_num_of_strides);
821 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
822 			mprq_caps.max_single_wqe_log_num_of_strides);
823 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
824 			mprq_caps.supported_qpts);
825 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
826 		mprq = 1;
827 		mprq_min_stride_size_n =
828 			mprq_caps.min_single_stride_log_num_of_bytes;
829 		mprq_max_stride_size_n =
830 			mprq_caps.max_single_stride_log_num_of_bytes;
831 		mprq_min_stride_num_n =
832 			mprq_caps.min_single_wqe_log_num_of_strides;
833 		mprq_max_stride_num_n =
834 			mprq_caps.max_single_wqe_log_num_of_strides;
835 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
836 						   mprq_min_stride_num_n);
837 	}
838 #endif
839 	if (RTE_CACHE_LINE_SIZE == 128 &&
840 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
841 		cqe_comp = 0;
842 	else
843 		cqe_comp = 1;
844 	config.cqe_comp = cqe_comp;
845 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
846 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
847 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
848 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
849 			     (dv_attr.tunnel_offloads_caps &
850 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
851 	}
852 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
853 		tunnel_en ? "" : "not ");
854 #else
855 	DRV_LOG(WARNING,
856 		"tunnel offloading disabled due to old OFED/rdma-core version");
857 #endif
858 	config.tunnel_en = tunnel_en;
859 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
860 	mpls_en = ((dv_attr.tunnel_offloads_caps &
861 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
862 		   (dv_attr.tunnel_offloads_caps &
863 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
864 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
865 		mpls_en ? "" : "not ");
866 #else
867 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
868 		" old OFED/rdma-core version or firmware configuration");
869 #endif
870 	config.mpls_en = mpls_en;
871 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr);
872 	if (err) {
873 		DEBUG("ibv_query_device_ex() failed");
874 		goto error;
875 	}
876 	if (!switch_info->representor)
877 		rte_strlcpy(name, dpdk_dev->name, sizeof(name));
878 	else
879 		snprintf(name, sizeof(name), "%s_representor_%u",
880 			 dpdk_dev->name, switch_info->port_name);
881 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
882 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
883 		eth_dev = rte_eth_dev_attach_secondary(name);
884 		if (eth_dev == NULL) {
885 			DRV_LOG(ERR, "can not attach rte ethdev");
886 			rte_errno = ENOMEM;
887 			err = rte_errno;
888 			goto error;
889 		}
890 		eth_dev->device = dpdk_dev;
891 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
892 		err = mlx5_uar_init_secondary(eth_dev);
893 		if (err) {
894 			err = rte_errno;
895 			goto error;
896 		}
897 		/* Receive command fd from primary process */
898 		err = mlx5_socket_connect(eth_dev);
899 		if (err < 0) {
900 			err = rte_errno;
901 			goto error;
902 		}
903 		/* Remap UAR for Tx queues. */
904 		err = mlx5_tx_uar_remap(eth_dev, err);
905 		if (err) {
906 			err = rte_errno;
907 			goto error;
908 		}
909 		/*
910 		 * Ethdev pointer is still required as input since
911 		 * the primary device is not accessible from the
912 		 * secondary process.
913 		 */
914 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
915 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
916 		claim_zero(mlx5_glue->close_device(ctx));
917 		return eth_dev;
918 	}
919 	/* Check port status. */
920 	err = mlx5_glue->query_port(ctx, 1, &port_attr);
921 	if (err) {
922 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
923 		goto error;
924 	}
925 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
926 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
927 		err = EINVAL;
928 		goto error;
929 	}
930 	if (port_attr.state != IBV_PORT_ACTIVE)
931 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
932 			mlx5_glue->port_state_str(port_attr.state),
933 			port_attr.state);
934 	/* Allocate protection domain. */
935 	pd = mlx5_glue->alloc_pd(ctx);
936 	if (pd == NULL) {
937 		DRV_LOG(ERR, "PD allocation failure");
938 		err = ENOMEM;
939 		goto error;
940 	}
941 	priv = rte_zmalloc("ethdev private structure",
942 			   sizeof(*priv),
943 			   RTE_CACHE_LINE_SIZE);
944 	if (priv == NULL) {
945 		DRV_LOG(ERR, "priv allocation failure");
946 		err = ENOMEM;
947 		goto error;
948 	}
949 	priv->ctx = ctx;
950 	strncpy(priv->ibdev_name, priv->ctx->device->name,
951 		sizeof(priv->ibdev_name));
952 	strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
953 		sizeof(priv->ibdev_path));
954 	priv->device_attr = attr;
955 	priv->pd = pd;
956 	priv->mtu = ETHER_MTU;
957 #ifndef RTE_ARCH_64
958 	/* Initialize UAR access locks for 32bit implementations. */
959 	rte_spinlock_init(&priv->uar_lock_cq);
960 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
961 		rte_spinlock_init(&priv->uar_lock[i]);
962 #endif
963 	/* Some internal functions rely on Netlink sockets, open them now. */
964 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
965 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
966 	priv->nl_sn = 0;
967 	priv->representor = !!switch_info->representor;
968 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
969 	priv->representor_id =
970 		switch_info->representor ? switch_info->port_name : -1;
971 	/*
972 	 * Look for sibling devices in order to reuse their switch domain
973 	 * if any, otherwise allocate one.
974 	 */
975 	i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
976 	if (i > 0) {
977 		uint16_t port_id[i];
978 
979 		i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
980 		while (i--) {
981 			const struct priv *opriv =
982 				rte_eth_devices[port_id[i]].data->dev_private;
983 
984 			if (!opriv ||
985 			    opriv->domain_id ==
986 			    RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
987 				continue;
988 			priv->domain_id = opriv->domain_id;
989 			break;
990 		}
991 	}
992 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
993 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
994 		if (err) {
995 			err = rte_errno;
996 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
997 				strerror(rte_errno));
998 			goto error;
999 		}
1000 		own_domain_id = 1;
1001 	}
1002 	err = mlx5_args(&config, dpdk_dev->devargs);
1003 	if (err) {
1004 		err = rte_errno;
1005 		DRV_LOG(ERR, "failed to process device arguments: %s",
1006 			strerror(rte_errno));
1007 		goto error;
1008 	}
1009 	config.hw_csum = !!(attr.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
1010 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1011 		(config.hw_csum ? "" : "not "));
1012 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1013 	config.flow_counter_en = !!attr.max_counter_sets;
1014 	mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
1015 	DRV_LOG(DEBUG, "counter type = %d, num of cs = %ld, attributes = %d",
1016 		cs_desc.counter_type, cs_desc.num_of_cs,
1017 		cs_desc.attributes);
1018 #endif
1019 	config.ind_table_max_size =
1020 		attr.rss_caps.max_rwq_indirection_table_size;
1021 	/*
1022 	 * Remove this check once DPDK supports larger/variable
1023 	 * indirection tables.
1024 	 */
1025 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1026 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1027 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1028 		config.ind_table_max_size);
1029 	config.hw_vlan_strip = !!(attr.raw_packet_caps &
1030 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1031 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1032 		(config.hw_vlan_strip ? "" : "not "));
1033 	config.hw_fcs_strip = !!(attr.raw_packet_caps &
1034 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1035 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1036 		(config.hw_fcs_strip ? "" : "not "));
1037 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
1038 	config.hw_padding = !!attr.rx_pad_end_addr_align;
1039 #endif
1040 	DRV_LOG(DEBUG, "hardware Rx end alignment padding is %ssupported",
1041 		(config.hw_padding ? "" : "not "));
1042 	config.tso = (attr.tso_caps.max_tso > 0 &&
1043 		      (attr.tso_caps.supported_qpts &
1044 		       (1 << IBV_QPT_RAW_PACKET)));
1045 	if (config.tso)
1046 		config.tso_max_payload_sz = attr.tso_caps.max_tso;
1047 	/*
1048 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1049 	 * by default.
1050 	 */
1051 	if (config.mps == MLX5_ARG_UNSET)
1052 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1053 							  MLX5_MPW_DISABLED;
1054 	else
1055 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1056 	DRV_LOG(INFO, "%sMPS is %s",
1057 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1058 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1059 	if (config.cqe_comp && !cqe_comp) {
1060 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1061 		config.cqe_comp = 0;
1062 	}
1063 	if (config.mprq.enabled && mprq) {
1064 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1065 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1066 			config.mprq.stride_num_n =
1067 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1068 					mprq_min_stride_num_n);
1069 			DRV_LOG(WARNING,
1070 				"the number of strides"
1071 				" for Multi-Packet RQ is out of range,"
1072 				" setting default value (%u)",
1073 				1 << config.mprq.stride_num_n);
1074 		}
1075 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1076 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1077 	} else if (config.mprq.enabled && !mprq) {
1078 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1079 		config.mprq.enabled = 0;
1080 	}
1081 	eth_dev = rte_eth_dev_allocate(name);
1082 	if (eth_dev == NULL) {
1083 		DRV_LOG(ERR, "can not allocate rte ethdev");
1084 		err = ENOMEM;
1085 		goto error;
1086 	}
1087 	if (priv->representor)
1088 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1089 	eth_dev->data->dev_private = priv;
1090 	priv->dev_data = eth_dev->data;
1091 	eth_dev->data->mac_addrs = priv->mac;
1092 	eth_dev->device = dpdk_dev;
1093 	err = mlx5_uar_init_primary(eth_dev);
1094 	if (err) {
1095 		err = rte_errno;
1096 		goto error;
1097 	}
1098 	/* Configure the first MAC address by default. */
1099 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1100 		DRV_LOG(ERR,
1101 			"port %u cannot get MAC address, is mlx5_en"
1102 			" loaded? (errno: %s)",
1103 			eth_dev->data->port_id, strerror(rte_errno));
1104 		err = ENODEV;
1105 		goto error;
1106 	}
1107 	DRV_LOG(INFO,
1108 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1109 		eth_dev->data->port_id,
1110 		mac.addr_bytes[0], mac.addr_bytes[1],
1111 		mac.addr_bytes[2], mac.addr_bytes[3],
1112 		mac.addr_bytes[4], mac.addr_bytes[5]);
1113 #ifndef NDEBUG
1114 	{
1115 		char ifname[IF_NAMESIZE];
1116 
1117 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1118 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1119 				eth_dev->data->port_id, ifname);
1120 		else
1121 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1122 				eth_dev->data->port_id);
1123 	}
1124 #endif
1125 	/* Get actual MTU if possible. */
1126 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1127 	if (err) {
1128 		err = rte_errno;
1129 		goto error;
1130 	}
1131 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1132 		priv->mtu);
1133 	/* Initialize burst functions to prevent crashes before link-up. */
1134 	eth_dev->rx_pkt_burst = removed_rx_burst;
1135 	eth_dev->tx_pkt_burst = removed_tx_burst;
1136 	eth_dev->dev_ops = &mlx5_dev_ops;
1137 	/* Register MAC address. */
1138 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1139 	if (vf && config.vf_nl_en)
1140 		mlx5_nl_mac_addr_sync(eth_dev);
1141 	priv->mnl_socket = mlx5_flow_tcf_socket_create();
1142 	if (!priv->mnl_socket) {
1143 		err = -rte_errno;
1144 		DRV_LOG(WARNING,
1145 			"flow rules relying on switch offloads will not be"
1146 			" supported: cannot open libmnl socket: %s",
1147 			strerror(rte_errno));
1148 	} else {
1149 		struct rte_flow_error error;
1150 		unsigned int ifindex = mlx5_ifindex(eth_dev);
1151 
1152 		if (!ifindex) {
1153 			err = -rte_errno;
1154 			error.message =
1155 				"cannot retrieve network interface index";
1156 		} else {
1157 			err = mlx5_flow_tcf_init(priv->mnl_socket, ifindex,
1158 						&error);
1159 		}
1160 		if (err) {
1161 			DRV_LOG(WARNING,
1162 				"flow rules relying on switch offloads will"
1163 				" not be supported: %s: %s",
1164 				error.message, strerror(rte_errno));
1165 			mlx5_flow_tcf_socket_destroy(priv->mnl_socket);
1166 			priv->mnl_socket = NULL;
1167 		}
1168 	}
1169 	TAILQ_INIT(&priv->flows);
1170 	TAILQ_INIT(&priv->ctrl_flows);
1171 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1172 	struct mlx5dv_ctx_allocators alctr = {
1173 		.alloc = &mlx5_alloc_verbs_buf,
1174 		.free = &mlx5_free_verbs_buf,
1175 		.data = priv,
1176 	};
1177 	mlx5_glue->dv_set_context_attr(ctx, MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1178 				       (void *)((uintptr_t)&alctr));
1179 	/* Bring Ethernet device up. */
1180 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1181 		eth_dev->data->port_id);
1182 	mlx5_set_link_up(eth_dev);
1183 	/*
1184 	 * Even though the interrupt handler is not installed yet,
1185 	 * interrupts will still trigger on the asyn_fd from
1186 	 * Verbs context returned by ibv_open_device().
1187 	 */
1188 	mlx5_link_update(eth_dev, 0);
1189 	/* Store device configuration on private structure. */
1190 	priv->config = config;
1191 	/* Supported Verbs flow priority number detection. */
1192 	err = mlx5_flow_discover_priorities(eth_dev);
1193 	if (err < 0)
1194 		goto error;
1195 	priv->config.flow_prio = err;
1196 	/*
1197 	 * Once the device is added to the list of memory event
1198 	 * callback, its global MR cache table cannot be expanded
1199 	 * on the fly because of deadlock. If it overflows, lookup
1200 	 * should be done by searching MR list linearly, which is slow.
1201 	 */
1202 	err = mlx5_mr_btree_init(&priv->mr.cache,
1203 				 MLX5_MR_BTREE_CACHE_N * 2,
1204 				 eth_dev->device->numa_node);
1205 	if (err) {
1206 		err = rte_errno;
1207 		goto error;
1208 	}
1209 	/* Add device to memory callback list. */
1210 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1211 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1212 			 priv, mem_event_cb);
1213 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1214 	return eth_dev;
1215 error:
1216 	if (priv) {
1217 		if (priv->nl_socket_route >= 0)
1218 			close(priv->nl_socket_route);
1219 		if (priv->nl_socket_rdma >= 0)
1220 			close(priv->nl_socket_rdma);
1221 		if (priv->mnl_socket)
1222 			mlx5_flow_tcf_socket_destroy(priv->mnl_socket);
1223 		if (own_domain_id)
1224 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1225 		rte_free(priv);
1226 		if (eth_dev != NULL)
1227 			eth_dev->data->dev_private = NULL;
1228 	}
1229 	if (pd)
1230 		claim_zero(mlx5_glue->dealloc_pd(pd));
1231 	if (eth_dev != NULL) {
1232 		/* mac_addrs must not be freed alone because part of dev_private */
1233 		eth_dev->data->mac_addrs = NULL;
1234 		rte_eth_dev_release_port(eth_dev);
1235 	}
1236 	if (ctx)
1237 		claim_zero(mlx5_glue->close_device(ctx));
1238 	assert(err > 0);
1239 	rte_errno = err;
1240 	return NULL;
1241 }
1242 
1243 /** Data associated with devices to spawn. */
1244 struct mlx5_dev_spawn_data {
1245 	unsigned int ifindex; /**< Network interface index. */
1246 	struct mlx5_switch_info info; /**< Switch information. */
1247 	struct ibv_device *ibv_dev; /**< Associated IB device. */
1248 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
1249 };
1250 
1251 /**
1252  * Comparison callback to sort device data.
1253  *
1254  * This is meant to be used with qsort().
1255  *
1256  * @param a[in]
1257  *   Pointer to pointer to first data object.
1258  * @param b[in]
1259  *   Pointer to pointer to second data object.
1260  *
1261  * @return
1262  *   0 if both objects are equal, less than 0 if the first argument is less
1263  *   than the second, greater than 0 otherwise.
1264  */
1265 static int
1266 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1267 {
1268 	const struct mlx5_switch_info *si_a =
1269 		&((const struct mlx5_dev_spawn_data *)a)->info;
1270 	const struct mlx5_switch_info *si_b =
1271 		&((const struct mlx5_dev_spawn_data *)b)->info;
1272 	int ret;
1273 
1274 	/* Master device first. */
1275 	ret = si_b->master - si_a->master;
1276 	if (ret)
1277 		return ret;
1278 	/* Then representor devices. */
1279 	ret = si_b->representor - si_a->representor;
1280 	if (ret)
1281 		return ret;
1282 	/* Unidentified devices come last in no specific order. */
1283 	if (!si_a->representor)
1284 		return 0;
1285 	/* Order representors by name. */
1286 	return si_a->port_name - si_b->port_name;
1287 }
1288 
1289 /**
1290  * DPDK callback to register a PCI device.
1291  *
1292  * This function spawns Ethernet devices out of a given PCI device.
1293  *
1294  * @param[in] pci_drv
1295  *   PCI driver structure (mlx5_driver).
1296  * @param[in] pci_dev
1297  *   PCI device information.
1298  *
1299  * @return
1300  *   0 on success, a negative errno value otherwise and rte_errno is set.
1301  */
1302 static int
1303 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1304 	       struct rte_pci_device *pci_dev)
1305 {
1306 	struct ibv_device **ibv_list;
1307 	unsigned int n = 0;
1308 	int vf;
1309 	int ret;
1310 
1311 	assert(pci_drv == &mlx5_driver);
1312 	errno = 0;
1313 	ibv_list = mlx5_glue->get_device_list(&ret);
1314 	if (!ibv_list) {
1315 		rte_errno = errno ? errno : ENOSYS;
1316 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1317 		return -rte_errno;
1318 	}
1319 
1320 	struct ibv_device *ibv_match[ret + 1];
1321 
1322 	while (ret-- > 0) {
1323 		struct rte_pci_addr pci_addr;
1324 
1325 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1326 		if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1327 			continue;
1328 		if (pci_dev->addr.domain != pci_addr.domain ||
1329 		    pci_dev->addr.bus != pci_addr.bus ||
1330 		    pci_dev->addr.devid != pci_addr.devid ||
1331 		    pci_dev->addr.function != pci_addr.function)
1332 			continue;
1333 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1334 			ibv_list[ret]->name);
1335 		ibv_match[n++] = ibv_list[ret];
1336 	}
1337 	ibv_match[n] = NULL;
1338 
1339 	struct mlx5_dev_spawn_data list[n];
1340 	int nl_route = n ? mlx5_nl_init(NETLINK_ROUTE) : -1;
1341 	int nl_rdma = n ? mlx5_nl_init(NETLINK_RDMA) : -1;
1342 	unsigned int i;
1343 	unsigned int u;
1344 
1345 	/*
1346 	 * The existence of several matching entries (n > 1) means port
1347 	 * representors have been instantiated. No existing Verbs call nor
1348 	 * /sys entries can tell them apart, this can only be done through
1349 	 * Netlink calls assuming kernel drivers are recent enough to
1350 	 * support them.
1351 	 *
1352 	 * In the event of identification failure through Netlink, try again
1353 	 * through sysfs, then either:
1354 	 *
1355 	 * 1. No device matches (n == 0), complain and bail out.
1356 	 * 2. A single IB device matches (n == 1) and is not a representor,
1357 	 *    assume no switch support.
1358 	 * 3. Otherwise no safe assumptions can be made; complain louder and
1359 	 *    bail out.
1360 	 */
1361 	for (i = 0; i != n; ++i) {
1362 		list[i].ibv_dev = ibv_match[i];
1363 		list[i].eth_dev = NULL;
1364 		if (nl_rdma < 0)
1365 			list[i].ifindex = 0;
1366 		else
1367 			list[i].ifindex = mlx5_nl_ifindex
1368 				(nl_rdma, list[i].ibv_dev->name);
1369 		if (nl_route < 0 ||
1370 		    !list[i].ifindex ||
1371 		    mlx5_nl_switch_info(nl_route, list[i].ifindex,
1372 					&list[i].info) ||
1373 		    ((!list[i].info.representor && !list[i].info.master) &&
1374 		     mlx5_sysfs_switch_info(list[i].ifindex, &list[i].info))) {
1375 			list[i].ifindex = 0;
1376 			memset(&list[i].info, 0, sizeof(list[i].info));
1377 			continue;
1378 		}
1379 	}
1380 	if (nl_rdma >= 0)
1381 		close(nl_rdma);
1382 	if (nl_route >= 0)
1383 		close(nl_route);
1384 	/* Count unidentified devices. */
1385 	for (u = 0, i = 0; i != n; ++i)
1386 		if (!list[i].info.master && !list[i].info.representor)
1387 			++u;
1388 	if (u) {
1389 		if (n == 1 && u == 1) {
1390 			/* Case #2. */
1391 			DRV_LOG(INFO, "no switch support detected");
1392 		} else {
1393 			/* Case #3. */
1394 			DRV_LOG(ERR,
1395 				"unable to tell which of the matching devices"
1396 				" is the master (lack of kernel support?)");
1397 			n = 0;
1398 		}
1399 	}
1400 	/*
1401 	 * Sort list to probe devices in natural order for users convenience
1402 	 * (i.e. master first, then representors from lowest to highest ID).
1403 	 */
1404 	if (n)
1405 		qsort(list, n, sizeof(*list), mlx5_dev_spawn_data_cmp);
1406 	switch (pci_dev->id.device_id) {
1407 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1408 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1409 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1410 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1411 		vf = 1;
1412 		break;
1413 	default:
1414 		vf = 0;
1415 	}
1416 	for (i = 0; i != n; ++i) {
1417 		uint32_t restore;
1418 
1419 		list[i].eth_dev = mlx5_dev_spawn
1420 			(&pci_dev->device, list[i].ibv_dev, vf, &list[i].info);
1421 		if (!list[i].eth_dev) {
1422 			if (rte_errno != EBUSY)
1423 				break;
1424 			/* Device is disabled, ignore it. */
1425 			continue;
1426 		}
1427 		restore = list[i].eth_dev->data->dev_flags;
1428 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1429 		/* Restore non-PCI flags cleared by the above call. */
1430 		list[i].eth_dev->data->dev_flags |= restore;
1431 		rte_eth_dev_probing_finish(list[i].eth_dev);
1432 	}
1433 	mlx5_glue->free_device_list(ibv_list);
1434 	if (!n) {
1435 		DRV_LOG(WARNING,
1436 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1437 			" are kernel drivers loaded?",
1438 			pci_dev->addr.domain, pci_dev->addr.bus,
1439 			pci_dev->addr.devid, pci_dev->addr.function);
1440 		rte_errno = ENOENT;
1441 		ret = -rte_errno;
1442 	} else if (i != n) {
1443 		DRV_LOG(ERR,
1444 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1445 			" encountering an error: %s",
1446 			pci_dev->addr.domain, pci_dev->addr.bus,
1447 			pci_dev->addr.devid, pci_dev->addr.function,
1448 			strerror(rte_errno));
1449 		ret = -rte_errno;
1450 		/* Roll back. */
1451 		while (i--) {
1452 			if (!list[i].eth_dev)
1453 				continue;
1454 			mlx5_dev_close(list[i].eth_dev);
1455 			/* mac_addrs must not be freed because in dev_private */
1456 			list[i].eth_dev->data->mac_addrs = NULL;
1457 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1458 		}
1459 		/* Restore original error. */
1460 		rte_errno = -ret;
1461 	} else {
1462 		ret = 0;
1463 	}
1464 	return ret;
1465 }
1466 
1467 static const struct rte_pci_id mlx5_pci_id_map[] = {
1468 	{
1469 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1470 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1471 	},
1472 	{
1473 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1474 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1475 	},
1476 	{
1477 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1478 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1479 	},
1480 	{
1481 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1482 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1483 	},
1484 	{
1485 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1486 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1487 	},
1488 	{
1489 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1490 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1491 	},
1492 	{
1493 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1494 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1495 	},
1496 	{
1497 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1498 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1499 	},
1500 	{
1501 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1502 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1503 	},
1504 	{
1505 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1506 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
1507 	},
1508 	{
1509 		.vendor_id = 0
1510 	}
1511 };
1512 
1513 static struct rte_pci_driver mlx5_driver = {
1514 	.driver = {
1515 		.name = MLX5_DRIVER_NAME
1516 	},
1517 	.id_table = mlx5_pci_id_map,
1518 	.probe = mlx5_pci_probe,
1519 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1520 };
1521 
1522 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1523 
1524 /**
1525  * Suffix RTE_EAL_PMD_PATH with "-glue".
1526  *
1527  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1528  * suffixing its last component.
1529  *
1530  * @param buf[out]
1531  *   Output buffer, should be large enough otherwise NULL is returned.
1532  * @param size
1533  *   Size of @p out.
1534  *
1535  * @return
1536  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1537  */
1538 static char *
1539 mlx5_glue_path(char *buf, size_t size)
1540 {
1541 	static const char *const bad[] = { "/", ".", "..", NULL };
1542 	const char *path = RTE_EAL_PMD_PATH;
1543 	size_t len = strlen(path);
1544 	size_t off;
1545 	int i;
1546 
1547 	while (len && path[len - 1] == '/')
1548 		--len;
1549 	for (off = len; off && path[off - 1] != '/'; --off)
1550 		;
1551 	for (i = 0; bad[i]; ++i)
1552 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1553 			goto error;
1554 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1555 	if (i == -1 || (size_t)i >= size)
1556 		goto error;
1557 	return buf;
1558 error:
1559 	DRV_LOG(ERR,
1560 		"unable to append \"-glue\" to last component of"
1561 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1562 		" please re-configure DPDK");
1563 	return NULL;
1564 }
1565 
1566 /**
1567  * Initialization routine for run-time dependency on rdma-core.
1568  */
1569 static int
1570 mlx5_glue_init(void)
1571 {
1572 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1573 	const char *path[] = {
1574 		/*
1575 		 * A basic security check is necessary before trusting
1576 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1577 		 */
1578 		(geteuid() == getuid() && getegid() == getgid() ?
1579 		 getenv("MLX5_GLUE_PATH") : NULL),
1580 		/*
1581 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1582 		 * variant, otherwise let dlopen() look up libraries on its
1583 		 * own.
1584 		 */
1585 		(*RTE_EAL_PMD_PATH ?
1586 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1587 	};
1588 	unsigned int i = 0;
1589 	void *handle = NULL;
1590 	void **sym;
1591 	const char *dlmsg;
1592 
1593 	while (!handle && i != RTE_DIM(path)) {
1594 		const char *end;
1595 		size_t len;
1596 		int ret;
1597 
1598 		if (!path[i]) {
1599 			++i;
1600 			continue;
1601 		}
1602 		end = strpbrk(path[i], ":;");
1603 		if (!end)
1604 			end = path[i] + strlen(path[i]);
1605 		len = end - path[i];
1606 		ret = 0;
1607 		do {
1608 			char name[ret + 1];
1609 
1610 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1611 				       (int)len, path[i],
1612 				       (!len || *(end - 1) == '/') ? "" : "/");
1613 			if (ret == -1)
1614 				break;
1615 			if (sizeof(name) != (size_t)ret + 1)
1616 				continue;
1617 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1618 				name);
1619 			handle = dlopen(name, RTLD_LAZY);
1620 			break;
1621 		} while (1);
1622 		path[i] = end + 1;
1623 		if (!*end)
1624 			++i;
1625 	}
1626 	if (!handle) {
1627 		rte_errno = EINVAL;
1628 		dlmsg = dlerror();
1629 		if (dlmsg)
1630 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1631 		goto glue_error;
1632 	}
1633 	sym = dlsym(handle, "mlx5_glue");
1634 	if (!sym || !*sym) {
1635 		rte_errno = EINVAL;
1636 		dlmsg = dlerror();
1637 		if (dlmsg)
1638 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1639 		goto glue_error;
1640 	}
1641 	mlx5_glue = *sym;
1642 	return 0;
1643 glue_error:
1644 	if (handle)
1645 		dlclose(handle);
1646 	DRV_LOG(WARNING,
1647 		"cannot initialize PMD due to missing run-time dependency on"
1648 		" rdma-core libraries (libibverbs, libmlx5)");
1649 	return -rte_errno;
1650 }
1651 
1652 #endif
1653 
1654 /**
1655  * Driver initialization routine.
1656  */
1657 RTE_INIT(rte_mlx5_pmd_init)
1658 {
1659 	/* Initialize driver log type. */
1660 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1661 	if (mlx5_logtype >= 0)
1662 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1663 
1664 	/* Build the static tables for Verbs conversion. */
1665 	mlx5_set_ptype_table();
1666 	mlx5_set_cksum_table();
1667 	mlx5_set_swp_types_table();
1668 	/*
1669 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1670 	 * huge pages. Calling ibv_fork_init() during init allows
1671 	 * applications to use fork() safely for purposes other than
1672 	 * using this PMD, which is not supported in forked processes.
1673 	 */
1674 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1675 	/* Match the size of Rx completion entry to the size of a cacheline. */
1676 	if (RTE_CACHE_LINE_SIZE == 128)
1677 		setenv("MLX5_CQE_SIZE", "128", 0);
1678 	/*
1679 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
1680 	 * cleanup all the Verbs resources even when the device was removed.
1681 	 */
1682 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
1683 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1684 	if (mlx5_glue_init())
1685 		return;
1686 	assert(mlx5_glue);
1687 #endif
1688 #ifndef NDEBUG
1689 	/* Glue structure must not contain any NULL pointers. */
1690 	{
1691 		unsigned int i;
1692 
1693 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1694 			assert(((const void *const *)mlx5_glue)[i]);
1695 	}
1696 #endif
1697 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1698 		DRV_LOG(ERR,
1699 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1700 			mlx5_glue->version, MLX5_GLUE_VERSION);
1701 		return;
1702 	}
1703 	mlx5_glue->fork_init();
1704 	rte_pci_register(&mlx5_driver);
1705 }
1706 
1707 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1708 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1709 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1710