xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 2a7bb4fdf61e9edfb7adbaecb50e728b82da9e23)
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 padding Rx packet to cacheline size. */
58 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
59 
60 /* Device parameter to enable Multi-Packet Rx queue. */
61 #define MLX5_RX_MPRQ_EN "mprq_en"
62 
63 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
64 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
65 
66 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
67 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
68 
69 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
70 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
71 
72 /* Device parameter to configure inline send. */
73 #define MLX5_TXQ_INLINE "txq_inline"
74 
75 /*
76  * Device parameter to configure the number of TX queues threshold for
77  * enabling inline send.
78  */
79 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
80 
81 /*
82  * Device parameter to configure the number of TX queues threshold for
83  * enabling vectorized Tx.
84  */
85 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
86 
87 /* Device parameter to enable multi-packet send WQEs. */
88 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
89 
90 /* Device parameter to include 2 dsegs in the title WQEBB. */
91 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
92 
93 /* Device parameter to limit the size of inlining packet. */
94 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
95 
96 /* Device parameter to enable hardware Tx vector. */
97 #define MLX5_TX_VEC_EN "tx_vec_en"
98 
99 /* Device parameter to enable hardware Rx vector. */
100 #define MLX5_RX_VEC_EN "rx_vec_en"
101 
102 /* Allow L3 VXLAN flow creation. */
103 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
104 
105 /* Activate DV flow steering. */
106 #define MLX5_DV_FLOW_EN "dv_flow_en"
107 
108 /* Activate Netlink support in VF mode. */
109 #define MLX5_VF_NL_EN "vf_nl_en"
110 
111 /* Select port representors to instantiate. */
112 #define MLX5_REPRESENTOR "representor"
113 
114 #ifndef HAVE_IBV_MLX5_MOD_MPW
115 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
116 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
117 #endif
118 
119 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
120 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
121 #endif
122 
123 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
124 
125 /* Shared memory between primary and secondary processes. */
126 struct mlx5_shared_data *mlx5_shared_data;
127 
128 /* Spinlock for mlx5_shared_data allocation. */
129 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
130 
131 /** Driver-specific log messages type. */
132 int mlx5_logtype;
133 
134 /**
135  * Prepare shared data between primary and secondary process.
136  */
137 static void
138 mlx5_prepare_shared_data(void)
139 {
140 	const struct rte_memzone *mz;
141 
142 	rte_spinlock_lock(&mlx5_shared_data_lock);
143 	if (mlx5_shared_data == NULL) {
144 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
145 			/* Allocate shared memory. */
146 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
147 						 sizeof(*mlx5_shared_data),
148 						 SOCKET_ID_ANY, 0);
149 		} else {
150 			/* Lookup allocated shared memory. */
151 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
152 		}
153 		if (mz == NULL)
154 			rte_panic("Cannot allocate mlx5 shared data\n");
155 		mlx5_shared_data = mz->addr;
156 		/* Initialize shared data. */
157 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
158 			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
159 			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
160 		}
161 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
162 						mlx5_mr_mem_event_cb, NULL);
163 	}
164 	rte_spinlock_unlock(&mlx5_shared_data_lock);
165 }
166 
167 /**
168  * Retrieve integer value from environment variable.
169  *
170  * @param[in] name
171  *   Environment variable name.
172  *
173  * @return
174  *   Integer value, 0 if the variable is not set.
175  */
176 int
177 mlx5_getenv_int(const char *name)
178 {
179 	const char *val = getenv(name);
180 
181 	if (val == NULL)
182 		return 0;
183 	return atoi(val);
184 }
185 
186 /**
187  * Verbs callback to allocate a memory. This function should allocate the space
188  * according to the size provided residing inside a huge page.
189  * Please note that all allocation must respect the alignment from libmlx5
190  * (i.e. currently sysconf(_SC_PAGESIZE)).
191  *
192  * @param[in] size
193  *   The size in bytes of the memory to allocate.
194  * @param[in] data
195  *   A pointer to the callback data.
196  *
197  * @return
198  *   Allocated buffer, NULL otherwise and rte_errno is set.
199  */
200 static void *
201 mlx5_alloc_verbs_buf(size_t size, void *data)
202 {
203 	struct priv *priv = data;
204 	void *ret;
205 	size_t alignment = sysconf(_SC_PAGESIZE);
206 	unsigned int socket = SOCKET_ID_ANY;
207 
208 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
209 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
210 
211 		socket = ctrl->socket;
212 	} else if (priv->verbs_alloc_ctx.type ==
213 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
214 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
215 
216 		socket = ctrl->socket;
217 	}
218 	assert(data != NULL);
219 	ret = rte_malloc_socket(__func__, size, alignment, socket);
220 	if (!ret && size)
221 		rte_errno = ENOMEM;
222 	return ret;
223 }
224 
225 /**
226  * Verbs callback to free a memory.
227  *
228  * @param[in] ptr
229  *   A pointer to the memory to free.
230  * @param[in] data
231  *   A pointer to the callback data.
232  */
233 static void
234 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
235 {
236 	assert(data != NULL);
237 	rte_free(ptr);
238 }
239 
240 /**
241  * DPDK callback to close the device.
242  *
243  * Destroy all queues and objects, free memory.
244  *
245  * @param dev
246  *   Pointer to Ethernet device structure.
247  */
248 static void
249 mlx5_dev_close(struct rte_eth_dev *dev)
250 {
251 	struct priv *priv = dev->data->dev_private;
252 	unsigned int i;
253 	int ret;
254 
255 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
256 		dev->data->port_id,
257 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
258 	/* In case mlx5_dev_stop() has not been called. */
259 	mlx5_dev_interrupt_handler_uninstall(dev);
260 	mlx5_traffic_disable(dev);
261 	mlx5_flow_flush(dev, NULL);
262 	/* Prevent crashes when queues are still in use. */
263 	dev->rx_pkt_burst = removed_rx_burst;
264 	dev->tx_pkt_burst = removed_tx_burst;
265 	if (priv->rxqs != NULL) {
266 		/* XXX race condition if mlx5_rx_burst() is still running. */
267 		usleep(1000);
268 		for (i = 0; (i != priv->rxqs_n); ++i)
269 			mlx5_rxq_release(dev, i);
270 		priv->rxqs_n = 0;
271 		priv->rxqs = NULL;
272 	}
273 	if (priv->txqs != NULL) {
274 		/* XXX race condition if mlx5_tx_burst() is still running. */
275 		usleep(1000);
276 		for (i = 0; (i != priv->txqs_n); ++i)
277 			mlx5_txq_release(dev, i);
278 		priv->txqs_n = 0;
279 		priv->txqs = NULL;
280 	}
281 	mlx5_mprq_free_mp(dev);
282 	mlx5_mr_release(dev);
283 	if (priv->pd != NULL) {
284 		assert(priv->ctx != NULL);
285 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
286 		claim_zero(mlx5_glue->close_device(priv->ctx));
287 	} else
288 		assert(priv->ctx == NULL);
289 	if (priv->rss_conf.rss_key != NULL)
290 		rte_free(priv->rss_conf.rss_key);
291 	if (priv->reta_idx != NULL)
292 		rte_free(priv->reta_idx);
293 	if (priv->primary_socket)
294 		mlx5_socket_uninit(dev);
295 	if (priv->config.vf)
296 		mlx5_nl_mac_addr_flush(dev);
297 	if (priv->nl_socket_route >= 0)
298 		close(priv->nl_socket_route);
299 	if (priv->nl_socket_rdma >= 0)
300 		close(priv->nl_socket_rdma);
301 	if (priv->tcf_context)
302 		mlx5_flow_tcf_context_destroy(priv->tcf_context);
303 	ret = mlx5_hrxq_ibv_verify(dev);
304 	if (ret)
305 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
306 			dev->data->port_id);
307 	ret = mlx5_ind_table_ibv_verify(dev);
308 	if (ret)
309 		DRV_LOG(WARNING, "port %u some indirection table still remain",
310 			dev->data->port_id);
311 	ret = mlx5_rxq_ibv_verify(dev);
312 	if (ret)
313 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
314 			dev->data->port_id);
315 	ret = mlx5_rxq_verify(dev);
316 	if (ret)
317 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
318 			dev->data->port_id);
319 	ret = mlx5_txq_ibv_verify(dev);
320 	if (ret)
321 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
322 			dev->data->port_id);
323 	ret = mlx5_txq_verify(dev);
324 	if (ret)
325 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
326 			dev->data->port_id);
327 	ret = mlx5_flow_verify(dev);
328 	if (ret)
329 		DRV_LOG(WARNING, "port %u some flows still remain",
330 			dev->data->port_id);
331 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
332 		unsigned int c = 0;
333 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
334 		uint16_t port_id[i];
335 
336 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
337 		while (i--) {
338 			struct priv *opriv =
339 				rte_eth_devices[port_id[i]].data->dev_private;
340 
341 			if (!opriv ||
342 			    opriv->domain_id != priv->domain_id ||
343 			    &rte_eth_devices[port_id[i]] == dev)
344 				continue;
345 			++c;
346 		}
347 		if (!c)
348 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
349 	}
350 	memset(priv, 0, sizeof(*priv));
351 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
352 	/*
353 	 * Reset mac_addrs to NULL such that it is not freed as part of
354 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
355 	 * it is freed when dev_private is freed.
356 	 */
357 	dev->data->mac_addrs = NULL;
358 }
359 
360 const struct eth_dev_ops mlx5_dev_ops = {
361 	.dev_configure = mlx5_dev_configure,
362 	.dev_start = mlx5_dev_start,
363 	.dev_stop = mlx5_dev_stop,
364 	.dev_set_link_down = mlx5_set_link_down,
365 	.dev_set_link_up = mlx5_set_link_up,
366 	.dev_close = mlx5_dev_close,
367 	.promiscuous_enable = mlx5_promiscuous_enable,
368 	.promiscuous_disable = mlx5_promiscuous_disable,
369 	.allmulticast_enable = mlx5_allmulticast_enable,
370 	.allmulticast_disable = mlx5_allmulticast_disable,
371 	.link_update = mlx5_link_update,
372 	.stats_get = mlx5_stats_get,
373 	.stats_reset = mlx5_stats_reset,
374 	.xstats_get = mlx5_xstats_get,
375 	.xstats_reset = mlx5_xstats_reset,
376 	.xstats_get_names = mlx5_xstats_get_names,
377 	.dev_infos_get = mlx5_dev_infos_get,
378 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
379 	.vlan_filter_set = mlx5_vlan_filter_set,
380 	.rx_queue_setup = mlx5_rx_queue_setup,
381 	.tx_queue_setup = mlx5_tx_queue_setup,
382 	.rx_queue_release = mlx5_rx_queue_release,
383 	.tx_queue_release = mlx5_tx_queue_release,
384 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
385 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
386 	.mac_addr_remove = mlx5_mac_addr_remove,
387 	.mac_addr_add = mlx5_mac_addr_add,
388 	.mac_addr_set = mlx5_mac_addr_set,
389 	.set_mc_addr_list = mlx5_set_mc_addr_list,
390 	.mtu_set = mlx5_dev_set_mtu,
391 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
392 	.vlan_offload_set = mlx5_vlan_offload_set,
393 	.reta_update = mlx5_dev_rss_reta_update,
394 	.reta_query = mlx5_dev_rss_reta_query,
395 	.rss_hash_update = mlx5_rss_hash_update,
396 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
397 	.filter_ctrl = mlx5_dev_filter_ctrl,
398 	.rx_descriptor_status = mlx5_rx_descriptor_status,
399 	.tx_descriptor_status = mlx5_tx_descriptor_status,
400 	.rx_queue_count = mlx5_rx_queue_count,
401 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
402 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
403 	.is_removed = mlx5_is_removed,
404 };
405 
406 static const struct eth_dev_ops mlx5_dev_sec_ops = {
407 	.stats_get = mlx5_stats_get,
408 	.stats_reset = mlx5_stats_reset,
409 	.xstats_get = mlx5_xstats_get,
410 	.xstats_reset = mlx5_xstats_reset,
411 	.xstats_get_names = mlx5_xstats_get_names,
412 	.dev_infos_get = mlx5_dev_infos_get,
413 	.rx_descriptor_status = mlx5_rx_descriptor_status,
414 	.tx_descriptor_status = mlx5_tx_descriptor_status,
415 };
416 
417 /* Available operators in flow isolated mode. */
418 const struct eth_dev_ops mlx5_dev_ops_isolate = {
419 	.dev_configure = mlx5_dev_configure,
420 	.dev_start = mlx5_dev_start,
421 	.dev_stop = mlx5_dev_stop,
422 	.dev_set_link_down = mlx5_set_link_down,
423 	.dev_set_link_up = mlx5_set_link_up,
424 	.dev_close = mlx5_dev_close,
425 	.promiscuous_enable = mlx5_promiscuous_enable,
426 	.promiscuous_disable = mlx5_promiscuous_disable,
427 	.allmulticast_enable = mlx5_allmulticast_enable,
428 	.allmulticast_disable = mlx5_allmulticast_disable,
429 	.link_update = mlx5_link_update,
430 	.stats_get = mlx5_stats_get,
431 	.stats_reset = mlx5_stats_reset,
432 	.xstats_get = mlx5_xstats_get,
433 	.xstats_reset = mlx5_xstats_reset,
434 	.xstats_get_names = mlx5_xstats_get_names,
435 	.dev_infos_get = mlx5_dev_infos_get,
436 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
437 	.vlan_filter_set = mlx5_vlan_filter_set,
438 	.rx_queue_setup = mlx5_rx_queue_setup,
439 	.tx_queue_setup = mlx5_tx_queue_setup,
440 	.rx_queue_release = mlx5_rx_queue_release,
441 	.tx_queue_release = mlx5_tx_queue_release,
442 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
443 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
444 	.mac_addr_remove = mlx5_mac_addr_remove,
445 	.mac_addr_add = mlx5_mac_addr_add,
446 	.mac_addr_set = mlx5_mac_addr_set,
447 	.set_mc_addr_list = mlx5_set_mc_addr_list,
448 	.mtu_set = mlx5_dev_set_mtu,
449 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
450 	.vlan_offload_set = mlx5_vlan_offload_set,
451 	.filter_ctrl = mlx5_dev_filter_ctrl,
452 	.rx_descriptor_status = mlx5_rx_descriptor_status,
453 	.tx_descriptor_status = mlx5_tx_descriptor_status,
454 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
455 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
456 	.is_removed = mlx5_is_removed,
457 };
458 
459 /**
460  * Verify and store value for device argument.
461  *
462  * @param[in] key
463  *   Key argument to verify.
464  * @param[in] val
465  *   Value associated with key.
466  * @param opaque
467  *   User data.
468  *
469  * @return
470  *   0 on success, a negative errno value otherwise and rte_errno is set.
471  */
472 static int
473 mlx5_args_check(const char *key, const char *val, void *opaque)
474 {
475 	struct mlx5_dev_config *config = opaque;
476 	unsigned long tmp;
477 
478 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
479 	if (!strcmp(MLX5_REPRESENTOR, key))
480 		return 0;
481 	errno = 0;
482 	tmp = strtoul(val, NULL, 0);
483 	if (errno) {
484 		rte_errno = errno;
485 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
486 		return -rte_errno;
487 	}
488 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
489 		config->cqe_comp = !!tmp;
490 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
491 		config->cqe_pad = !!tmp;
492 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
493 		config->hw_padding = !!tmp;
494 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
495 		config->mprq.enabled = !!tmp;
496 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
497 		config->mprq.stride_num_n = tmp;
498 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
499 		config->mprq.max_memcpy_len = tmp;
500 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
501 		config->mprq.min_rxqs_num = tmp;
502 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
503 		config->txq_inline = tmp;
504 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
505 		config->txqs_inline = tmp;
506 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
507 		config->txqs_vec = tmp;
508 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
509 		config->mps = !!tmp;
510 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
511 		config->mpw_hdr_dseg = !!tmp;
512 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
513 		config->inline_max_packet_sz = tmp;
514 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
515 		config->tx_vec_en = !!tmp;
516 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
517 		config->rx_vec_en = !!tmp;
518 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
519 		config->l3_vxlan_en = !!tmp;
520 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
521 		config->vf_nl_en = !!tmp;
522 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
523 		config->dv_flow_en = !!tmp;
524 	} else {
525 		DRV_LOG(WARNING, "%s: unknown parameter", key);
526 		rte_errno = EINVAL;
527 		return -rte_errno;
528 	}
529 	return 0;
530 }
531 
532 /**
533  * Parse device parameters.
534  *
535  * @param config
536  *   Pointer to device configuration structure.
537  * @param devargs
538  *   Device arguments structure.
539  *
540  * @return
541  *   0 on success, a negative errno value otherwise and rte_errno is set.
542  */
543 static int
544 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
545 {
546 	const char **params = (const char *[]){
547 		MLX5_RXQ_CQE_COMP_EN,
548 		MLX5_RXQ_CQE_PAD_EN,
549 		MLX5_RXQ_PKT_PAD_EN,
550 		MLX5_RX_MPRQ_EN,
551 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
552 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
553 		MLX5_RXQS_MIN_MPRQ,
554 		MLX5_TXQ_INLINE,
555 		MLX5_TXQS_MIN_INLINE,
556 		MLX5_TXQS_MAX_VEC,
557 		MLX5_TXQ_MPW_EN,
558 		MLX5_TXQ_MPW_HDR_DSEG_EN,
559 		MLX5_TXQ_MAX_INLINE_LEN,
560 		MLX5_TX_VEC_EN,
561 		MLX5_RX_VEC_EN,
562 		MLX5_L3_VXLAN_EN,
563 		MLX5_VF_NL_EN,
564 		MLX5_DV_FLOW_EN,
565 		MLX5_REPRESENTOR,
566 		NULL,
567 	};
568 	struct rte_kvargs *kvlist;
569 	int ret = 0;
570 	int i;
571 
572 	if (devargs == NULL)
573 		return 0;
574 	/* Following UGLY cast is done to pass checkpatch. */
575 	kvlist = rte_kvargs_parse(devargs->args, params);
576 	if (kvlist == NULL)
577 		return 0;
578 	/* Process parameters. */
579 	for (i = 0; (params[i] != NULL); ++i) {
580 		if (rte_kvargs_count(kvlist, params[i])) {
581 			ret = rte_kvargs_process(kvlist, params[i],
582 						 mlx5_args_check, config);
583 			if (ret) {
584 				rte_errno = EINVAL;
585 				rte_kvargs_free(kvlist);
586 				return -rte_errno;
587 			}
588 		}
589 	}
590 	rte_kvargs_free(kvlist);
591 	return 0;
592 }
593 
594 static struct rte_pci_driver mlx5_driver;
595 
596 /*
597  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
598  * local resource used by both primary and secondary to avoid duplicate
599  * reservation.
600  * The space has to be available on both primary and secondary process,
601  * TXQ UAR maps to this area using fixed mmap w/o double check.
602  */
603 static void *uar_base;
604 
605 static int
606 find_lower_va_bound(const struct rte_memseg_list *msl,
607 		const struct rte_memseg *ms, void *arg)
608 {
609 	void **addr = arg;
610 
611 	if (msl->external)
612 		return 0;
613 	if (*addr == NULL)
614 		*addr = ms->addr;
615 	else
616 		*addr = RTE_MIN(*addr, ms->addr);
617 
618 	return 0;
619 }
620 
621 /**
622  * Reserve UAR address space for primary process.
623  *
624  * @param[in] dev
625  *   Pointer to Ethernet device.
626  *
627  * @return
628  *   0 on success, a negative errno value otherwise and rte_errno is set.
629  */
630 static int
631 mlx5_uar_init_primary(struct rte_eth_dev *dev)
632 {
633 	struct priv *priv = dev->data->dev_private;
634 	void *addr = (void *)0;
635 
636 	if (uar_base) { /* UAR address space mapped. */
637 		priv->uar_base = uar_base;
638 		return 0;
639 	}
640 	/* find out lower bound of hugepage segments */
641 	rte_memseg_walk(find_lower_va_bound, &addr);
642 
643 	/* keep distance to hugepages to minimize potential conflicts. */
644 	addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
645 	/* anonymous mmap, no real memory consumption. */
646 	addr = mmap(addr, MLX5_UAR_SIZE,
647 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
648 	if (addr == MAP_FAILED) {
649 		DRV_LOG(ERR,
650 			"port %u failed to reserve UAR address space, please"
651 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
652 			dev->data->port_id);
653 		rte_errno = ENOMEM;
654 		return -rte_errno;
655 	}
656 	/* Accept either same addr or a new addr returned from mmap if target
657 	 * range occupied.
658 	 */
659 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
660 		dev->data->port_id, addr);
661 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
662 	uar_base = addr; /* process local, don't reserve again. */
663 	return 0;
664 }
665 
666 /**
667  * Reserve UAR address space for secondary process, align with
668  * primary process.
669  *
670  * @param[in] dev
671  *   Pointer to Ethernet device.
672  *
673  * @return
674  *   0 on success, a negative errno value otherwise and rte_errno is set.
675  */
676 static int
677 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
678 {
679 	struct priv *priv = dev->data->dev_private;
680 	void *addr;
681 
682 	assert(priv->uar_base);
683 	if (uar_base) { /* already reserved. */
684 		assert(uar_base == priv->uar_base);
685 		return 0;
686 	}
687 	/* anonymous mmap, no real memory consumption. */
688 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
689 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
690 	if (addr == MAP_FAILED) {
691 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
692 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
693 		rte_errno = ENXIO;
694 		return -rte_errno;
695 	}
696 	if (priv->uar_base != addr) {
697 		DRV_LOG(ERR,
698 			"port %u UAR address %p size %llu occupied, please"
699 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
700 			" --base-virtaddr",
701 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
702 		rte_errno = ENXIO;
703 		return -rte_errno;
704 	}
705 	uar_base = addr; /* process local, don't reserve again */
706 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
707 		dev->data->port_id, addr);
708 	return 0;
709 }
710 
711 /**
712  * Spawn an Ethernet device from Verbs information.
713  *
714  * @param dpdk_dev
715  *   Backing DPDK device.
716  * @param ibv_dev
717  *   Verbs device.
718  * @param config
719  *   Device configuration parameters.
720  * @param[in] switch_info
721  *   Switch properties of Ethernet device.
722  *
723  * @return
724  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
725  *   is set. The following errors are defined:
726  *
727  *   EBUSY: device is not supposed to be spawned.
728  *   EEXIST: device is already spawned
729  */
730 static struct rte_eth_dev *
731 mlx5_dev_spawn(struct rte_device *dpdk_dev,
732 	       struct ibv_device *ibv_dev,
733 	       struct mlx5_dev_config config,
734 	       const struct mlx5_switch_info *switch_info)
735 {
736 	struct ibv_context *ctx = NULL;
737 	struct ibv_device_attr_ex attr;
738 	struct ibv_port_attr port_attr;
739 	struct ibv_pd *pd = NULL;
740 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
741 	struct rte_eth_dev *eth_dev = NULL;
742 	struct priv *priv = NULL;
743 	int err = 0;
744 	unsigned int hw_padding = 0;
745 	unsigned int mps;
746 	unsigned int cqe_comp;
747 	unsigned int cqe_pad = 0;
748 	unsigned int tunnel_en = 0;
749 	unsigned int mpls_en = 0;
750 	unsigned int swp = 0;
751 	unsigned int mprq = 0;
752 	unsigned int mprq_min_stride_size_n = 0;
753 	unsigned int mprq_max_stride_size_n = 0;
754 	unsigned int mprq_min_stride_num_n = 0;
755 	unsigned int mprq_max_stride_num_n = 0;
756 	struct ether_addr mac;
757 	char name[RTE_ETH_NAME_MAX_LEN];
758 	int own_domain_id = 0;
759 	uint16_t port_id;
760 	unsigned int i;
761 
762 	/* Determine if this port representor is supposed to be spawned. */
763 	if (switch_info->representor && dpdk_dev->devargs) {
764 		struct rte_eth_devargs eth_da;
765 
766 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
767 		if (err) {
768 			rte_errno = -err;
769 			DRV_LOG(ERR, "failed to process device arguments: %s",
770 				strerror(rte_errno));
771 			return NULL;
772 		}
773 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
774 			if (eth_da.representor_ports[i] ==
775 			    (uint16_t)switch_info->port_name)
776 				break;
777 		if (i == eth_da.nb_representor_ports) {
778 			rte_errno = EBUSY;
779 			return NULL;
780 		}
781 	}
782 	/* Build device name. */
783 	if (!switch_info->representor)
784 		rte_strlcpy(name, dpdk_dev->name, sizeof(name));
785 	else
786 		snprintf(name, sizeof(name), "%s_representor_%u",
787 			 dpdk_dev->name, switch_info->port_name);
788 	/* check if the device is already spawned */
789 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
790 		rte_errno = EEXIST;
791 		return NULL;
792 	}
793 	/* Prepare shared data between primary and secondary process. */
794 	mlx5_prepare_shared_data();
795 	errno = 0;
796 	ctx = mlx5_glue->dv_open_device(ibv_dev);
797 	if (ctx) {
798 		config.devx = 1;
799 		DRV_LOG(DEBUG, "DEVX is supported");
800 	} else {
801 		ctx = mlx5_glue->open_device(ibv_dev);
802 		if (!ctx) {
803 			rte_errno = errno ? errno : ENODEV;
804 			return NULL;
805 		}
806 	}
807 #ifdef HAVE_IBV_MLX5_MOD_SWP
808 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
809 #endif
810 	/*
811 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
812 	 * as all ConnectX-5 devices.
813 	 */
814 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
815 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
816 #endif
817 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
818 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
819 #endif
820 	mlx5_glue->dv_query_device(ctx, &dv_attr);
821 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
822 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
823 			DRV_LOG(DEBUG, "enhanced MPW is supported");
824 			mps = MLX5_MPW_ENHANCED;
825 		} else {
826 			DRV_LOG(DEBUG, "MPW is supported");
827 			mps = MLX5_MPW;
828 		}
829 	} else {
830 		DRV_LOG(DEBUG, "MPW isn't supported");
831 		mps = MLX5_MPW_DISABLED;
832 	}
833 #ifdef HAVE_IBV_MLX5_MOD_SWP
834 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
835 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
836 	DRV_LOG(DEBUG, "SWP support: %u", swp);
837 #endif
838 	config.swp = !!swp;
839 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
840 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
841 		struct mlx5dv_striding_rq_caps mprq_caps =
842 			dv_attr.striding_rq_caps;
843 
844 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
845 			mprq_caps.min_single_stride_log_num_of_bytes);
846 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
847 			mprq_caps.max_single_stride_log_num_of_bytes);
848 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
849 			mprq_caps.min_single_wqe_log_num_of_strides);
850 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
851 			mprq_caps.max_single_wqe_log_num_of_strides);
852 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
853 			mprq_caps.supported_qpts);
854 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
855 		mprq = 1;
856 		mprq_min_stride_size_n =
857 			mprq_caps.min_single_stride_log_num_of_bytes;
858 		mprq_max_stride_size_n =
859 			mprq_caps.max_single_stride_log_num_of_bytes;
860 		mprq_min_stride_num_n =
861 			mprq_caps.min_single_wqe_log_num_of_strides;
862 		mprq_max_stride_num_n =
863 			mprq_caps.max_single_wqe_log_num_of_strides;
864 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
865 						   mprq_min_stride_num_n);
866 	}
867 #endif
868 	if (RTE_CACHE_LINE_SIZE == 128 &&
869 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
870 		cqe_comp = 0;
871 	else
872 		cqe_comp = 1;
873 	config.cqe_comp = cqe_comp;
874 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
875 	/* Whether device supports 128B Rx CQE padding. */
876 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
877 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
878 #endif
879 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
880 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
881 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
882 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
883 			     (dv_attr.tunnel_offloads_caps &
884 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
885 	}
886 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
887 		tunnel_en ? "" : "not ");
888 #else
889 	DRV_LOG(WARNING,
890 		"tunnel offloading disabled due to old OFED/rdma-core version");
891 #endif
892 	config.tunnel_en = tunnel_en;
893 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
894 	mpls_en = ((dv_attr.tunnel_offloads_caps &
895 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
896 		   (dv_attr.tunnel_offloads_caps &
897 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
898 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
899 		mpls_en ? "" : "not ");
900 #else
901 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
902 		" old OFED/rdma-core version or firmware configuration");
903 #endif
904 	config.mpls_en = mpls_en;
905 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr);
906 	if (err) {
907 		DEBUG("ibv_query_device_ex() failed");
908 		goto error;
909 	}
910 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
911 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
912 		eth_dev = rte_eth_dev_attach_secondary(name);
913 		if (eth_dev == NULL) {
914 			DRV_LOG(ERR, "can not attach rte ethdev");
915 			rte_errno = ENOMEM;
916 			err = rte_errno;
917 			goto error;
918 		}
919 		eth_dev->device = dpdk_dev;
920 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
921 		err = mlx5_uar_init_secondary(eth_dev);
922 		if (err) {
923 			err = rte_errno;
924 			goto error;
925 		}
926 		/* Receive command fd from primary process */
927 		err = mlx5_socket_connect(eth_dev);
928 		if (err < 0) {
929 			err = rte_errno;
930 			goto error;
931 		}
932 		/* Remap UAR for Tx queues. */
933 		err = mlx5_tx_uar_remap(eth_dev, err);
934 		if (err) {
935 			err = rte_errno;
936 			goto error;
937 		}
938 		/*
939 		 * Ethdev pointer is still required as input since
940 		 * the primary device is not accessible from the
941 		 * secondary process.
942 		 */
943 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
944 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
945 		claim_zero(mlx5_glue->close_device(ctx));
946 		return eth_dev;
947 	}
948 	/* Check port status. */
949 	err = mlx5_glue->query_port(ctx, 1, &port_attr);
950 	if (err) {
951 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
952 		goto error;
953 	}
954 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
955 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
956 		err = EINVAL;
957 		goto error;
958 	}
959 	if (port_attr.state != IBV_PORT_ACTIVE)
960 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
961 			mlx5_glue->port_state_str(port_attr.state),
962 			port_attr.state);
963 	/* Allocate protection domain. */
964 	pd = mlx5_glue->alloc_pd(ctx);
965 	if (pd == NULL) {
966 		DRV_LOG(ERR, "PD allocation failure");
967 		err = ENOMEM;
968 		goto error;
969 	}
970 	priv = rte_zmalloc("ethdev private structure",
971 			   sizeof(*priv),
972 			   RTE_CACHE_LINE_SIZE);
973 	if (priv == NULL) {
974 		DRV_LOG(ERR, "priv allocation failure");
975 		err = ENOMEM;
976 		goto error;
977 	}
978 	priv->ctx = ctx;
979 	strncpy(priv->ibdev_name, priv->ctx->device->name,
980 		sizeof(priv->ibdev_name));
981 	strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
982 		sizeof(priv->ibdev_path));
983 	priv->device_attr = attr;
984 	priv->pd = pd;
985 	priv->mtu = ETHER_MTU;
986 #ifndef RTE_ARCH_64
987 	/* Initialize UAR access locks for 32bit implementations. */
988 	rte_spinlock_init(&priv->uar_lock_cq);
989 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
990 		rte_spinlock_init(&priv->uar_lock[i]);
991 #endif
992 	/* Some internal functions rely on Netlink sockets, open them now. */
993 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
994 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
995 	priv->nl_sn = 0;
996 	priv->representor = !!switch_info->representor;
997 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
998 	priv->representor_id =
999 		switch_info->representor ? switch_info->port_name : -1;
1000 	/*
1001 	 * Look for sibling devices in order to reuse their switch domain
1002 	 * if any, otherwise allocate one.
1003 	 */
1004 	i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
1005 	if (i > 0) {
1006 		uint16_t port_id[i];
1007 
1008 		i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
1009 		while (i--) {
1010 			const struct priv *opriv =
1011 				rte_eth_devices[port_id[i]].data->dev_private;
1012 
1013 			if (!opriv ||
1014 			    opriv->domain_id ==
1015 			    RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1016 				continue;
1017 			priv->domain_id = opriv->domain_id;
1018 			break;
1019 		}
1020 	}
1021 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1022 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1023 		if (err) {
1024 			err = rte_errno;
1025 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1026 				strerror(rte_errno));
1027 			goto error;
1028 		}
1029 		own_domain_id = 1;
1030 	}
1031 	err = mlx5_args(&config, dpdk_dev->devargs);
1032 	if (err) {
1033 		err = rte_errno;
1034 		DRV_LOG(ERR, "failed to process device arguments: %s",
1035 			strerror(rte_errno));
1036 		goto error;
1037 	}
1038 	config.hw_csum = !!(attr.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
1039 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1040 		(config.hw_csum ? "" : "not "));
1041 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1042 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1043 	DRV_LOG(DEBUG, "counters are not supported");
1044 #endif
1045 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1046 	if (config.dv_flow_en) {
1047 		DRV_LOG(WARNING, "DV flow is not supported");
1048 		config.dv_flow_en = 0;
1049 	}
1050 #endif
1051 	config.ind_table_max_size =
1052 		attr.rss_caps.max_rwq_indirection_table_size;
1053 	/*
1054 	 * Remove this check once DPDK supports larger/variable
1055 	 * indirection tables.
1056 	 */
1057 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1058 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1059 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1060 		config.ind_table_max_size);
1061 	config.hw_vlan_strip = !!(attr.raw_packet_caps &
1062 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1063 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1064 		(config.hw_vlan_strip ? "" : "not "));
1065 	config.hw_fcs_strip = !!(attr.raw_packet_caps &
1066 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1067 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1068 		(config.hw_fcs_strip ? "" : "not "));
1069 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1070 	hw_padding = !!attr.rx_pad_end_addr_align;
1071 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1072 	hw_padding = !!(attr.device_cap_flags_ex &
1073 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1074 #endif
1075 	if (config.hw_padding && !hw_padding) {
1076 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1077 		config.hw_padding = 0;
1078 	} else if (config.hw_padding) {
1079 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1080 	}
1081 	config.tso = (attr.tso_caps.max_tso > 0 &&
1082 		      (attr.tso_caps.supported_qpts &
1083 		       (1 << IBV_QPT_RAW_PACKET)));
1084 	if (config.tso)
1085 		config.tso_max_payload_sz = attr.tso_caps.max_tso;
1086 	/*
1087 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1088 	 * by default.
1089 	 */
1090 	if (config.mps == MLX5_ARG_UNSET)
1091 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1092 							  MLX5_MPW_DISABLED;
1093 	else
1094 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1095 	DRV_LOG(INFO, "%sMPS is %s",
1096 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1097 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1098 	if (config.cqe_comp && !cqe_comp) {
1099 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1100 		config.cqe_comp = 0;
1101 	}
1102 	if (config.cqe_pad && !cqe_pad) {
1103 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1104 		config.cqe_pad = 0;
1105 	} else if (config.cqe_pad) {
1106 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1107 	}
1108 	if (config.mprq.enabled && mprq) {
1109 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1110 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1111 			config.mprq.stride_num_n =
1112 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1113 					mprq_min_stride_num_n);
1114 			DRV_LOG(WARNING,
1115 				"the number of strides"
1116 				" for Multi-Packet RQ is out of range,"
1117 				" setting default value (%u)",
1118 				1 << config.mprq.stride_num_n);
1119 		}
1120 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1121 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1122 	} else if (config.mprq.enabled && !mprq) {
1123 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1124 		config.mprq.enabled = 0;
1125 	}
1126 	eth_dev = rte_eth_dev_allocate(name);
1127 	if (eth_dev == NULL) {
1128 		DRV_LOG(ERR, "can not allocate rte ethdev");
1129 		err = ENOMEM;
1130 		goto error;
1131 	}
1132 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1133 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1134 	if (priv->representor) {
1135 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1136 		eth_dev->data->representor_id = priv->representor_id;
1137 	}
1138 	eth_dev->data->dev_private = priv;
1139 	priv->dev_data = eth_dev->data;
1140 	eth_dev->data->mac_addrs = priv->mac;
1141 	eth_dev->device = dpdk_dev;
1142 	err = mlx5_uar_init_primary(eth_dev);
1143 	if (err) {
1144 		err = rte_errno;
1145 		goto error;
1146 	}
1147 	/* Configure the first MAC address by default. */
1148 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1149 		DRV_LOG(ERR,
1150 			"port %u cannot get MAC address, is mlx5_en"
1151 			" loaded? (errno: %s)",
1152 			eth_dev->data->port_id, strerror(rte_errno));
1153 		err = ENODEV;
1154 		goto error;
1155 	}
1156 	DRV_LOG(INFO,
1157 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1158 		eth_dev->data->port_id,
1159 		mac.addr_bytes[0], mac.addr_bytes[1],
1160 		mac.addr_bytes[2], mac.addr_bytes[3],
1161 		mac.addr_bytes[4], mac.addr_bytes[5]);
1162 #ifndef NDEBUG
1163 	{
1164 		char ifname[IF_NAMESIZE];
1165 
1166 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1167 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1168 				eth_dev->data->port_id, ifname);
1169 		else
1170 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1171 				eth_dev->data->port_id);
1172 	}
1173 #endif
1174 	/* Get actual MTU if possible. */
1175 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1176 	if (err) {
1177 		err = rte_errno;
1178 		goto error;
1179 	}
1180 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1181 		priv->mtu);
1182 	/* Initialize burst functions to prevent crashes before link-up. */
1183 	eth_dev->rx_pkt_burst = removed_rx_burst;
1184 	eth_dev->tx_pkt_burst = removed_tx_burst;
1185 	eth_dev->dev_ops = &mlx5_dev_ops;
1186 	/* Register MAC address. */
1187 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1188 	if (config.vf && config.vf_nl_en)
1189 		mlx5_nl_mac_addr_sync(eth_dev);
1190 	priv->tcf_context = mlx5_flow_tcf_context_create();
1191 	if (!priv->tcf_context) {
1192 		err = -rte_errno;
1193 		DRV_LOG(WARNING,
1194 			"flow rules relying on switch offloads will not be"
1195 			" supported: cannot open libmnl socket: %s",
1196 			strerror(rte_errno));
1197 	} else {
1198 		struct rte_flow_error error;
1199 		unsigned int ifindex = mlx5_ifindex(eth_dev);
1200 
1201 		if (!ifindex) {
1202 			err = -rte_errno;
1203 			error.message =
1204 				"cannot retrieve network interface index";
1205 		} else {
1206 			err = mlx5_flow_tcf_init(priv->tcf_context,
1207 						 ifindex, &error);
1208 		}
1209 		if (err) {
1210 			DRV_LOG(WARNING,
1211 				"flow rules relying on switch offloads will"
1212 				" not be supported: %s: %s",
1213 				error.message, strerror(rte_errno));
1214 			mlx5_flow_tcf_context_destroy(priv->tcf_context);
1215 			priv->tcf_context = NULL;
1216 		}
1217 	}
1218 	TAILQ_INIT(&priv->flows);
1219 	TAILQ_INIT(&priv->ctrl_flows);
1220 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1221 	struct mlx5dv_ctx_allocators alctr = {
1222 		.alloc = &mlx5_alloc_verbs_buf,
1223 		.free = &mlx5_free_verbs_buf,
1224 		.data = priv,
1225 	};
1226 	mlx5_glue->dv_set_context_attr(ctx, MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1227 				       (void *)((uintptr_t)&alctr));
1228 	/* Bring Ethernet device up. */
1229 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1230 		eth_dev->data->port_id);
1231 	mlx5_set_link_up(eth_dev);
1232 	/*
1233 	 * Even though the interrupt handler is not installed yet,
1234 	 * interrupts will still trigger on the asyn_fd from
1235 	 * Verbs context returned by ibv_open_device().
1236 	 */
1237 	mlx5_link_update(eth_dev, 0);
1238 	/* Store device configuration on private structure. */
1239 	priv->config = config;
1240 	/* Supported Verbs flow priority number detection. */
1241 	err = mlx5_flow_discover_priorities(eth_dev);
1242 	if (err < 0)
1243 		goto error;
1244 	priv->config.flow_prio = err;
1245 	/*
1246 	 * Once the device is added to the list of memory event
1247 	 * callback, its global MR cache table cannot be expanded
1248 	 * on the fly because of deadlock. If it overflows, lookup
1249 	 * should be done by searching MR list linearly, which is slow.
1250 	 */
1251 	err = mlx5_mr_btree_init(&priv->mr.cache,
1252 				 MLX5_MR_BTREE_CACHE_N * 2,
1253 				 eth_dev->device->numa_node);
1254 	if (err) {
1255 		err = rte_errno;
1256 		goto error;
1257 	}
1258 	/* Add device to memory callback list. */
1259 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1260 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1261 			 priv, mem_event_cb);
1262 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1263 	return eth_dev;
1264 error:
1265 	if (priv) {
1266 		if (priv->nl_socket_route >= 0)
1267 			close(priv->nl_socket_route);
1268 		if (priv->nl_socket_rdma >= 0)
1269 			close(priv->nl_socket_rdma);
1270 		if (priv->tcf_context)
1271 			mlx5_flow_tcf_context_destroy(priv->tcf_context);
1272 		if (own_domain_id)
1273 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1274 		rte_free(priv);
1275 		if (eth_dev != NULL)
1276 			eth_dev->data->dev_private = NULL;
1277 	}
1278 	if (pd)
1279 		claim_zero(mlx5_glue->dealloc_pd(pd));
1280 	if (eth_dev != NULL) {
1281 		/* mac_addrs must not be freed alone because part of dev_private */
1282 		eth_dev->data->mac_addrs = NULL;
1283 		rte_eth_dev_release_port(eth_dev);
1284 	}
1285 	if (ctx)
1286 		claim_zero(mlx5_glue->close_device(ctx));
1287 	assert(err > 0);
1288 	rte_errno = err;
1289 	return NULL;
1290 }
1291 
1292 /** Data associated with devices to spawn. */
1293 struct mlx5_dev_spawn_data {
1294 	unsigned int ifindex; /**< Network interface index. */
1295 	struct mlx5_switch_info info; /**< Switch information. */
1296 	struct ibv_device *ibv_dev; /**< Associated IB device. */
1297 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
1298 };
1299 
1300 /**
1301  * Comparison callback to sort device data.
1302  *
1303  * This is meant to be used with qsort().
1304  *
1305  * @param a[in]
1306  *   Pointer to pointer to first data object.
1307  * @param b[in]
1308  *   Pointer to pointer to second data object.
1309  *
1310  * @return
1311  *   0 if both objects are equal, less than 0 if the first argument is less
1312  *   than the second, greater than 0 otherwise.
1313  */
1314 static int
1315 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1316 {
1317 	const struct mlx5_switch_info *si_a =
1318 		&((const struct mlx5_dev_spawn_data *)a)->info;
1319 	const struct mlx5_switch_info *si_b =
1320 		&((const struct mlx5_dev_spawn_data *)b)->info;
1321 	int ret;
1322 
1323 	/* Master device first. */
1324 	ret = si_b->master - si_a->master;
1325 	if (ret)
1326 		return ret;
1327 	/* Then representor devices. */
1328 	ret = si_b->representor - si_a->representor;
1329 	if (ret)
1330 		return ret;
1331 	/* Unidentified devices come last in no specific order. */
1332 	if (!si_a->representor)
1333 		return 0;
1334 	/* Order representors by name. */
1335 	return si_a->port_name - si_b->port_name;
1336 }
1337 
1338 /**
1339  * DPDK callback to register a PCI device.
1340  *
1341  * This function spawns Ethernet devices out of a given PCI device.
1342  *
1343  * @param[in] pci_drv
1344  *   PCI driver structure (mlx5_driver).
1345  * @param[in] pci_dev
1346  *   PCI device information.
1347  *
1348  * @return
1349  *   0 on success, a negative errno value otherwise and rte_errno is set.
1350  */
1351 static int
1352 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1353 	       struct rte_pci_device *pci_dev)
1354 {
1355 	struct ibv_device **ibv_list;
1356 	unsigned int n = 0;
1357 	struct mlx5_dev_config dev_config;
1358 	int ret;
1359 
1360 	assert(pci_drv == &mlx5_driver);
1361 	errno = 0;
1362 	ibv_list = mlx5_glue->get_device_list(&ret);
1363 	if (!ibv_list) {
1364 		rte_errno = errno ? errno : ENOSYS;
1365 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1366 		return -rte_errno;
1367 	}
1368 
1369 	struct ibv_device *ibv_match[ret + 1];
1370 
1371 	while (ret-- > 0) {
1372 		struct rte_pci_addr pci_addr;
1373 
1374 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1375 		if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1376 			continue;
1377 		if (pci_dev->addr.domain != pci_addr.domain ||
1378 		    pci_dev->addr.bus != pci_addr.bus ||
1379 		    pci_dev->addr.devid != pci_addr.devid ||
1380 		    pci_dev->addr.function != pci_addr.function)
1381 			continue;
1382 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1383 			ibv_list[ret]->name);
1384 		ibv_match[n++] = ibv_list[ret];
1385 	}
1386 	ibv_match[n] = NULL;
1387 
1388 	struct mlx5_dev_spawn_data list[n];
1389 	int nl_route = n ? mlx5_nl_init(NETLINK_ROUTE) : -1;
1390 	int nl_rdma = n ? mlx5_nl_init(NETLINK_RDMA) : -1;
1391 	unsigned int i;
1392 	unsigned int u;
1393 
1394 	/*
1395 	 * The existence of several matching entries (n > 1) means port
1396 	 * representors have been instantiated. No existing Verbs call nor
1397 	 * /sys entries can tell them apart, this can only be done through
1398 	 * Netlink calls assuming kernel drivers are recent enough to
1399 	 * support them.
1400 	 *
1401 	 * In the event of identification failure through Netlink, try again
1402 	 * through sysfs, then either:
1403 	 *
1404 	 * 1. No device matches (n == 0), complain and bail out.
1405 	 * 2. A single IB device matches (n == 1) and is not a representor,
1406 	 *    assume no switch support.
1407 	 * 3. Otherwise no safe assumptions can be made; complain louder and
1408 	 *    bail out.
1409 	 */
1410 	for (i = 0; i != n; ++i) {
1411 		list[i].ibv_dev = ibv_match[i];
1412 		list[i].eth_dev = NULL;
1413 		if (nl_rdma < 0)
1414 			list[i].ifindex = 0;
1415 		else
1416 			list[i].ifindex = mlx5_nl_ifindex
1417 				(nl_rdma, list[i].ibv_dev->name);
1418 		if (nl_route < 0 ||
1419 		    !list[i].ifindex ||
1420 		    mlx5_nl_switch_info(nl_route, list[i].ifindex,
1421 					&list[i].info) ||
1422 		    ((!list[i].info.representor && !list[i].info.master) &&
1423 		     mlx5_sysfs_switch_info(list[i].ifindex, &list[i].info))) {
1424 			list[i].ifindex = 0;
1425 			memset(&list[i].info, 0, sizeof(list[i].info));
1426 			continue;
1427 		}
1428 	}
1429 	if (nl_rdma >= 0)
1430 		close(nl_rdma);
1431 	if (nl_route >= 0)
1432 		close(nl_route);
1433 	/* Count unidentified devices. */
1434 	for (u = 0, i = 0; i != n; ++i)
1435 		if (!list[i].info.master && !list[i].info.representor)
1436 			++u;
1437 	if (u) {
1438 		if (n == 1 && u == 1) {
1439 			/* Case #2. */
1440 			DRV_LOG(INFO, "no switch support detected");
1441 		} else {
1442 			/* Case #3. */
1443 			DRV_LOG(ERR,
1444 				"unable to tell which of the matching devices"
1445 				" is the master (lack of kernel support?)");
1446 			n = 0;
1447 		}
1448 	}
1449 	/*
1450 	 * Sort list to probe devices in natural order for users convenience
1451 	 * (i.e. master first, then representors from lowest to highest ID).
1452 	 */
1453 	if (n)
1454 		qsort(list, n, sizeof(*list), mlx5_dev_spawn_data_cmp);
1455 	/* Default configuration. */
1456 	dev_config = (struct mlx5_dev_config){
1457 		.hw_padding = 0,
1458 		.mps = MLX5_ARG_UNSET,
1459 		.tx_vec_en = 1,
1460 		.rx_vec_en = 1,
1461 		.txq_inline = MLX5_ARG_UNSET,
1462 		.txqs_inline = MLX5_ARG_UNSET,
1463 		.txqs_vec = MLX5_ARG_UNSET,
1464 		.inline_max_packet_sz = MLX5_ARG_UNSET,
1465 		.vf_nl_en = 1,
1466 		.mprq = {
1467 			.enabled = 0, /* Disabled by default. */
1468 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1469 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1470 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1471 		},
1472 	};
1473 	/* Device speicific configuration. */
1474 	switch (pci_dev->id.device_id) {
1475 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1476 		dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1477 		break;
1478 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1479 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1480 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1481 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1482 		dev_config.vf = 1;
1483 		break;
1484 	default:
1485 		break;
1486 	}
1487 	/* Set architecture-dependent default value if unset. */
1488 	if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1489 		dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1490 	for (i = 0; i != n; ++i) {
1491 		uint32_t restore;
1492 
1493 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1494 						 list[i].ibv_dev, dev_config,
1495 						 &list[i].info);
1496 		if (!list[i].eth_dev) {
1497 			if (rte_errno != EBUSY && rte_errno != EEXIST)
1498 				break;
1499 			/* Device is disabled or already spawned. Ignore it. */
1500 			continue;
1501 		}
1502 		restore = list[i].eth_dev->data->dev_flags;
1503 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1504 		/* Restore non-PCI flags cleared by the above call. */
1505 		list[i].eth_dev->data->dev_flags |= restore;
1506 		rte_eth_dev_probing_finish(list[i].eth_dev);
1507 	}
1508 	mlx5_glue->free_device_list(ibv_list);
1509 	if (!n) {
1510 		DRV_LOG(WARNING,
1511 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1512 			" are kernel drivers loaded?",
1513 			pci_dev->addr.domain, pci_dev->addr.bus,
1514 			pci_dev->addr.devid, pci_dev->addr.function);
1515 		rte_errno = ENOENT;
1516 		ret = -rte_errno;
1517 	} else if (i != n) {
1518 		DRV_LOG(ERR,
1519 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1520 			" encountering an error: %s",
1521 			pci_dev->addr.domain, pci_dev->addr.bus,
1522 			pci_dev->addr.devid, pci_dev->addr.function,
1523 			strerror(rte_errno));
1524 		ret = -rte_errno;
1525 		/* Roll back. */
1526 		while (i--) {
1527 			if (!list[i].eth_dev)
1528 				continue;
1529 			mlx5_dev_close(list[i].eth_dev);
1530 			/* mac_addrs must not be freed because in dev_private */
1531 			list[i].eth_dev->data->mac_addrs = NULL;
1532 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1533 		}
1534 		/* Restore original error. */
1535 		rte_errno = -ret;
1536 	} else {
1537 		ret = 0;
1538 	}
1539 	return ret;
1540 }
1541 
1542 /**
1543  * DPDK callback to remove a PCI device.
1544  *
1545  * This function removes all Ethernet devices belong to a given PCI device.
1546  *
1547  * @param[in] pci_dev
1548  *   Pointer to the PCI device.
1549  *
1550  * @return
1551  *   0 on success, the function cannot fail.
1552  */
1553 static int
1554 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1555 {
1556 	uint16_t port_id;
1557 	struct rte_eth_dev *port;
1558 
1559 	for (port_id = 0; port_id < RTE_MAX_ETHPORTS; port_id++) {
1560 		port = &rte_eth_devices[port_id];
1561 		if (port->state != RTE_ETH_DEV_UNUSED &&
1562 				port->device == &pci_dev->device)
1563 			rte_eth_dev_close(port_id);
1564 	}
1565 	return 0;
1566 }
1567 
1568 static const struct rte_pci_id mlx5_pci_id_map[] = {
1569 	{
1570 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1571 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1572 	},
1573 	{
1574 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1575 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1576 	},
1577 	{
1578 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1579 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1580 	},
1581 	{
1582 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1583 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1584 	},
1585 	{
1586 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1587 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1588 	},
1589 	{
1590 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1591 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1592 	},
1593 	{
1594 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1595 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1596 	},
1597 	{
1598 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1599 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1600 	},
1601 	{
1602 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1603 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1604 	},
1605 	{
1606 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1607 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
1608 	},
1609 	{
1610 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1611 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
1612 	},
1613 	{
1614 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1615 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
1616 	},
1617 	{
1618 		.vendor_id = 0
1619 	}
1620 };
1621 
1622 static struct rte_pci_driver mlx5_driver = {
1623 	.driver = {
1624 		.name = MLX5_DRIVER_NAME
1625 	},
1626 	.id_table = mlx5_pci_id_map,
1627 	.probe = mlx5_pci_probe,
1628 	.remove = mlx5_pci_remove,
1629 	.drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
1630 		      RTE_PCI_DRV_PROBE_AGAIN),
1631 };
1632 
1633 #ifdef RTE_IBVERBS_LINK_DLOPEN
1634 
1635 /**
1636  * Suffix RTE_EAL_PMD_PATH with "-glue".
1637  *
1638  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1639  * suffixing its last component.
1640  *
1641  * @param buf[out]
1642  *   Output buffer, should be large enough otherwise NULL is returned.
1643  * @param size
1644  *   Size of @p out.
1645  *
1646  * @return
1647  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1648  */
1649 static char *
1650 mlx5_glue_path(char *buf, size_t size)
1651 {
1652 	static const char *const bad[] = { "/", ".", "..", NULL };
1653 	const char *path = RTE_EAL_PMD_PATH;
1654 	size_t len = strlen(path);
1655 	size_t off;
1656 	int i;
1657 
1658 	while (len && path[len - 1] == '/')
1659 		--len;
1660 	for (off = len; off && path[off - 1] != '/'; --off)
1661 		;
1662 	for (i = 0; bad[i]; ++i)
1663 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1664 			goto error;
1665 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1666 	if (i == -1 || (size_t)i >= size)
1667 		goto error;
1668 	return buf;
1669 error:
1670 	DRV_LOG(ERR,
1671 		"unable to append \"-glue\" to last component of"
1672 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1673 		" please re-configure DPDK");
1674 	return NULL;
1675 }
1676 
1677 /**
1678  * Initialization routine for run-time dependency on rdma-core.
1679  */
1680 static int
1681 mlx5_glue_init(void)
1682 {
1683 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1684 	const char *path[] = {
1685 		/*
1686 		 * A basic security check is necessary before trusting
1687 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1688 		 */
1689 		(geteuid() == getuid() && getegid() == getgid() ?
1690 		 getenv("MLX5_GLUE_PATH") : NULL),
1691 		/*
1692 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1693 		 * variant, otherwise let dlopen() look up libraries on its
1694 		 * own.
1695 		 */
1696 		(*RTE_EAL_PMD_PATH ?
1697 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1698 	};
1699 	unsigned int i = 0;
1700 	void *handle = NULL;
1701 	void **sym;
1702 	const char *dlmsg;
1703 
1704 	while (!handle && i != RTE_DIM(path)) {
1705 		const char *end;
1706 		size_t len;
1707 		int ret;
1708 
1709 		if (!path[i]) {
1710 			++i;
1711 			continue;
1712 		}
1713 		end = strpbrk(path[i], ":;");
1714 		if (!end)
1715 			end = path[i] + strlen(path[i]);
1716 		len = end - path[i];
1717 		ret = 0;
1718 		do {
1719 			char name[ret + 1];
1720 
1721 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1722 				       (int)len, path[i],
1723 				       (!len || *(end - 1) == '/') ? "" : "/");
1724 			if (ret == -1)
1725 				break;
1726 			if (sizeof(name) != (size_t)ret + 1)
1727 				continue;
1728 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1729 				name);
1730 			handle = dlopen(name, RTLD_LAZY);
1731 			break;
1732 		} while (1);
1733 		path[i] = end + 1;
1734 		if (!*end)
1735 			++i;
1736 	}
1737 	if (!handle) {
1738 		rte_errno = EINVAL;
1739 		dlmsg = dlerror();
1740 		if (dlmsg)
1741 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1742 		goto glue_error;
1743 	}
1744 	sym = dlsym(handle, "mlx5_glue");
1745 	if (!sym || !*sym) {
1746 		rte_errno = EINVAL;
1747 		dlmsg = dlerror();
1748 		if (dlmsg)
1749 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1750 		goto glue_error;
1751 	}
1752 	mlx5_glue = *sym;
1753 	return 0;
1754 glue_error:
1755 	if (handle)
1756 		dlclose(handle);
1757 	DRV_LOG(WARNING,
1758 		"cannot initialize PMD due to missing run-time dependency on"
1759 		" rdma-core libraries (libibverbs, libmlx5)");
1760 	return -rte_errno;
1761 }
1762 
1763 #endif
1764 
1765 /**
1766  * Driver initialization routine.
1767  */
1768 RTE_INIT(rte_mlx5_pmd_init)
1769 {
1770 	/* Initialize driver log type. */
1771 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1772 	if (mlx5_logtype >= 0)
1773 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1774 
1775 	/* Build the static tables for Verbs conversion. */
1776 	mlx5_set_ptype_table();
1777 	mlx5_set_cksum_table();
1778 	mlx5_set_swp_types_table();
1779 	/*
1780 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1781 	 * huge pages. Calling ibv_fork_init() during init allows
1782 	 * applications to use fork() safely for purposes other than
1783 	 * using this PMD, which is not supported in forked processes.
1784 	 */
1785 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1786 	/* Match the size of Rx completion entry to the size of a cacheline. */
1787 	if (RTE_CACHE_LINE_SIZE == 128)
1788 		setenv("MLX5_CQE_SIZE", "128", 0);
1789 	/*
1790 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
1791 	 * cleanup all the Verbs resources even when the device was removed.
1792 	 */
1793 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
1794 #ifdef RTE_IBVERBS_LINK_DLOPEN
1795 	if (mlx5_glue_init())
1796 		return;
1797 	assert(mlx5_glue);
1798 #endif
1799 #ifndef NDEBUG
1800 	/* Glue structure must not contain any NULL pointers. */
1801 	{
1802 		unsigned int i;
1803 
1804 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1805 			assert(((const void *const *)mlx5_glue)[i]);
1806 	}
1807 #endif
1808 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1809 		DRV_LOG(ERR,
1810 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1811 			mlx5_glue->version, MLX5_GLUE_VERSION);
1812 		return;
1813 	}
1814 	mlx5_glue->fork_init();
1815 	rte_pci_register(&mlx5_driver);
1816 }
1817 
1818 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1819 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1820 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1821