xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 3caef7e1625f9ef1cf6785522917cb844eab5168)
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 RX completion entry padding to 128B. */
55 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
56 
57 /* Device parameter to enable Multi-Packet Rx queue. */
58 #define MLX5_RX_MPRQ_EN "mprq_en"
59 
60 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
61 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
62 
63 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
64 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
65 
66 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
67 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
68 
69 /* Device parameter to configure inline send. */
70 #define MLX5_TXQ_INLINE "txq_inline"
71 
72 /*
73  * Device parameter to configure the number of TX queues threshold for
74  * enabling inline send.
75  */
76 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
77 
78 /*
79  * Device parameter to configure the number of TX queues threshold for
80  * enabling vectorized Tx.
81  */
82 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
83 
84 /* Device parameter to enable multi-packet send WQEs. */
85 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
86 
87 /* Device parameter to include 2 dsegs in the title WQEBB. */
88 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
89 
90 /* Device parameter to limit the size of inlining packet. */
91 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
92 
93 /* Device parameter to enable hardware Tx vector. */
94 #define MLX5_TX_VEC_EN "tx_vec_en"
95 
96 /* Device parameter to enable hardware Rx vector. */
97 #define MLX5_RX_VEC_EN "rx_vec_en"
98 
99 /* Allow L3 VXLAN flow creation. */
100 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
101 
102 /* Activate DV flow steering. */
103 #define MLX5_DV_FLOW_EN "dv_flow_en"
104 
105 /* Activate Netlink support in VF mode. */
106 #define MLX5_VF_NL_EN "vf_nl_en"
107 
108 /* Select port representors to instantiate. */
109 #define MLX5_REPRESENTOR "representor"
110 
111 #ifndef HAVE_IBV_MLX5_MOD_MPW
112 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
113 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
114 #endif
115 
116 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
117 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
118 #endif
119 
120 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
121 
122 /* Shared memory between primary and secondary processes. */
123 struct mlx5_shared_data *mlx5_shared_data;
124 
125 /* Spinlock for mlx5_shared_data allocation. */
126 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
127 
128 /** Driver-specific log messages type. */
129 int mlx5_logtype;
130 
131 /**
132  * Prepare shared data between primary and secondary process.
133  */
134 static void
135 mlx5_prepare_shared_data(void)
136 {
137 	const struct rte_memzone *mz;
138 
139 	rte_spinlock_lock(&mlx5_shared_data_lock);
140 	if (mlx5_shared_data == NULL) {
141 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
142 			/* Allocate shared memory. */
143 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
144 						 sizeof(*mlx5_shared_data),
145 						 SOCKET_ID_ANY, 0);
146 		} else {
147 			/* Lookup allocated shared memory. */
148 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
149 		}
150 		if (mz == NULL)
151 			rte_panic("Cannot allocate mlx5 shared data\n");
152 		mlx5_shared_data = mz->addr;
153 		/* Initialize shared data. */
154 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
155 			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
156 			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
157 		}
158 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
159 						mlx5_mr_mem_event_cb, NULL);
160 	}
161 	rte_spinlock_unlock(&mlx5_shared_data_lock);
162 }
163 
164 /**
165  * Retrieve integer value from environment variable.
166  *
167  * @param[in] name
168  *   Environment variable name.
169  *
170  * @return
171  *   Integer value, 0 if the variable is not set.
172  */
173 int
174 mlx5_getenv_int(const char *name)
175 {
176 	const char *val = getenv(name);
177 
178 	if (val == NULL)
179 		return 0;
180 	return atoi(val);
181 }
182 
183 /**
184  * Verbs callback to allocate a memory. This function should allocate the space
185  * according to the size provided residing inside a huge page.
186  * Please note that all allocation must respect the alignment from libmlx5
187  * (i.e. currently sysconf(_SC_PAGESIZE)).
188  *
189  * @param[in] size
190  *   The size in bytes of the memory to allocate.
191  * @param[in] data
192  *   A pointer to the callback data.
193  *
194  * @return
195  *   Allocated buffer, NULL otherwise and rte_errno is set.
196  */
197 static void *
198 mlx5_alloc_verbs_buf(size_t size, void *data)
199 {
200 	struct priv *priv = data;
201 	void *ret;
202 	size_t alignment = sysconf(_SC_PAGESIZE);
203 	unsigned int socket = SOCKET_ID_ANY;
204 
205 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
206 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
207 
208 		socket = ctrl->socket;
209 	} else if (priv->verbs_alloc_ctx.type ==
210 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
211 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
212 
213 		socket = ctrl->socket;
214 	}
215 	assert(data != NULL);
216 	ret = rte_malloc_socket(__func__, size, alignment, socket);
217 	if (!ret && size)
218 		rte_errno = ENOMEM;
219 	return ret;
220 }
221 
222 /**
223  * Verbs callback to free a memory.
224  *
225  * @param[in] ptr
226  *   A pointer to the memory to free.
227  * @param[in] data
228  *   A pointer to the callback data.
229  */
230 static void
231 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
232 {
233 	assert(data != NULL);
234 	rte_free(ptr);
235 }
236 
237 /**
238  * DPDK callback to close the device.
239  *
240  * Destroy all queues and objects, free memory.
241  *
242  * @param dev
243  *   Pointer to Ethernet device structure.
244  */
245 static void
246 mlx5_dev_close(struct rte_eth_dev *dev)
247 {
248 	struct priv *priv = dev->data->dev_private;
249 	unsigned int i;
250 	int ret;
251 
252 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
253 		dev->data->port_id,
254 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
255 	/* In case mlx5_dev_stop() has not been called. */
256 	mlx5_dev_interrupt_handler_uninstall(dev);
257 	mlx5_traffic_disable(dev);
258 	mlx5_flow_flush(dev, NULL);
259 	/* Prevent crashes when queues are still in use. */
260 	dev->rx_pkt_burst = removed_rx_burst;
261 	dev->tx_pkt_burst = removed_tx_burst;
262 	if (priv->rxqs != NULL) {
263 		/* XXX race condition if mlx5_rx_burst() is still running. */
264 		usleep(1000);
265 		for (i = 0; (i != priv->rxqs_n); ++i)
266 			mlx5_rxq_release(dev, i);
267 		priv->rxqs_n = 0;
268 		priv->rxqs = NULL;
269 	}
270 	if (priv->txqs != NULL) {
271 		/* XXX race condition if mlx5_tx_burst() is still running. */
272 		usleep(1000);
273 		for (i = 0; (i != priv->txqs_n); ++i)
274 			mlx5_txq_release(dev, i);
275 		priv->txqs_n = 0;
276 		priv->txqs = NULL;
277 	}
278 	mlx5_mprq_free_mp(dev);
279 	mlx5_mr_release(dev);
280 	if (priv->pd != NULL) {
281 		assert(priv->ctx != NULL);
282 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
283 		claim_zero(mlx5_glue->close_device(priv->ctx));
284 	} else
285 		assert(priv->ctx == NULL);
286 	if (priv->rss_conf.rss_key != NULL)
287 		rte_free(priv->rss_conf.rss_key);
288 	if (priv->reta_idx != NULL)
289 		rte_free(priv->reta_idx);
290 	if (priv->primary_socket)
291 		mlx5_socket_uninit(dev);
292 	if (priv->config.vf)
293 		mlx5_nl_mac_addr_flush(dev);
294 	if (priv->nl_socket_route >= 0)
295 		close(priv->nl_socket_route);
296 	if (priv->nl_socket_rdma >= 0)
297 		close(priv->nl_socket_rdma);
298 	if (priv->tcf_context)
299 		mlx5_flow_tcf_context_destroy(priv->tcf_context);
300 	ret = mlx5_hrxq_ibv_verify(dev);
301 	if (ret)
302 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
303 			dev->data->port_id);
304 	ret = mlx5_ind_table_ibv_verify(dev);
305 	if (ret)
306 		DRV_LOG(WARNING, "port %u some indirection table still remain",
307 			dev->data->port_id);
308 	ret = mlx5_rxq_ibv_verify(dev);
309 	if (ret)
310 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
311 			dev->data->port_id);
312 	ret = mlx5_rxq_verify(dev);
313 	if (ret)
314 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
315 			dev->data->port_id);
316 	ret = mlx5_txq_ibv_verify(dev);
317 	if (ret)
318 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
319 			dev->data->port_id);
320 	ret = mlx5_txq_verify(dev);
321 	if (ret)
322 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
323 			dev->data->port_id);
324 	ret = mlx5_flow_verify(dev);
325 	if (ret)
326 		DRV_LOG(WARNING, "port %u some flows still remain",
327 			dev->data->port_id);
328 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
329 		unsigned int c = 0;
330 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
331 		uint16_t port_id[i];
332 
333 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
334 		while (i--) {
335 			struct priv *opriv =
336 				rte_eth_devices[port_id[i]].data->dev_private;
337 
338 			if (!opriv ||
339 			    opriv->domain_id != priv->domain_id ||
340 			    &rte_eth_devices[port_id[i]] == dev)
341 				continue;
342 			++c;
343 		}
344 		if (!c)
345 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
346 	}
347 	memset(priv, 0, sizeof(*priv));
348 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
349 	/*
350 	 * Reset mac_addrs to NULL such that it is not freed as part of
351 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
352 	 * it is freed when dev_private is freed.
353 	 */
354 	dev->data->mac_addrs = NULL;
355 }
356 
357 const struct eth_dev_ops mlx5_dev_ops = {
358 	.dev_configure = mlx5_dev_configure,
359 	.dev_start = mlx5_dev_start,
360 	.dev_stop = mlx5_dev_stop,
361 	.dev_set_link_down = mlx5_set_link_down,
362 	.dev_set_link_up = mlx5_set_link_up,
363 	.dev_close = mlx5_dev_close,
364 	.promiscuous_enable = mlx5_promiscuous_enable,
365 	.promiscuous_disable = mlx5_promiscuous_disable,
366 	.allmulticast_enable = mlx5_allmulticast_enable,
367 	.allmulticast_disable = mlx5_allmulticast_disable,
368 	.link_update = mlx5_link_update,
369 	.stats_get = mlx5_stats_get,
370 	.stats_reset = mlx5_stats_reset,
371 	.xstats_get = mlx5_xstats_get,
372 	.xstats_reset = mlx5_xstats_reset,
373 	.xstats_get_names = mlx5_xstats_get_names,
374 	.dev_infos_get = mlx5_dev_infos_get,
375 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
376 	.vlan_filter_set = mlx5_vlan_filter_set,
377 	.rx_queue_setup = mlx5_rx_queue_setup,
378 	.tx_queue_setup = mlx5_tx_queue_setup,
379 	.rx_queue_release = mlx5_rx_queue_release,
380 	.tx_queue_release = mlx5_tx_queue_release,
381 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
382 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
383 	.mac_addr_remove = mlx5_mac_addr_remove,
384 	.mac_addr_add = mlx5_mac_addr_add,
385 	.mac_addr_set = mlx5_mac_addr_set,
386 	.set_mc_addr_list = mlx5_set_mc_addr_list,
387 	.mtu_set = mlx5_dev_set_mtu,
388 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
389 	.vlan_offload_set = mlx5_vlan_offload_set,
390 	.reta_update = mlx5_dev_rss_reta_update,
391 	.reta_query = mlx5_dev_rss_reta_query,
392 	.rss_hash_update = mlx5_rss_hash_update,
393 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
394 	.filter_ctrl = mlx5_dev_filter_ctrl,
395 	.rx_descriptor_status = mlx5_rx_descriptor_status,
396 	.tx_descriptor_status = mlx5_tx_descriptor_status,
397 	.rx_queue_count = mlx5_rx_queue_count,
398 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
399 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
400 	.is_removed = mlx5_is_removed,
401 };
402 
403 static const struct eth_dev_ops mlx5_dev_sec_ops = {
404 	.stats_get = mlx5_stats_get,
405 	.stats_reset = mlx5_stats_reset,
406 	.xstats_get = mlx5_xstats_get,
407 	.xstats_reset = mlx5_xstats_reset,
408 	.xstats_get_names = mlx5_xstats_get_names,
409 	.dev_infos_get = mlx5_dev_infos_get,
410 	.rx_descriptor_status = mlx5_rx_descriptor_status,
411 	.tx_descriptor_status = mlx5_tx_descriptor_status,
412 };
413 
414 /* Available operators in flow isolated mode. */
415 const struct eth_dev_ops mlx5_dev_ops_isolate = {
416 	.dev_configure = mlx5_dev_configure,
417 	.dev_start = mlx5_dev_start,
418 	.dev_stop = mlx5_dev_stop,
419 	.dev_set_link_down = mlx5_set_link_down,
420 	.dev_set_link_up = mlx5_set_link_up,
421 	.dev_close = mlx5_dev_close,
422 	.promiscuous_enable = mlx5_promiscuous_enable,
423 	.promiscuous_disable = mlx5_promiscuous_disable,
424 	.allmulticast_enable = mlx5_allmulticast_enable,
425 	.allmulticast_disable = mlx5_allmulticast_disable,
426 	.link_update = mlx5_link_update,
427 	.stats_get = mlx5_stats_get,
428 	.stats_reset = mlx5_stats_reset,
429 	.xstats_get = mlx5_xstats_get,
430 	.xstats_reset = mlx5_xstats_reset,
431 	.xstats_get_names = mlx5_xstats_get_names,
432 	.dev_infos_get = mlx5_dev_infos_get,
433 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
434 	.vlan_filter_set = mlx5_vlan_filter_set,
435 	.rx_queue_setup = mlx5_rx_queue_setup,
436 	.tx_queue_setup = mlx5_tx_queue_setup,
437 	.rx_queue_release = mlx5_rx_queue_release,
438 	.tx_queue_release = mlx5_tx_queue_release,
439 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
440 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
441 	.mac_addr_remove = mlx5_mac_addr_remove,
442 	.mac_addr_add = mlx5_mac_addr_add,
443 	.mac_addr_set = mlx5_mac_addr_set,
444 	.set_mc_addr_list = mlx5_set_mc_addr_list,
445 	.mtu_set = mlx5_dev_set_mtu,
446 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
447 	.vlan_offload_set = mlx5_vlan_offload_set,
448 	.filter_ctrl = mlx5_dev_filter_ctrl,
449 	.rx_descriptor_status = mlx5_rx_descriptor_status,
450 	.tx_descriptor_status = mlx5_tx_descriptor_status,
451 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
452 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
453 	.is_removed = mlx5_is_removed,
454 };
455 
456 /**
457  * Verify and store value for device argument.
458  *
459  * @param[in] key
460  *   Key argument to verify.
461  * @param[in] val
462  *   Value associated with key.
463  * @param opaque
464  *   User data.
465  *
466  * @return
467  *   0 on success, a negative errno value otherwise and rte_errno is set.
468  */
469 static int
470 mlx5_args_check(const char *key, const char *val, void *opaque)
471 {
472 	struct mlx5_dev_config *config = opaque;
473 	unsigned long tmp;
474 
475 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
476 	if (!strcmp(MLX5_REPRESENTOR, key))
477 		return 0;
478 	errno = 0;
479 	tmp = strtoul(val, NULL, 0);
480 	if (errno) {
481 		rte_errno = errno;
482 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
483 		return -rte_errno;
484 	}
485 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
486 		config->cqe_comp = !!tmp;
487 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
488 		config->cqe_pad = !!tmp;
489 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
490 		config->mprq.enabled = !!tmp;
491 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
492 		config->mprq.stride_num_n = tmp;
493 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
494 		config->mprq.max_memcpy_len = tmp;
495 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
496 		config->mprq.min_rxqs_num = tmp;
497 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
498 		config->txq_inline = tmp;
499 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
500 		config->txqs_inline = tmp;
501 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
502 		config->txqs_vec = tmp;
503 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
504 		config->mps = !!tmp;
505 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
506 		config->mpw_hdr_dseg = !!tmp;
507 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
508 		config->inline_max_packet_sz = tmp;
509 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
510 		config->tx_vec_en = !!tmp;
511 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
512 		config->rx_vec_en = !!tmp;
513 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
514 		config->l3_vxlan_en = !!tmp;
515 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
516 		config->vf_nl_en = !!tmp;
517 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
518 		config->dv_flow_en = !!tmp;
519 	} else {
520 		DRV_LOG(WARNING, "%s: unknown parameter", key);
521 		rte_errno = EINVAL;
522 		return -rte_errno;
523 	}
524 	return 0;
525 }
526 
527 /**
528  * Parse device parameters.
529  *
530  * @param config
531  *   Pointer to device configuration structure.
532  * @param devargs
533  *   Device arguments structure.
534  *
535  * @return
536  *   0 on success, a negative errno value otherwise and rte_errno is set.
537  */
538 static int
539 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
540 {
541 	const char **params = (const char *[]){
542 		MLX5_RXQ_CQE_COMP_EN,
543 		MLX5_RXQ_CQE_PAD_EN,
544 		MLX5_RX_MPRQ_EN,
545 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
546 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
547 		MLX5_RXQS_MIN_MPRQ,
548 		MLX5_TXQ_INLINE,
549 		MLX5_TXQS_MIN_INLINE,
550 		MLX5_TXQS_MAX_VEC,
551 		MLX5_TXQ_MPW_EN,
552 		MLX5_TXQ_MPW_HDR_DSEG_EN,
553 		MLX5_TXQ_MAX_INLINE_LEN,
554 		MLX5_TX_VEC_EN,
555 		MLX5_RX_VEC_EN,
556 		MLX5_L3_VXLAN_EN,
557 		MLX5_VF_NL_EN,
558 		MLX5_DV_FLOW_EN,
559 		MLX5_REPRESENTOR,
560 		NULL,
561 	};
562 	struct rte_kvargs *kvlist;
563 	int ret = 0;
564 	int i;
565 
566 	if (devargs == NULL)
567 		return 0;
568 	/* Following UGLY cast is done to pass checkpatch. */
569 	kvlist = rte_kvargs_parse(devargs->args, params);
570 	if (kvlist == NULL)
571 		return 0;
572 	/* Process parameters. */
573 	for (i = 0; (params[i] != NULL); ++i) {
574 		if (rte_kvargs_count(kvlist, params[i])) {
575 			ret = rte_kvargs_process(kvlist, params[i],
576 						 mlx5_args_check, config);
577 			if (ret) {
578 				rte_errno = EINVAL;
579 				rte_kvargs_free(kvlist);
580 				return -rte_errno;
581 			}
582 		}
583 	}
584 	rte_kvargs_free(kvlist);
585 	return 0;
586 }
587 
588 static struct rte_pci_driver mlx5_driver;
589 
590 /*
591  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
592  * local resource used by both primary and secondary to avoid duplicate
593  * reservation.
594  * The space has to be available on both primary and secondary process,
595  * TXQ UAR maps to this area using fixed mmap w/o double check.
596  */
597 static void *uar_base;
598 
599 static int
600 find_lower_va_bound(const struct rte_memseg_list *msl,
601 		const struct rte_memseg *ms, void *arg)
602 {
603 	void **addr = arg;
604 
605 	if (msl->external)
606 		return 0;
607 	if (*addr == NULL)
608 		*addr = ms->addr;
609 	else
610 		*addr = RTE_MIN(*addr, ms->addr);
611 
612 	return 0;
613 }
614 
615 /**
616  * Reserve UAR address space for primary process.
617  *
618  * @param[in] dev
619  *   Pointer to Ethernet device.
620  *
621  * @return
622  *   0 on success, a negative errno value otherwise and rte_errno is set.
623  */
624 static int
625 mlx5_uar_init_primary(struct rte_eth_dev *dev)
626 {
627 	struct priv *priv = dev->data->dev_private;
628 	void *addr = (void *)0;
629 
630 	if (uar_base) { /* UAR address space mapped. */
631 		priv->uar_base = uar_base;
632 		return 0;
633 	}
634 	/* find out lower bound of hugepage segments */
635 	rte_memseg_walk(find_lower_va_bound, &addr);
636 
637 	/* keep distance to hugepages to minimize potential conflicts. */
638 	addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
639 	/* anonymous mmap, no real memory consumption. */
640 	addr = mmap(addr, MLX5_UAR_SIZE,
641 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
642 	if (addr == MAP_FAILED) {
643 		DRV_LOG(ERR,
644 			"port %u failed to reserve UAR address space, please"
645 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
646 			dev->data->port_id);
647 		rte_errno = ENOMEM;
648 		return -rte_errno;
649 	}
650 	/* Accept either same addr or a new addr returned from mmap if target
651 	 * range occupied.
652 	 */
653 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
654 		dev->data->port_id, addr);
655 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
656 	uar_base = addr; /* process local, don't reserve again. */
657 	return 0;
658 }
659 
660 /**
661  * Reserve UAR address space for secondary process, align with
662  * primary process.
663  *
664  * @param[in] dev
665  *   Pointer to Ethernet device.
666  *
667  * @return
668  *   0 on success, a negative errno value otherwise and rte_errno is set.
669  */
670 static int
671 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
672 {
673 	struct priv *priv = dev->data->dev_private;
674 	void *addr;
675 
676 	assert(priv->uar_base);
677 	if (uar_base) { /* already reserved. */
678 		assert(uar_base == priv->uar_base);
679 		return 0;
680 	}
681 	/* anonymous mmap, no real memory consumption. */
682 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
683 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
684 	if (addr == MAP_FAILED) {
685 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
686 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
687 		rte_errno = ENXIO;
688 		return -rte_errno;
689 	}
690 	if (priv->uar_base != addr) {
691 		DRV_LOG(ERR,
692 			"port %u UAR address %p size %llu occupied, please"
693 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
694 			" --base-virtaddr",
695 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
696 		rte_errno = ENXIO;
697 		return -rte_errno;
698 	}
699 	uar_base = addr; /* process local, don't reserve again */
700 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
701 		dev->data->port_id, addr);
702 	return 0;
703 }
704 
705 /**
706  * Spawn an Ethernet device from Verbs information.
707  *
708  * @param dpdk_dev
709  *   Backing DPDK device.
710  * @param ibv_dev
711  *   Verbs device.
712  * @param config
713  *   Device configuration parameters.
714  * @param[in] switch_info
715  *   Switch properties of Ethernet device.
716  *
717  * @return
718  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
719  *   is set. The following errors are defined:
720  *
721  *   EBUSY: device is not supposed to be spawned.
722  *   EEXIST: device is already spawned
723  */
724 static struct rte_eth_dev *
725 mlx5_dev_spawn(struct rte_device *dpdk_dev,
726 	       struct ibv_device *ibv_dev,
727 	       struct mlx5_dev_config config,
728 	       const struct mlx5_switch_info *switch_info)
729 {
730 	struct ibv_context *ctx = NULL;
731 	struct ibv_device_attr_ex attr;
732 	struct ibv_port_attr port_attr;
733 	struct ibv_pd *pd = NULL;
734 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
735 	struct rte_eth_dev *eth_dev = NULL;
736 	struct priv *priv = NULL;
737 	int err = 0;
738 	unsigned int mps;
739 	unsigned int cqe_comp;
740 	unsigned int cqe_pad = 0;
741 	unsigned int tunnel_en = 0;
742 	unsigned int mpls_en = 0;
743 	unsigned int swp = 0;
744 	unsigned int mprq = 0;
745 	unsigned int mprq_min_stride_size_n = 0;
746 	unsigned int mprq_max_stride_size_n = 0;
747 	unsigned int mprq_min_stride_num_n = 0;
748 	unsigned int mprq_max_stride_num_n = 0;
749 	struct ether_addr mac;
750 	char name[RTE_ETH_NAME_MAX_LEN];
751 	int own_domain_id = 0;
752 	uint16_t port_id;
753 	unsigned int i;
754 
755 	/* Determine if this port representor is supposed to be spawned. */
756 	if (switch_info->representor && dpdk_dev->devargs) {
757 		struct rte_eth_devargs eth_da;
758 
759 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
760 		if (err) {
761 			rte_errno = -err;
762 			DRV_LOG(ERR, "failed to process device arguments: %s",
763 				strerror(rte_errno));
764 			return NULL;
765 		}
766 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
767 			if (eth_da.representor_ports[i] ==
768 			    (uint16_t)switch_info->port_name)
769 				break;
770 		if (i == eth_da.nb_representor_ports) {
771 			rte_errno = EBUSY;
772 			return NULL;
773 		}
774 	}
775 	/* Build device name. */
776 	if (!switch_info->representor)
777 		rte_strlcpy(name, dpdk_dev->name, sizeof(name));
778 	else
779 		snprintf(name, sizeof(name), "%s_representor_%u",
780 			 dpdk_dev->name, switch_info->port_name);
781 	/* check if the device is already spawned */
782 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
783 		rte_errno = EEXIST;
784 		return NULL;
785 	}
786 	/* Prepare shared data between primary and secondary process. */
787 	mlx5_prepare_shared_data();
788 	errno = 0;
789 	ctx = mlx5_glue->dv_open_device(ibv_dev);
790 	if (ctx) {
791 		config.devx = 1;
792 		DRV_LOG(DEBUG, "DEVX is supported");
793 	} else {
794 		ctx = mlx5_glue->open_device(ibv_dev);
795 		if (!ctx) {
796 			rte_errno = errno ? errno : ENODEV;
797 			return NULL;
798 		}
799 	}
800 #ifdef HAVE_IBV_MLX5_MOD_SWP
801 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
802 #endif
803 	/*
804 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
805 	 * as all ConnectX-5 devices.
806 	 */
807 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
808 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
809 #endif
810 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
811 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
812 #endif
813 	mlx5_glue->dv_query_device(ctx, &dv_attr);
814 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
815 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
816 			DRV_LOG(DEBUG, "enhanced MPW is supported");
817 			mps = MLX5_MPW_ENHANCED;
818 		} else {
819 			DRV_LOG(DEBUG, "MPW is supported");
820 			mps = MLX5_MPW;
821 		}
822 	} else {
823 		DRV_LOG(DEBUG, "MPW isn't supported");
824 		mps = MLX5_MPW_DISABLED;
825 	}
826 #ifdef HAVE_IBV_MLX5_MOD_SWP
827 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
828 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
829 	DRV_LOG(DEBUG, "SWP support: %u", swp);
830 #endif
831 	config.swp = !!swp;
832 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
833 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
834 		struct mlx5dv_striding_rq_caps mprq_caps =
835 			dv_attr.striding_rq_caps;
836 
837 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
838 			mprq_caps.min_single_stride_log_num_of_bytes);
839 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
840 			mprq_caps.max_single_stride_log_num_of_bytes);
841 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
842 			mprq_caps.min_single_wqe_log_num_of_strides);
843 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
844 			mprq_caps.max_single_wqe_log_num_of_strides);
845 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
846 			mprq_caps.supported_qpts);
847 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
848 		mprq = 1;
849 		mprq_min_stride_size_n =
850 			mprq_caps.min_single_stride_log_num_of_bytes;
851 		mprq_max_stride_size_n =
852 			mprq_caps.max_single_stride_log_num_of_bytes;
853 		mprq_min_stride_num_n =
854 			mprq_caps.min_single_wqe_log_num_of_strides;
855 		mprq_max_stride_num_n =
856 			mprq_caps.max_single_wqe_log_num_of_strides;
857 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
858 						   mprq_min_stride_num_n);
859 	}
860 #endif
861 	if (RTE_CACHE_LINE_SIZE == 128 &&
862 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
863 		cqe_comp = 0;
864 	else
865 		cqe_comp = 1;
866 	config.cqe_comp = cqe_comp;
867 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
868 	/* Whether device supports 128B Rx CQE padding. */
869 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
870 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
871 #endif
872 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
873 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
874 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
875 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
876 			     (dv_attr.tunnel_offloads_caps &
877 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
878 	}
879 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
880 		tunnel_en ? "" : "not ");
881 #else
882 	DRV_LOG(WARNING,
883 		"tunnel offloading disabled due to old OFED/rdma-core version");
884 #endif
885 	config.tunnel_en = tunnel_en;
886 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
887 	mpls_en = ((dv_attr.tunnel_offloads_caps &
888 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
889 		   (dv_attr.tunnel_offloads_caps &
890 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
891 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
892 		mpls_en ? "" : "not ");
893 #else
894 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
895 		" old OFED/rdma-core version or firmware configuration");
896 #endif
897 	config.mpls_en = mpls_en;
898 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr);
899 	if (err) {
900 		DEBUG("ibv_query_device_ex() failed");
901 		goto error;
902 	}
903 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
904 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
905 		eth_dev = rte_eth_dev_attach_secondary(name);
906 		if (eth_dev == NULL) {
907 			DRV_LOG(ERR, "can not attach rte ethdev");
908 			rte_errno = ENOMEM;
909 			err = rte_errno;
910 			goto error;
911 		}
912 		eth_dev->device = dpdk_dev;
913 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
914 		err = mlx5_uar_init_secondary(eth_dev);
915 		if (err) {
916 			err = rte_errno;
917 			goto error;
918 		}
919 		/* Receive command fd from primary process */
920 		err = mlx5_socket_connect(eth_dev);
921 		if (err < 0) {
922 			err = rte_errno;
923 			goto error;
924 		}
925 		/* Remap UAR for Tx queues. */
926 		err = mlx5_tx_uar_remap(eth_dev, err);
927 		if (err) {
928 			err = rte_errno;
929 			goto error;
930 		}
931 		/*
932 		 * Ethdev pointer is still required as input since
933 		 * the primary device is not accessible from the
934 		 * secondary process.
935 		 */
936 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
937 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
938 		claim_zero(mlx5_glue->close_device(ctx));
939 		return eth_dev;
940 	}
941 	/* Check port status. */
942 	err = mlx5_glue->query_port(ctx, 1, &port_attr);
943 	if (err) {
944 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
945 		goto error;
946 	}
947 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
948 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
949 		err = EINVAL;
950 		goto error;
951 	}
952 	if (port_attr.state != IBV_PORT_ACTIVE)
953 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
954 			mlx5_glue->port_state_str(port_attr.state),
955 			port_attr.state);
956 	/* Allocate protection domain. */
957 	pd = mlx5_glue->alloc_pd(ctx);
958 	if (pd == NULL) {
959 		DRV_LOG(ERR, "PD allocation failure");
960 		err = ENOMEM;
961 		goto error;
962 	}
963 	priv = rte_zmalloc("ethdev private structure",
964 			   sizeof(*priv),
965 			   RTE_CACHE_LINE_SIZE);
966 	if (priv == NULL) {
967 		DRV_LOG(ERR, "priv allocation failure");
968 		err = ENOMEM;
969 		goto error;
970 	}
971 	priv->ctx = ctx;
972 	strncpy(priv->ibdev_name, priv->ctx->device->name,
973 		sizeof(priv->ibdev_name));
974 	strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
975 		sizeof(priv->ibdev_path));
976 	priv->device_attr = attr;
977 	priv->pd = pd;
978 	priv->mtu = ETHER_MTU;
979 #ifndef RTE_ARCH_64
980 	/* Initialize UAR access locks for 32bit implementations. */
981 	rte_spinlock_init(&priv->uar_lock_cq);
982 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
983 		rte_spinlock_init(&priv->uar_lock[i]);
984 #endif
985 	/* Some internal functions rely on Netlink sockets, open them now. */
986 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
987 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
988 	priv->nl_sn = 0;
989 	priv->representor = !!switch_info->representor;
990 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
991 	priv->representor_id =
992 		switch_info->representor ? switch_info->port_name : -1;
993 	/*
994 	 * Look for sibling devices in order to reuse their switch domain
995 	 * if any, otherwise allocate one.
996 	 */
997 	i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
998 	if (i > 0) {
999 		uint16_t port_id[i];
1000 
1001 		i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
1002 		while (i--) {
1003 			const struct priv *opriv =
1004 				rte_eth_devices[port_id[i]].data->dev_private;
1005 
1006 			if (!opriv ||
1007 			    opriv->domain_id ==
1008 			    RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1009 				continue;
1010 			priv->domain_id = opriv->domain_id;
1011 			break;
1012 		}
1013 	}
1014 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1015 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1016 		if (err) {
1017 			err = rte_errno;
1018 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1019 				strerror(rte_errno));
1020 			goto error;
1021 		}
1022 		own_domain_id = 1;
1023 	}
1024 	err = mlx5_args(&config, dpdk_dev->devargs);
1025 	if (err) {
1026 		err = rte_errno;
1027 		DRV_LOG(ERR, "failed to process device arguments: %s",
1028 			strerror(rte_errno));
1029 		goto error;
1030 	}
1031 	config.hw_csum = !!(attr.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
1032 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1033 		(config.hw_csum ? "" : "not "));
1034 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1035 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1036 	DRV_LOG(DEBUG, "counters are not supported");
1037 #endif
1038 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1039 	if (config.dv_flow_en) {
1040 		DRV_LOG(WARNING, "DV flow is not supported");
1041 		config.dv_flow_en = 0;
1042 	}
1043 #endif
1044 	config.ind_table_max_size =
1045 		attr.rss_caps.max_rwq_indirection_table_size;
1046 	/*
1047 	 * Remove this check once DPDK supports larger/variable
1048 	 * indirection tables.
1049 	 */
1050 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1051 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1052 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1053 		config.ind_table_max_size);
1054 	config.hw_vlan_strip = !!(attr.raw_packet_caps &
1055 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1056 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1057 		(config.hw_vlan_strip ? "" : "not "));
1058 	config.hw_fcs_strip = !!(attr.raw_packet_caps &
1059 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1060 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1061 		(config.hw_fcs_strip ? "" : "not "));
1062 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
1063 	config.hw_padding = !!attr.rx_pad_end_addr_align;
1064 #endif
1065 	DRV_LOG(DEBUG, "hardware Rx end alignment padding is %ssupported",
1066 		(config.hw_padding ? "" : "not "));
1067 	config.tso = (attr.tso_caps.max_tso > 0 &&
1068 		      (attr.tso_caps.supported_qpts &
1069 		       (1 << IBV_QPT_RAW_PACKET)));
1070 	if (config.tso)
1071 		config.tso_max_payload_sz = attr.tso_caps.max_tso;
1072 	/*
1073 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1074 	 * by default.
1075 	 */
1076 	if (config.mps == MLX5_ARG_UNSET)
1077 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1078 							  MLX5_MPW_DISABLED;
1079 	else
1080 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1081 	DRV_LOG(INFO, "%sMPS is %s",
1082 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1083 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1084 	if (config.cqe_comp && !cqe_comp) {
1085 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1086 		config.cqe_comp = 0;
1087 	}
1088 	if (config.cqe_pad && !cqe_pad) {
1089 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1090 		config.cqe_pad = 0;
1091 	} else if (config.cqe_pad) {
1092 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1093 	}
1094 	if (config.mprq.enabled && mprq) {
1095 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1096 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1097 			config.mprq.stride_num_n =
1098 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1099 					mprq_min_stride_num_n);
1100 			DRV_LOG(WARNING,
1101 				"the number of strides"
1102 				" for Multi-Packet RQ is out of range,"
1103 				" setting default value (%u)",
1104 				1 << config.mprq.stride_num_n);
1105 		}
1106 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1107 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1108 	} else if (config.mprq.enabled && !mprq) {
1109 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1110 		config.mprq.enabled = 0;
1111 	}
1112 	eth_dev = rte_eth_dev_allocate(name);
1113 	if (eth_dev == NULL) {
1114 		DRV_LOG(ERR, "can not allocate rte ethdev");
1115 		err = ENOMEM;
1116 		goto error;
1117 	}
1118 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1119 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1120 	if (priv->representor) {
1121 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1122 		eth_dev->data->representor_id = priv->representor_id;
1123 	}
1124 	eth_dev->data->dev_private = priv;
1125 	priv->dev_data = eth_dev->data;
1126 	eth_dev->data->mac_addrs = priv->mac;
1127 	eth_dev->device = dpdk_dev;
1128 	err = mlx5_uar_init_primary(eth_dev);
1129 	if (err) {
1130 		err = rte_errno;
1131 		goto error;
1132 	}
1133 	/* Configure the first MAC address by default. */
1134 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1135 		DRV_LOG(ERR,
1136 			"port %u cannot get MAC address, is mlx5_en"
1137 			" loaded? (errno: %s)",
1138 			eth_dev->data->port_id, strerror(rte_errno));
1139 		err = ENODEV;
1140 		goto error;
1141 	}
1142 	DRV_LOG(INFO,
1143 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1144 		eth_dev->data->port_id,
1145 		mac.addr_bytes[0], mac.addr_bytes[1],
1146 		mac.addr_bytes[2], mac.addr_bytes[3],
1147 		mac.addr_bytes[4], mac.addr_bytes[5]);
1148 #ifndef NDEBUG
1149 	{
1150 		char ifname[IF_NAMESIZE];
1151 
1152 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1153 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1154 				eth_dev->data->port_id, ifname);
1155 		else
1156 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1157 				eth_dev->data->port_id);
1158 	}
1159 #endif
1160 	/* Get actual MTU if possible. */
1161 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1162 	if (err) {
1163 		err = rte_errno;
1164 		goto error;
1165 	}
1166 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1167 		priv->mtu);
1168 	/* Initialize burst functions to prevent crashes before link-up. */
1169 	eth_dev->rx_pkt_burst = removed_rx_burst;
1170 	eth_dev->tx_pkt_burst = removed_tx_burst;
1171 	eth_dev->dev_ops = &mlx5_dev_ops;
1172 	/* Register MAC address. */
1173 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1174 	if (config.vf && config.vf_nl_en)
1175 		mlx5_nl_mac_addr_sync(eth_dev);
1176 	priv->tcf_context = mlx5_flow_tcf_context_create();
1177 	if (!priv->tcf_context) {
1178 		err = -rte_errno;
1179 		DRV_LOG(WARNING,
1180 			"flow rules relying on switch offloads will not be"
1181 			" supported: cannot open libmnl socket: %s",
1182 			strerror(rte_errno));
1183 	} else {
1184 		struct rte_flow_error error;
1185 		unsigned int ifindex = mlx5_ifindex(eth_dev);
1186 
1187 		if (!ifindex) {
1188 			err = -rte_errno;
1189 			error.message =
1190 				"cannot retrieve network interface index";
1191 		} else {
1192 			err = mlx5_flow_tcf_init(priv->tcf_context,
1193 						 ifindex, &error);
1194 		}
1195 		if (err) {
1196 			DRV_LOG(WARNING,
1197 				"flow rules relying on switch offloads will"
1198 				" not be supported: %s: %s",
1199 				error.message, strerror(rte_errno));
1200 			mlx5_flow_tcf_context_destroy(priv->tcf_context);
1201 			priv->tcf_context = NULL;
1202 		}
1203 	}
1204 	TAILQ_INIT(&priv->flows);
1205 	TAILQ_INIT(&priv->ctrl_flows);
1206 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1207 	struct mlx5dv_ctx_allocators alctr = {
1208 		.alloc = &mlx5_alloc_verbs_buf,
1209 		.free = &mlx5_free_verbs_buf,
1210 		.data = priv,
1211 	};
1212 	mlx5_glue->dv_set_context_attr(ctx, MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1213 				       (void *)((uintptr_t)&alctr));
1214 	/* Bring Ethernet device up. */
1215 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1216 		eth_dev->data->port_id);
1217 	mlx5_set_link_up(eth_dev);
1218 	/*
1219 	 * Even though the interrupt handler is not installed yet,
1220 	 * interrupts will still trigger on the asyn_fd from
1221 	 * Verbs context returned by ibv_open_device().
1222 	 */
1223 	mlx5_link_update(eth_dev, 0);
1224 	/* Store device configuration on private structure. */
1225 	priv->config = config;
1226 	/* Supported Verbs flow priority number detection. */
1227 	err = mlx5_flow_discover_priorities(eth_dev);
1228 	if (err < 0)
1229 		goto error;
1230 	priv->config.flow_prio = err;
1231 	/*
1232 	 * Once the device is added to the list of memory event
1233 	 * callback, its global MR cache table cannot be expanded
1234 	 * on the fly because of deadlock. If it overflows, lookup
1235 	 * should be done by searching MR list linearly, which is slow.
1236 	 */
1237 	err = mlx5_mr_btree_init(&priv->mr.cache,
1238 				 MLX5_MR_BTREE_CACHE_N * 2,
1239 				 eth_dev->device->numa_node);
1240 	if (err) {
1241 		err = rte_errno;
1242 		goto error;
1243 	}
1244 	/* Add device to memory callback list. */
1245 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1246 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1247 			 priv, mem_event_cb);
1248 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1249 	return eth_dev;
1250 error:
1251 	if (priv) {
1252 		if (priv->nl_socket_route >= 0)
1253 			close(priv->nl_socket_route);
1254 		if (priv->nl_socket_rdma >= 0)
1255 			close(priv->nl_socket_rdma);
1256 		if (priv->tcf_context)
1257 			mlx5_flow_tcf_context_destroy(priv->tcf_context);
1258 		if (own_domain_id)
1259 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1260 		rte_free(priv);
1261 		if (eth_dev != NULL)
1262 			eth_dev->data->dev_private = NULL;
1263 	}
1264 	if (pd)
1265 		claim_zero(mlx5_glue->dealloc_pd(pd));
1266 	if (eth_dev != NULL) {
1267 		/* mac_addrs must not be freed alone because part of dev_private */
1268 		eth_dev->data->mac_addrs = NULL;
1269 		rte_eth_dev_release_port(eth_dev);
1270 	}
1271 	if (ctx)
1272 		claim_zero(mlx5_glue->close_device(ctx));
1273 	assert(err > 0);
1274 	rte_errno = err;
1275 	return NULL;
1276 }
1277 
1278 /** Data associated with devices to spawn. */
1279 struct mlx5_dev_spawn_data {
1280 	unsigned int ifindex; /**< Network interface index. */
1281 	struct mlx5_switch_info info; /**< Switch information. */
1282 	struct ibv_device *ibv_dev; /**< Associated IB device. */
1283 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
1284 };
1285 
1286 /**
1287  * Comparison callback to sort device data.
1288  *
1289  * This is meant to be used with qsort().
1290  *
1291  * @param a[in]
1292  *   Pointer to pointer to first data object.
1293  * @param b[in]
1294  *   Pointer to pointer to second data object.
1295  *
1296  * @return
1297  *   0 if both objects are equal, less than 0 if the first argument is less
1298  *   than the second, greater than 0 otherwise.
1299  */
1300 static int
1301 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1302 {
1303 	const struct mlx5_switch_info *si_a =
1304 		&((const struct mlx5_dev_spawn_data *)a)->info;
1305 	const struct mlx5_switch_info *si_b =
1306 		&((const struct mlx5_dev_spawn_data *)b)->info;
1307 	int ret;
1308 
1309 	/* Master device first. */
1310 	ret = si_b->master - si_a->master;
1311 	if (ret)
1312 		return ret;
1313 	/* Then representor devices. */
1314 	ret = si_b->representor - si_a->representor;
1315 	if (ret)
1316 		return ret;
1317 	/* Unidentified devices come last in no specific order. */
1318 	if (!si_a->representor)
1319 		return 0;
1320 	/* Order representors by name. */
1321 	return si_a->port_name - si_b->port_name;
1322 }
1323 
1324 /**
1325  * DPDK callback to register a PCI device.
1326  *
1327  * This function spawns Ethernet devices out of a given PCI device.
1328  *
1329  * @param[in] pci_drv
1330  *   PCI driver structure (mlx5_driver).
1331  * @param[in] pci_dev
1332  *   PCI device information.
1333  *
1334  * @return
1335  *   0 on success, a negative errno value otherwise and rte_errno is set.
1336  */
1337 static int
1338 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1339 	       struct rte_pci_device *pci_dev)
1340 {
1341 	struct ibv_device **ibv_list;
1342 	unsigned int n = 0;
1343 	struct mlx5_dev_config dev_config;
1344 	int ret;
1345 
1346 	assert(pci_drv == &mlx5_driver);
1347 	errno = 0;
1348 	ibv_list = mlx5_glue->get_device_list(&ret);
1349 	if (!ibv_list) {
1350 		rte_errno = errno ? errno : ENOSYS;
1351 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1352 		return -rte_errno;
1353 	}
1354 
1355 	struct ibv_device *ibv_match[ret + 1];
1356 
1357 	while (ret-- > 0) {
1358 		struct rte_pci_addr pci_addr;
1359 
1360 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1361 		if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1362 			continue;
1363 		if (pci_dev->addr.domain != pci_addr.domain ||
1364 		    pci_dev->addr.bus != pci_addr.bus ||
1365 		    pci_dev->addr.devid != pci_addr.devid ||
1366 		    pci_dev->addr.function != pci_addr.function)
1367 			continue;
1368 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1369 			ibv_list[ret]->name);
1370 		ibv_match[n++] = ibv_list[ret];
1371 	}
1372 	ibv_match[n] = NULL;
1373 
1374 	struct mlx5_dev_spawn_data list[n];
1375 	int nl_route = n ? mlx5_nl_init(NETLINK_ROUTE) : -1;
1376 	int nl_rdma = n ? mlx5_nl_init(NETLINK_RDMA) : -1;
1377 	unsigned int i;
1378 	unsigned int u;
1379 
1380 	/*
1381 	 * The existence of several matching entries (n > 1) means port
1382 	 * representors have been instantiated. No existing Verbs call nor
1383 	 * /sys entries can tell them apart, this can only be done through
1384 	 * Netlink calls assuming kernel drivers are recent enough to
1385 	 * support them.
1386 	 *
1387 	 * In the event of identification failure through Netlink, try again
1388 	 * through sysfs, then either:
1389 	 *
1390 	 * 1. No device matches (n == 0), complain and bail out.
1391 	 * 2. A single IB device matches (n == 1) and is not a representor,
1392 	 *    assume no switch support.
1393 	 * 3. Otherwise no safe assumptions can be made; complain louder and
1394 	 *    bail out.
1395 	 */
1396 	for (i = 0; i != n; ++i) {
1397 		list[i].ibv_dev = ibv_match[i];
1398 		list[i].eth_dev = NULL;
1399 		if (nl_rdma < 0)
1400 			list[i].ifindex = 0;
1401 		else
1402 			list[i].ifindex = mlx5_nl_ifindex
1403 				(nl_rdma, list[i].ibv_dev->name);
1404 		if (nl_route < 0 ||
1405 		    !list[i].ifindex ||
1406 		    mlx5_nl_switch_info(nl_route, list[i].ifindex,
1407 					&list[i].info) ||
1408 		    ((!list[i].info.representor && !list[i].info.master) &&
1409 		     mlx5_sysfs_switch_info(list[i].ifindex, &list[i].info))) {
1410 			list[i].ifindex = 0;
1411 			memset(&list[i].info, 0, sizeof(list[i].info));
1412 			continue;
1413 		}
1414 	}
1415 	if (nl_rdma >= 0)
1416 		close(nl_rdma);
1417 	if (nl_route >= 0)
1418 		close(nl_route);
1419 	/* Count unidentified devices. */
1420 	for (u = 0, i = 0; i != n; ++i)
1421 		if (!list[i].info.master && !list[i].info.representor)
1422 			++u;
1423 	if (u) {
1424 		if (n == 1 && u == 1) {
1425 			/* Case #2. */
1426 			DRV_LOG(INFO, "no switch support detected");
1427 		} else {
1428 			/* Case #3. */
1429 			DRV_LOG(ERR,
1430 				"unable to tell which of the matching devices"
1431 				" is the master (lack of kernel support?)");
1432 			n = 0;
1433 		}
1434 	}
1435 	/*
1436 	 * Sort list to probe devices in natural order for users convenience
1437 	 * (i.e. master first, then representors from lowest to highest ID).
1438 	 */
1439 	if (n)
1440 		qsort(list, n, sizeof(*list), mlx5_dev_spawn_data_cmp);
1441 	/* Default configuration. */
1442 	dev_config = (struct mlx5_dev_config){
1443 		.mps = MLX5_ARG_UNSET,
1444 		.tx_vec_en = 1,
1445 		.rx_vec_en = 1,
1446 		.txq_inline = MLX5_ARG_UNSET,
1447 		.txqs_inline = MLX5_ARG_UNSET,
1448 		.txqs_vec = MLX5_ARG_UNSET,
1449 		.inline_max_packet_sz = MLX5_ARG_UNSET,
1450 		.vf_nl_en = 1,
1451 		.mprq = {
1452 			.enabled = 0, /* Disabled by default. */
1453 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1454 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1455 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1456 		},
1457 	};
1458 	/* Device speicific configuration. */
1459 	switch (pci_dev->id.device_id) {
1460 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1461 		dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1462 		break;
1463 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1464 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1465 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1466 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1467 		dev_config.vf = 1;
1468 		break;
1469 	default:
1470 		break;
1471 	}
1472 	/* Set architecture-dependent default value if unset. */
1473 	if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1474 		dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1475 	for (i = 0; i != n; ++i) {
1476 		uint32_t restore;
1477 
1478 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1479 						 list[i].ibv_dev, dev_config,
1480 						 &list[i].info);
1481 		if (!list[i].eth_dev) {
1482 			if (rte_errno != EBUSY && rte_errno != EEXIST)
1483 				break;
1484 			/* Device is disabled or already spawned. Ignore it. */
1485 			continue;
1486 		}
1487 		restore = list[i].eth_dev->data->dev_flags;
1488 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1489 		/* Restore non-PCI flags cleared by the above call. */
1490 		list[i].eth_dev->data->dev_flags |= restore;
1491 		rte_eth_dev_probing_finish(list[i].eth_dev);
1492 	}
1493 	mlx5_glue->free_device_list(ibv_list);
1494 	if (!n) {
1495 		DRV_LOG(WARNING,
1496 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1497 			" are kernel drivers loaded?",
1498 			pci_dev->addr.domain, pci_dev->addr.bus,
1499 			pci_dev->addr.devid, pci_dev->addr.function);
1500 		rte_errno = ENOENT;
1501 		ret = -rte_errno;
1502 	} else if (i != n) {
1503 		DRV_LOG(ERR,
1504 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1505 			" encountering an error: %s",
1506 			pci_dev->addr.domain, pci_dev->addr.bus,
1507 			pci_dev->addr.devid, pci_dev->addr.function,
1508 			strerror(rte_errno));
1509 		ret = -rte_errno;
1510 		/* Roll back. */
1511 		while (i--) {
1512 			if (!list[i].eth_dev)
1513 				continue;
1514 			mlx5_dev_close(list[i].eth_dev);
1515 			/* mac_addrs must not be freed because in dev_private */
1516 			list[i].eth_dev->data->mac_addrs = NULL;
1517 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1518 		}
1519 		/* Restore original error. */
1520 		rte_errno = -ret;
1521 	} else {
1522 		ret = 0;
1523 	}
1524 	return ret;
1525 }
1526 
1527 /**
1528  * DPDK callback to remove a PCI device.
1529  *
1530  * This function removes all Ethernet devices belong to a given PCI device.
1531  *
1532  * @param[in] pci_dev
1533  *   Pointer to the PCI device.
1534  *
1535  * @return
1536  *   0 on success, the function cannot fail.
1537  */
1538 static int
1539 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1540 {
1541 	uint16_t port_id;
1542 	struct rte_eth_dev *port;
1543 
1544 	for (port_id = 0; port_id < RTE_MAX_ETHPORTS; port_id++) {
1545 		port = &rte_eth_devices[port_id];
1546 		if (port->state != RTE_ETH_DEV_UNUSED &&
1547 				port->device == &pci_dev->device)
1548 			rte_eth_dev_close(port_id);
1549 	}
1550 	return 0;
1551 }
1552 
1553 static const struct rte_pci_id mlx5_pci_id_map[] = {
1554 	{
1555 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1556 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1557 	},
1558 	{
1559 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1560 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1561 	},
1562 	{
1563 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1564 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1565 	},
1566 	{
1567 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1568 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1569 	},
1570 	{
1571 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1572 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1573 	},
1574 	{
1575 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1576 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1577 	},
1578 	{
1579 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1580 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1581 	},
1582 	{
1583 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1584 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1585 	},
1586 	{
1587 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1588 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1589 	},
1590 	{
1591 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1592 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
1593 	},
1594 	{
1595 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1596 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
1597 	},
1598 	{
1599 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1600 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
1601 	},
1602 	{
1603 		.vendor_id = 0
1604 	}
1605 };
1606 
1607 static struct rte_pci_driver mlx5_driver = {
1608 	.driver = {
1609 		.name = MLX5_DRIVER_NAME
1610 	},
1611 	.id_table = mlx5_pci_id_map,
1612 	.probe = mlx5_pci_probe,
1613 	.remove = mlx5_pci_remove,
1614 	.drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
1615 		      RTE_PCI_DRV_PROBE_AGAIN),
1616 };
1617 
1618 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1619 
1620 /**
1621  * Suffix RTE_EAL_PMD_PATH with "-glue".
1622  *
1623  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1624  * suffixing its last component.
1625  *
1626  * @param buf[out]
1627  *   Output buffer, should be large enough otherwise NULL is returned.
1628  * @param size
1629  *   Size of @p out.
1630  *
1631  * @return
1632  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1633  */
1634 static char *
1635 mlx5_glue_path(char *buf, size_t size)
1636 {
1637 	static const char *const bad[] = { "/", ".", "..", NULL };
1638 	const char *path = RTE_EAL_PMD_PATH;
1639 	size_t len = strlen(path);
1640 	size_t off;
1641 	int i;
1642 
1643 	while (len && path[len - 1] == '/')
1644 		--len;
1645 	for (off = len; off && path[off - 1] != '/'; --off)
1646 		;
1647 	for (i = 0; bad[i]; ++i)
1648 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1649 			goto error;
1650 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1651 	if (i == -1 || (size_t)i >= size)
1652 		goto error;
1653 	return buf;
1654 error:
1655 	DRV_LOG(ERR,
1656 		"unable to append \"-glue\" to last component of"
1657 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1658 		" please re-configure DPDK");
1659 	return NULL;
1660 }
1661 
1662 /**
1663  * Initialization routine for run-time dependency on rdma-core.
1664  */
1665 static int
1666 mlx5_glue_init(void)
1667 {
1668 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1669 	const char *path[] = {
1670 		/*
1671 		 * A basic security check is necessary before trusting
1672 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1673 		 */
1674 		(geteuid() == getuid() && getegid() == getgid() ?
1675 		 getenv("MLX5_GLUE_PATH") : NULL),
1676 		/*
1677 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1678 		 * variant, otherwise let dlopen() look up libraries on its
1679 		 * own.
1680 		 */
1681 		(*RTE_EAL_PMD_PATH ?
1682 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1683 	};
1684 	unsigned int i = 0;
1685 	void *handle = NULL;
1686 	void **sym;
1687 	const char *dlmsg;
1688 
1689 	while (!handle && i != RTE_DIM(path)) {
1690 		const char *end;
1691 		size_t len;
1692 		int ret;
1693 
1694 		if (!path[i]) {
1695 			++i;
1696 			continue;
1697 		}
1698 		end = strpbrk(path[i], ":;");
1699 		if (!end)
1700 			end = path[i] + strlen(path[i]);
1701 		len = end - path[i];
1702 		ret = 0;
1703 		do {
1704 			char name[ret + 1];
1705 
1706 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1707 				       (int)len, path[i],
1708 				       (!len || *(end - 1) == '/') ? "" : "/");
1709 			if (ret == -1)
1710 				break;
1711 			if (sizeof(name) != (size_t)ret + 1)
1712 				continue;
1713 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1714 				name);
1715 			handle = dlopen(name, RTLD_LAZY);
1716 			break;
1717 		} while (1);
1718 		path[i] = end + 1;
1719 		if (!*end)
1720 			++i;
1721 	}
1722 	if (!handle) {
1723 		rte_errno = EINVAL;
1724 		dlmsg = dlerror();
1725 		if (dlmsg)
1726 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1727 		goto glue_error;
1728 	}
1729 	sym = dlsym(handle, "mlx5_glue");
1730 	if (!sym || !*sym) {
1731 		rte_errno = EINVAL;
1732 		dlmsg = dlerror();
1733 		if (dlmsg)
1734 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1735 		goto glue_error;
1736 	}
1737 	mlx5_glue = *sym;
1738 	return 0;
1739 glue_error:
1740 	if (handle)
1741 		dlclose(handle);
1742 	DRV_LOG(WARNING,
1743 		"cannot initialize PMD due to missing run-time dependency on"
1744 		" rdma-core libraries (libibverbs, libmlx5)");
1745 	return -rte_errno;
1746 }
1747 
1748 #endif
1749 
1750 /**
1751  * Driver initialization routine.
1752  */
1753 RTE_INIT(rte_mlx5_pmd_init)
1754 {
1755 	/* Initialize driver log type. */
1756 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1757 	if (mlx5_logtype >= 0)
1758 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1759 
1760 	/* Build the static tables for Verbs conversion. */
1761 	mlx5_set_ptype_table();
1762 	mlx5_set_cksum_table();
1763 	mlx5_set_swp_types_table();
1764 	/*
1765 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1766 	 * huge pages. Calling ibv_fork_init() during init allows
1767 	 * applications to use fork() safely for purposes other than
1768 	 * using this PMD, which is not supported in forked processes.
1769 	 */
1770 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1771 	/* Match the size of Rx completion entry to the size of a cacheline. */
1772 	if (RTE_CACHE_LINE_SIZE == 128)
1773 		setenv("MLX5_CQE_SIZE", "128", 0);
1774 	/*
1775 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
1776 	 * cleanup all the Verbs resources even when the device was removed.
1777 	 */
1778 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
1779 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1780 	if (mlx5_glue_init())
1781 		return;
1782 	assert(mlx5_glue);
1783 #endif
1784 #ifndef NDEBUG
1785 	/* Glue structure must not contain any NULL pointers. */
1786 	{
1787 		unsigned int i;
1788 
1789 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1790 			assert(((const void *const *)mlx5_glue)[i]);
1791 	}
1792 #endif
1793 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1794 		DRV_LOG(ERR,
1795 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1796 			mlx5_glue->version, MLX5_GLUE_VERSION);
1797 		return;
1798 	}
1799 	mlx5_glue->fork_init();
1800 	rte_pci_register(&mlx5_driver);
1801 }
1802 
1803 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1804 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1805 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1806