xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 2d0c29a37a9c080c1cccb1ad7941aba2ccf5437e)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <assert.h>
10 #include <dlfcn.h>
11 #include <stdint.h>
12 #include <stdlib.h>
13 #include <errno.h>
14 #include <net/if.h>
15 #include <sys/mman.h>
16 #include <linux/rtnetlink.h>
17 
18 /* Verbs header. */
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27 
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
31 #include <rte_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_eal_memconfig.h>
36 #include <rte_kvargs.h>
37 #include <rte_rwlock.h>
38 #include <rte_spinlock.h>
39 #include <rte_string_fns.h>
40 
41 #include "mlx5.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_autoconf.h"
45 #include "mlx5_defs.h"
46 #include "mlx5_glue.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49 
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52 
53 /* Device parameter to enable RX completion entry padding to 128B. */
54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
55 
56 /* Device parameter to enable padding Rx packet to cacheline size. */
57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
58 
59 /* Device parameter to enable Multi-Packet Rx queue. */
60 #define MLX5_RX_MPRQ_EN "mprq_en"
61 
62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
64 
65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
67 
68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
70 
71 /* Device parameter to configure inline send. */
72 #define MLX5_TXQ_INLINE "txq_inline"
73 
74 /*
75  * Device parameter to configure the number of TX queues threshold for
76  * enabling inline send.
77  */
78 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
79 
80 /*
81  * Device parameter to configure the number of TX queues threshold for
82  * enabling vectorized Tx.
83  */
84 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
85 
86 /* Device parameter to enable multi-packet send WQEs. */
87 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
88 
89 /* Device parameter to include 2 dsegs in the title WQEBB. */
90 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
91 
92 /* Device parameter to limit the size of inlining packet. */
93 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
94 
95 /* Device parameter to enable hardware Tx vector. */
96 #define MLX5_TX_VEC_EN "tx_vec_en"
97 
98 /* Device parameter to enable hardware Rx vector. */
99 #define MLX5_RX_VEC_EN "rx_vec_en"
100 
101 /* Allow L3 VXLAN flow creation. */
102 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
103 
104 /* Activate DV flow steering. */
105 #define MLX5_DV_FLOW_EN "dv_flow_en"
106 
107 /* Activate Netlink support in VF mode. */
108 #define MLX5_VF_NL_EN "vf_nl_en"
109 
110 /* Select port representors to instantiate. */
111 #define MLX5_REPRESENTOR "representor"
112 
113 #ifndef HAVE_IBV_MLX5_MOD_MPW
114 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
115 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
116 #endif
117 
118 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
119 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
120 #endif
121 
122 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
123 
124 /* Shared memory between primary and secondary processes. */
125 struct mlx5_shared_data *mlx5_shared_data;
126 
127 /* Spinlock for mlx5_shared_data allocation. */
128 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
129 
130 /** Driver-specific log messages type. */
131 int mlx5_logtype;
132 
133 /** Data associated with devices to spawn. */
134 struct mlx5_dev_spawn_data {
135 	uint32_t ifindex; /**< Network interface index. */
136 	uint32_t max_port; /**< IB device maximal port index. */
137 	uint32_t ibv_port; /**< IB device physical port index. */
138 	struct mlx5_switch_info info; /**< Switch information. */
139 	struct ibv_device *ibv_dev; /**< Associated IB device. */
140 	struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
141 };
142 
143 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
144 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
145 
146 /**
147  * Allocate shared IB device context. If there is multiport device the
148  * master and representors will share this context, if there is single
149  * port dedicated IB device, the context will be used by only given
150  * port due to unification.
151  *
152  * Routine first searches the context for the spesified IB device name,
153  * if found the shared context assumed and reference counter is incremented.
154  * If no context found the new one is created and initialized with specified
155  * IB device context and parameters.
156  *
157  * @param[in] spawn
158  *   Pointer to the IB device attributes (name, port, etc).
159  *
160  * @return
161  *   Pointer to mlx5_ibv_shared object on success,
162  *   otherwise NULL and rte_errno is set.
163  */
164 static struct mlx5_ibv_shared *
165 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
166 {
167 	struct mlx5_ibv_shared *sh;
168 	int err = 0;
169 	uint32_t i;
170 
171 	assert(spawn);
172 	/* Secondary process should not create the shared context. */
173 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
174 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
175 	/* Search for IB context by device name. */
176 	LIST_FOREACH(sh, &mlx5_ibv_list, next) {
177 		if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
178 			sh->refcnt++;
179 			goto exit;
180 		}
181 	}
182 	/* No device found, we have to create new sharted context. */
183 	assert(spawn->max_port);
184 	sh = rte_zmalloc("ethdev shared ib context",
185 			 sizeof(struct mlx5_ibv_shared) +
186 			 spawn->max_port *
187 			 sizeof(struct mlx5_ibv_shared_port),
188 			 RTE_CACHE_LINE_SIZE);
189 	if (!sh) {
190 		DRV_LOG(ERR, "shared context allocation failure");
191 		rte_errno  = ENOMEM;
192 		goto exit;
193 	}
194 	/* Try to open IB device with DV first, then usual Verbs. */
195 	errno = 0;
196 	sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
197 	if (sh->ctx) {
198 		sh->devx = 1;
199 		DRV_LOG(DEBUG, "DevX is supported");
200 	} else {
201 		sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
202 		if (!sh->ctx) {
203 			err = errno ? errno : ENODEV;
204 			goto error;
205 		}
206 		DRV_LOG(DEBUG, "DevX is NOT supported");
207 	}
208 	err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
209 	if (err) {
210 		DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
211 		goto error;
212 	}
213 	sh->refcnt = 1;
214 	sh->max_port = spawn->max_port;
215 	strncpy(sh->ibdev_name, sh->ctx->device->name,
216 		sizeof(sh->ibdev_name));
217 	strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
218 		sizeof(sh->ibdev_path));
219 	pthread_mutex_init(&sh->intr_mutex, NULL);
220 	/*
221 	 * Setting port_id to max unallowed value means
222 	 * there is no interrupt subhandler installed for
223 	 * the given port index i.
224 	 */
225 	for (i = 0; i < sh->max_port; i++)
226 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
227 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
228 	if (sh->pd == NULL) {
229 		DRV_LOG(ERR, "PD allocation failure");
230 		err = ENOMEM;
231 		goto error;
232 	}
233 	LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
234 exit:
235 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
236 	return sh;
237 error:
238 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
239 	assert(sh);
240 	if (sh->pd)
241 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
242 	if (sh->ctx)
243 		claim_zero(mlx5_glue->close_device(sh->ctx));
244 	rte_free(sh);
245 	assert(err > 0);
246 	rte_errno = err;
247 	return NULL;
248 }
249 
250 /**
251  * Free shared IB device context. Decrement counter and if zero free
252  * all allocated resources and close handles.
253  *
254  * @param[in] sh
255  *   Pointer to mlx5_ibv_shared object to free
256  */
257 static void
258 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
259 {
260 	pthread_mutex_lock(&mlx5_ibv_list_mutex);
261 #ifndef NDEBUG
262 	/* Check the object presence in the list. */
263 	struct mlx5_ibv_shared *lctx;
264 
265 	LIST_FOREACH(lctx, &mlx5_ibv_list, next)
266 		if (lctx == sh)
267 			break;
268 	assert(lctx);
269 	if (lctx != sh) {
270 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
271 		goto exit;
272 	}
273 #endif
274 	assert(sh);
275 	assert(sh->refcnt);
276 	/* Secondary process should not free the shared context. */
277 	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
278 	if (--sh->refcnt)
279 		goto exit;
280 	LIST_REMOVE(sh, next);
281 	/*
282 	 *  Ensure there is no async event handler installed.
283 	 *  Only primary process handles async device events.
284 	 **/
285 	assert(!sh->intr_cnt);
286 	if (sh->intr_cnt)
287 		rte_intr_callback_unregister
288 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
289 	pthread_mutex_destroy(&sh->intr_mutex);
290 	if (sh->pd)
291 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
292 	if (sh->ctx)
293 		claim_zero(mlx5_glue->close_device(sh->ctx));
294 	rte_free(sh);
295 exit:
296 	pthread_mutex_unlock(&mlx5_ibv_list_mutex);
297 }
298 
299 /**
300  * Prepare shared data between primary and secondary process.
301  */
302 static void
303 mlx5_prepare_shared_data(void)
304 {
305 	const struct rte_memzone *mz;
306 
307 	rte_spinlock_lock(&mlx5_shared_data_lock);
308 	if (mlx5_shared_data == NULL) {
309 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
310 			/* Allocate shared memory. */
311 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
312 						 sizeof(*mlx5_shared_data),
313 						 SOCKET_ID_ANY, 0);
314 		} else {
315 			/* Lookup allocated shared memory. */
316 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
317 		}
318 		if (mz == NULL)
319 			rte_panic("Cannot allocate mlx5 shared data\n");
320 		mlx5_shared_data = mz->addr;
321 		/* Initialize shared data. */
322 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
323 			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
324 			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
325 		}
326 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
327 						mlx5_mr_mem_event_cb, NULL);
328 	}
329 	rte_spinlock_unlock(&mlx5_shared_data_lock);
330 }
331 
332 /**
333  * Retrieve integer value from environment variable.
334  *
335  * @param[in] name
336  *   Environment variable name.
337  *
338  * @return
339  *   Integer value, 0 if the variable is not set.
340  */
341 int
342 mlx5_getenv_int(const char *name)
343 {
344 	const char *val = getenv(name);
345 
346 	if (val == NULL)
347 		return 0;
348 	return atoi(val);
349 }
350 
351 /**
352  * Verbs callback to allocate a memory. This function should allocate the space
353  * according to the size provided residing inside a huge page.
354  * Please note that all allocation must respect the alignment from libmlx5
355  * (i.e. currently sysconf(_SC_PAGESIZE)).
356  *
357  * @param[in] size
358  *   The size in bytes of the memory to allocate.
359  * @param[in] data
360  *   A pointer to the callback data.
361  *
362  * @return
363  *   Allocated buffer, NULL otherwise and rte_errno is set.
364  */
365 static void *
366 mlx5_alloc_verbs_buf(size_t size, void *data)
367 {
368 	struct mlx5_priv *priv = data;
369 	void *ret;
370 	size_t alignment = sysconf(_SC_PAGESIZE);
371 	unsigned int socket = SOCKET_ID_ANY;
372 
373 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
374 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
375 
376 		socket = ctrl->socket;
377 	} else if (priv->verbs_alloc_ctx.type ==
378 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
379 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
380 
381 		socket = ctrl->socket;
382 	}
383 	assert(data != NULL);
384 	ret = rte_malloc_socket(__func__, size, alignment, socket);
385 	if (!ret && size)
386 		rte_errno = ENOMEM;
387 	return ret;
388 }
389 
390 /**
391  * Verbs callback to free a memory.
392  *
393  * @param[in] ptr
394  *   A pointer to the memory to free.
395  * @param[in] data
396  *   A pointer to the callback data.
397  */
398 static void
399 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
400 {
401 	assert(data != NULL);
402 	rte_free(ptr);
403 }
404 
405 /**
406  * DPDK callback to close the device.
407  *
408  * Destroy all queues and objects, free memory.
409  *
410  * @param dev
411  *   Pointer to Ethernet device structure.
412  */
413 static void
414 mlx5_dev_close(struct rte_eth_dev *dev)
415 {
416 	struct mlx5_priv *priv = dev->data->dev_private;
417 	unsigned int i;
418 	int ret;
419 
420 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
421 		dev->data->port_id,
422 		((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
423 	/* In case mlx5_dev_stop() has not been called. */
424 	mlx5_dev_interrupt_handler_uninstall(dev);
425 	mlx5_traffic_disable(dev);
426 	mlx5_flow_flush(dev, NULL);
427 	/* Prevent crashes when queues are still in use. */
428 	dev->rx_pkt_burst = removed_rx_burst;
429 	dev->tx_pkt_burst = removed_tx_burst;
430 	if (priv->rxqs != NULL) {
431 		/* XXX race condition if mlx5_rx_burst() is still running. */
432 		usleep(1000);
433 		for (i = 0; (i != priv->rxqs_n); ++i)
434 			mlx5_rxq_release(dev, i);
435 		priv->rxqs_n = 0;
436 		priv->rxqs = NULL;
437 	}
438 	if (priv->txqs != NULL) {
439 		/* XXX race condition if mlx5_tx_burst() is still running. */
440 		usleep(1000);
441 		for (i = 0; (i != priv->txqs_n); ++i)
442 			mlx5_txq_release(dev, i);
443 		priv->txqs_n = 0;
444 		priv->txqs = NULL;
445 	}
446 	mlx5_mprq_free_mp(dev);
447 	mlx5_mr_release(dev);
448 	assert(priv->sh);
449 	if (priv->sh)
450 		mlx5_free_shared_ibctx(priv->sh);
451 	priv->sh = NULL;
452 	if (priv->rss_conf.rss_key != NULL)
453 		rte_free(priv->rss_conf.rss_key);
454 	if (priv->reta_idx != NULL)
455 		rte_free(priv->reta_idx);
456 	if (priv->primary_socket)
457 		mlx5_socket_uninit(dev);
458 	if (priv->config.vf)
459 		mlx5_nl_mac_addr_flush(dev);
460 	if (priv->nl_socket_route >= 0)
461 		close(priv->nl_socket_route);
462 	if (priv->nl_socket_rdma >= 0)
463 		close(priv->nl_socket_rdma);
464 	if (priv->tcf_context)
465 		mlx5_flow_tcf_context_destroy(priv->tcf_context);
466 	ret = mlx5_hrxq_ibv_verify(dev);
467 	if (ret)
468 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
469 			dev->data->port_id);
470 	ret = mlx5_ind_table_ibv_verify(dev);
471 	if (ret)
472 		DRV_LOG(WARNING, "port %u some indirection table still remain",
473 			dev->data->port_id);
474 	ret = mlx5_rxq_ibv_verify(dev);
475 	if (ret)
476 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
477 			dev->data->port_id);
478 	ret = mlx5_rxq_verify(dev);
479 	if (ret)
480 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
481 			dev->data->port_id);
482 	ret = mlx5_txq_ibv_verify(dev);
483 	if (ret)
484 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
485 			dev->data->port_id);
486 	ret = mlx5_txq_verify(dev);
487 	if (ret)
488 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
489 			dev->data->port_id);
490 	ret = mlx5_flow_verify(dev);
491 	if (ret)
492 		DRV_LOG(WARNING, "port %u some flows still remain",
493 			dev->data->port_id);
494 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
495 		unsigned int c = 0;
496 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
497 		uint16_t port_id[i];
498 
499 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
500 		while (i--) {
501 			struct mlx5_priv *opriv =
502 				rte_eth_devices[port_id[i]].data->dev_private;
503 
504 			if (!opriv ||
505 			    opriv->domain_id != priv->domain_id ||
506 			    &rte_eth_devices[port_id[i]] == dev)
507 				continue;
508 			++c;
509 		}
510 		if (!c)
511 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
512 	}
513 	memset(priv, 0, sizeof(*priv));
514 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
515 	/*
516 	 * Reset mac_addrs to NULL such that it is not freed as part of
517 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
518 	 * it is freed when dev_private is freed.
519 	 */
520 	dev->data->mac_addrs = NULL;
521 }
522 
523 const struct eth_dev_ops mlx5_dev_ops = {
524 	.dev_configure = mlx5_dev_configure,
525 	.dev_start = mlx5_dev_start,
526 	.dev_stop = mlx5_dev_stop,
527 	.dev_set_link_down = mlx5_set_link_down,
528 	.dev_set_link_up = mlx5_set_link_up,
529 	.dev_close = mlx5_dev_close,
530 	.promiscuous_enable = mlx5_promiscuous_enable,
531 	.promiscuous_disable = mlx5_promiscuous_disable,
532 	.allmulticast_enable = mlx5_allmulticast_enable,
533 	.allmulticast_disable = mlx5_allmulticast_disable,
534 	.link_update = mlx5_link_update,
535 	.stats_get = mlx5_stats_get,
536 	.stats_reset = mlx5_stats_reset,
537 	.xstats_get = mlx5_xstats_get,
538 	.xstats_reset = mlx5_xstats_reset,
539 	.xstats_get_names = mlx5_xstats_get_names,
540 	.fw_version_get = mlx5_fw_version_get,
541 	.dev_infos_get = mlx5_dev_infos_get,
542 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
543 	.vlan_filter_set = mlx5_vlan_filter_set,
544 	.rx_queue_setup = mlx5_rx_queue_setup,
545 	.tx_queue_setup = mlx5_tx_queue_setup,
546 	.rx_queue_release = mlx5_rx_queue_release,
547 	.tx_queue_release = mlx5_tx_queue_release,
548 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
549 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
550 	.mac_addr_remove = mlx5_mac_addr_remove,
551 	.mac_addr_add = mlx5_mac_addr_add,
552 	.mac_addr_set = mlx5_mac_addr_set,
553 	.set_mc_addr_list = mlx5_set_mc_addr_list,
554 	.mtu_set = mlx5_dev_set_mtu,
555 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
556 	.vlan_offload_set = mlx5_vlan_offload_set,
557 	.reta_update = mlx5_dev_rss_reta_update,
558 	.reta_query = mlx5_dev_rss_reta_query,
559 	.rss_hash_update = mlx5_rss_hash_update,
560 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
561 	.filter_ctrl = mlx5_dev_filter_ctrl,
562 	.rx_descriptor_status = mlx5_rx_descriptor_status,
563 	.tx_descriptor_status = mlx5_tx_descriptor_status,
564 	.rx_queue_count = mlx5_rx_queue_count,
565 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
566 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
567 	.is_removed = mlx5_is_removed,
568 };
569 
570 /* Available operations from secondary process. */
571 static const struct eth_dev_ops mlx5_dev_sec_ops = {
572 	.stats_get = mlx5_stats_get,
573 	.stats_reset = mlx5_stats_reset,
574 	.xstats_get = mlx5_xstats_get,
575 	.xstats_reset = mlx5_xstats_reset,
576 	.xstats_get_names = mlx5_xstats_get_names,
577 	.fw_version_get = mlx5_fw_version_get,
578 	.dev_infos_get = mlx5_dev_infos_get,
579 	.rx_descriptor_status = mlx5_rx_descriptor_status,
580 	.tx_descriptor_status = mlx5_tx_descriptor_status,
581 };
582 
583 /* Available operations in flow isolated mode. */
584 const struct eth_dev_ops mlx5_dev_ops_isolate = {
585 	.dev_configure = mlx5_dev_configure,
586 	.dev_start = mlx5_dev_start,
587 	.dev_stop = mlx5_dev_stop,
588 	.dev_set_link_down = mlx5_set_link_down,
589 	.dev_set_link_up = mlx5_set_link_up,
590 	.dev_close = mlx5_dev_close,
591 	.promiscuous_enable = mlx5_promiscuous_enable,
592 	.promiscuous_disable = mlx5_promiscuous_disable,
593 	.allmulticast_enable = mlx5_allmulticast_enable,
594 	.allmulticast_disable = mlx5_allmulticast_disable,
595 	.link_update = mlx5_link_update,
596 	.stats_get = mlx5_stats_get,
597 	.stats_reset = mlx5_stats_reset,
598 	.xstats_get = mlx5_xstats_get,
599 	.xstats_reset = mlx5_xstats_reset,
600 	.xstats_get_names = mlx5_xstats_get_names,
601 	.fw_version_get = mlx5_fw_version_get,
602 	.dev_infos_get = mlx5_dev_infos_get,
603 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
604 	.vlan_filter_set = mlx5_vlan_filter_set,
605 	.rx_queue_setup = mlx5_rx_queue_setup,
606 	.tx_queue_setup = mlx5_tx_queue_setup,
607 	.rx_queue_release = mlx5_rx_queue_release,
608 	.tx_queue_release = mlx5_tx_queue_release,
609 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
610 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
611 	.mac_addr_remove = mlx5_mac_addr_remove,
612 	.mac_addr_add = mlx5_mac_addr_add,
613 	.mac_addr_set = mlx5_mac_addr_set,
614 	.set_mc_addr_list = mlx5_set_mc_addr_list,
615 	.mtu_set = mlx5_dev_set_mtu,
616 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
617 	.vlan_offload_set = mlx5_vlan_offload_set,
618 	.filter_ctrl = mlx5_dev_filter_ctrl,
619 	.rx_descriptor_status = mlx5_rx_descriptor_status,
620 	.tx_descriptor_status = mlx5_tx_descriptor_status,
621 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
622 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
623 	.is_removed = mlx5_is_removed,
624 };
625 
626 /**
627  * Verify and store value for device argument.
628  *
629  * @param[in] key
630  *   Key argument to verify.
631  * @param[in] val
632  *   Value associated with key.
633  * @param opaque
634  *   User data.
635  *
636  * @return
637  *   0 on success, a negative errno value otherwise and rte_errno is set.
638  */
639 static int
640 mlx5_args_check(const char *key, const char *val, void *opaque)
641 {
642 	struct mlx5_dev_config *config = opaque;
643 	unsigned long tmp;
644 
645 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
646 	if (!strcmp(MLX5_REPRESENTOR, key))
647 		return 0;
648 	errno = 0;
649 	tmp = strtoul(val, NULL, 0);
650 	if (errno) {
651 		rte_errno = errno;
652 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
653 		return -rte_errno;
654 	}
655 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
656 		config->cqe_comp = !!tmp;
657 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
658 		config->cqe_pad = !!tmp;
659 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
660 		config->hw_padding = !!tmp;
661 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
662 		config->mprq.enabled = !!tmp;
663 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
664 		config->mprq.stride_num_n = tmp;
665 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
666 		config->mprq.max_memcpy_len = tmp;
667 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
668 		config->mprq.min_rxqs_num = tmp;
669 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
670 		config->txq_inline = tmp;
671 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
672 		config->txqs_inline = tmp;
673 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
674 		config->txqs_vec = tmp;
675 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
676 		config->mps = !!tmp;
677 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
678 		config->mpw_hdr_dseg = !!tmp;
679 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
680 		config->inline_max_packet_sz = tmp;
681 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
682 		config->tx_vec_en = !!tmp;
683 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
684 		config->rx_vec_en = !!tmp;
685 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
686 		config->l3_vxlan_en = !!tmp;
687 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
688 		config->vf_nl_en = !!tmp;
689 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
690 		config->dv_flow_en = !!tmp;
691 	} else {
692 		DRV_LOG(WARNING, "%s: unknown parameter", key);
693 		rte_errno = EINVAL;
694 		return -rte_errno;
695 	}
696 	return 0;
697 }
698 
699 /**
700  * Parse device parameters.
701  *
702  * @param config
703  *   Pointer to device configuration structure.
704  * @param devargs
705  *   Device arguments structure.
706  *
707  * @return
708  *   0 on success, a negative errno value otherwise and rte_errno is set.
709  */
710 static int
711 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
712 {
713 	const char **params = (const char *[]){
714 		MLX5_RXQ_CQE_COMP_EN,
715 		MLX5_RXQ_CQE_PAD_EN,
716 		MLX5_RXQ_PKT_PAD_EN,
717 		MLX5_RX_MPRQ_EN,
718 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
719 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
720 		MLX5_RXQS_MIN_MPRQ,
721 		MLX5_TXQ_INLINE,
722 		MLX5_TXQS_MIN_INLINE,
723 		MLX5_TXQS_MAX_VEC,
724 		MLX5_TXQ_MPW_EN,
725 		MLX5_TXQ_MPW_HDR_DSEG_EN,
726 		MLX5_TXQ_MAX_INLINE_LEN,
727 		MLX5_TX_VEC_EN,
728 		MLX5_RX_VEC_EN,
729 		MLX5_L3_VXLAN_EN,
730 		MLX5_VF_NL_EN,
731 		MLX5_DV_FLOW_EN,
732 		MLX5_REPRESENTOR,
733 		NULL,
734 	};
735 	struct rte_kvargs *kvlist;
736 	int ret = 0;
737 	int i;
738 
739 	if (devargs == NULL)
740 		return 0;
741 	/* Following UGLY cast is done to pass checkpatch. */
742 	kvlist = rte_kvargs_parse(devargs->args, params);
743 	if (kvlist == NULL)
744 		return 0;
745 	/* Process parameters. */
746 	for (i = 0; (params[i] != NULL); ++i) {
747 		if (rte_kvargs_count(kvlist, params[i])) {
748 			ret = rte_kvargs_process(kvlist, params[i],
749 						 mlx5_args_check, config);
750 			if (ret) {
751 				rte_errno = EINVAL;
752 				rte_kvargs_free(kvlist);
753 				return -rte_errno;
754 			}
755 		}
756 	}
757 	rte_kvargs_free(kvlist);
758 	return 0;
759 }
760 
761 static struct rte_pci_driver mlx5_driver;
762 
763 /*
764  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
765  * local resource used by both primary and secondary to avoid duplicate
766  * reservation.
767  * The space has to be available on both primary and secondary process,
768  * TXQ UAR maps to this area using fixed mmap w/o double check.
769  */
770 static void *uar_base;
771 
772 static int
773 find_lower_va_bound(const struct rte_memseg_list *msl,
774 		const struct rte_memseg *ms, void *arg)
775 {
776 	void **addr = arg;
777 
778 	if (msl->external)
779 		return 0;
780 	if (*addr == NULL)
781 		*addr = ms->addr;
782 	else
783 		*addr = RTE_MIN(*addr, ms->addr);
784 
785 	return 0;
786 }
787 
788 /**
789  * Reserve UAR address space for primary process.
790  *
791  * @param[in] dev
792  *   Pointer to Ethernet device.
793  *
794  * @return
795  *   0 on success, a negative errno value otherwise and rte_errno is set.
796  */
797 static int
798 mlx5_uar_init_primary(struct rte_eth_dev *dev)
799 {
800 	struct mlx5_priv *priv = dev->data->dev_private;
801 	void *addr = (void *)0;
802 
803 	if (uar_base) { /* UAR address space mapped. */
804 		priv->uar_base = uar_base;
805 		return 0;
806 	}
807 	/* find out lower bound of hugepage segments */
808 	rte_memseg_walk(find_lower_va_bound, &addr);
809 
810 	/* keep distance to hugepages to minimize potential conflicts. */
811 	addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
812 	/* anonymous mmap, no real memory consumption. */
813 	addr = mmap(addr, MLX5_UAR_SIZE,
814 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
815 	if (addr == MAP_FAILED) {
816 		DRV_LOG(ERR,
817 			"port %u failed to reserve UAR address space, please"
818 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
819 			dev->data->port_id);
820 		rte_errno = ENOMEM;
821 		return -rte_errno;
822 	}
823 	/* Accept either same addr or a new addr returned from mmap if target
824 	 * range occupied.
825 	 */
826 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
827 		dev->data->port_id, addr);
828 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
829 	uar_base = addr; /* process local, don't reserve again. */
830 	return 0;
831 }
832 
833 /**
834  * Reserve UAR address space for secondary process, align with
835  * primary process.
836  *
837  * @param[in] dev
838  *   Pointer to Ethernet device.
839  *
840  * @return
841  *   0 on success, a negative errno value otherwise and rte_errno is set.
842  */
843 static int
844 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
845 {
846 	struct mlx5_priv *priv = dev->data->dev_private;
847 	void *addr;
848 
849 	assert(priv->uar_base);
850 	if (uar_base) { /* already reserved. */
851 		assert(uar_base == priv->uar_base);
852 		return 0;
853 	}
854 	/* anonymous mmap, no real memory consumption. */
855 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
856 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
857 	if (addr == MAP_FAILED) {
858 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
859 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
860 		rte_errno = ENXIO;
861 		return -rte_errno;
862 	}
863 	if (priv->uar_base != addr) {
864 		DRV_LOG(ERR,
865 			"port %u UAR address %p size %llu occupied, please"
866 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
867 			" --base-virtaddr",
868 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
869 		rte_errno = ENXIO;
870 		return -rte_errno;
871 	}
872 	uar_base = addr; /* process local, don't reserve again */
873 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
874 		dev->data->port_id, addr);
875 	return 0;
876 }
877 
878 /**
879  * Spawn an Ethernet device from Verbs information.
880  *
881  * @param dpdk_dev
882  *   Backing DPDK device.
883  * @param spawn
884  *   Verbs device parameters (name, port, switch_info) to spawn.
885  * @param config
886  *   Device configuration parameters.
887  *
888  * @return
889  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
890  *   is set. The following errors are defined:
891  *
892  *   EBUSY: device is not supposed to be spawned.
893  *   EEXIST: device is already spawned
894  */
895 static struct rte_eth_dev *
896 mlx5_dev_spawn(struct rte_device *dpdk_dev,
897 	       struct mlx5_dev_spawn_data *spawn,
898 	       struct mlx5_dev_config config)
899 {
900 	const struct mlx5_switch_info *switch_info = &spawn->info;
901 	struct mlx5_ibv_shared *sh = NULL;
902 	struct ibv_port_attr port_attr;
903 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
904 	struct rte_eth_dev *eth_dev = NULL;
905 	struct mlx5_priv *priv = NULL;
906 	int err = 0;
907 	unsigned int hw_padding = 0;
908 	unsigned int mps;
909 	unsigned int cqe_comp;
910 	unsigned int cqe_pad = 0;
911 	unsigned int tunnel_en = 0;
912 	unsigned int mpls_en = 0;
913 	unsigned int swp = 0;
914 	unsigned int mprq = 0;
915 	unsigned int mprq_min_stride_size_n = 0;
916 	unsigned int mprq_max_stride_size_n = 0;
917 	unsigned int mprq_min_stride_num_n = 0;
918 	unsigned int mprq_max_stride_num_n = 0;
919 	struct ether_addr mac;
920 	char name[RTE_ETH_NAME_MAX_LEN];
921 	int own_domain_id = 0;
922 	uint16_t port_id;
923 	unsigned int i;
924 
925 	/* Determine if this port representor is supposed to be spawned. */
926 	if (switch_info->representor && dpdk_dev->devargs) {
927 		struct rte_eth_devargs eth_da;
928 
929 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
930 		if (err) {
931 			rte_errno = -err;
932 			DRV_LOG(ERR, "failed to process device arguments: %s",
933 				strerror(rte_errno));
934 			return NULL;
935 		}
936 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
937 			if (eth_da.representor_ports[i] ==
938 			    (uint16_t)switch_info->port_name)
939 				break;
940 		if (i == eth_da.nb_representor_ports) {
941 			rte_errno = EBUSY;
942 			return NULL;
943 		}
944 	}
945 	/* Build device name. */
946 	if (!switch_info->representor)
947 		strlcpy(name, dpdk_dev->name, sizeof(name));
948 	else
949 		snprintf(name, sizeof(name), "%s_representor_%u",
950 			 dpdk_dev->name, switch_info->port_name);
951 	/* check if the device is already spawned */
952 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
953 		rte_errno = EEXIST;
954 		return NULL;
955 	}
956 	/* Prepare shared data between primary and secondary process. */
957 	mlx5_prepare_shared_data();
958 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
959 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
960 		eth_dev = rte_eth_dev_attach_secondary(name);
961 		if (eth_dev == NULL) {
962 			DRV_LOG(ERR, "can not attach rte ethdev");
963 			rte_errno = ENOMEM;
964 			return NULL;
965 		}
966 		eth_dev->device = dpdk_dev;
967 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
968 		err = mlx5_uar_init_secondary(eth_dev);
969 		if (err)
970 			return NULL;
971 		/* Receive command fd from primary process */
972 		err = mlx5_socket_connect(eth_dev);
973 		if (err < 0)
974 			return NULL;
975 		/* Remap UAR for Tx queues. */
976 		err = mlx5_tx_uar_remap(eth_dev, err);
977 		if (err)
978 			return NULL;
979 		/*
980 		 * Ethdev pointer is still required as input since
981 		 * the primary device is not accessible from the
982 		 * secondary process.
983 		 */
984 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
985 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
986 		return eth_dev;
987 	}
988 	sh = mlx5_alloc_shared_ibctx(spawn);
989 	if (!sh)
990 		return NULL;
991 	config.devx = sh->devx;
992 #ifdef HAVE_IBV_MLX5_MOD_SWP
993 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
994 #endif
995 	/*
996 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
997 	 * as all ConnectX-5 devices.
998 	 */
999 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1000 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1001 #endif
1002 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1003 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1004 #endif
1005 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1006 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1007 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1008 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1009 			mps = MLX5_MPW_ENHANCED;
1010 		} else {
1011 			DRV_LOG(DEBUG, "MPW is supported");
1012 			mps = MLX5_MPW;
1013 		}
1014 	} else {
1015 		DRV_LOG(DEBUG, "MPW isn't supported");
1016 		mps = MLX5_MPW_DISABLED;
1017 	}
1018 #ifdef HAVE_IBV_MLX5_MOD_SWP
1019 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1020 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1021 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1022 #endif
1023 	config.swp = !!swp;
1024 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1025 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1026 		struct mlx5dv_striding_rq_caps mprq_caps =
1027 			dv_attr.striding_rq_caps;
1028 
1029 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1030 			mprq_caps.min_single_stride_log_num_of_bytes);
1031 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1032 			mprq_caps.max_single_stride_log_num_of_bytes);
1033 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1034 			mprq_caps.min_single_wqe_log_num_of_strides);
1035 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1036 			mprq_caps.max_single_wqe_log_num_of_strides);
1037 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1038 			mprq_caps.supported_qpts);
1039 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1040 		mprq = 1;
1041 		mprq_min_stride_size_n =
1042 			mprq_caps.min_single_stride_log_num_of_bytes;
1043 		mprq_max_stride_size_n =
1044 			mprq_caps.max_single_stride_log_num_of_bytes;
1045 		mprq_min_stride_num_n =
1046 			mprq_caps.min_single_wqe_log_num_of_strides;
1047 		mprq_max_stride_num_n =
1048 			mprq_caps.max_single_wqe_log_num_of_strides;
1049 		config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1050 						   mprq_min_stride_num_n);
1051 	}
1052 #endif
1053 	if (RTE_CACHE_LINE_SIZE == 128 &&
1054 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1055 		cqe_comp = 0;
1056 	else
1057 		cqe_comp = 1;
1058 	config.cqe_comp = cqe_comp;
1059 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1060 	/* Whether device supports 128B Rx CQE padding. */
1061 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1062 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1063 #endif
1064 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1065 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1066 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1067 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1068 			     (dv_attr.tunnel_offloads_caps &
1069 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1070 	}
1071 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1072 		tunnel_en ? "" : "not ");
1073 #else
1074 	DRV_LOG(WARNING,
1075 		"tunnel offloading disabled due to old OFED/rdma-core version");
1076 #endif
1077 	config.tunnel_en = tunnel_en;
1078 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1079 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1080 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1081 		   (dv_attr.tunnel_offloads_caps &
1082 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1083 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1084 		mpls_en ? "" : "not ");
1085 #else
1086 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1087 		" old OFED/rdma-core version or firmware configuration");
1088 #endif
1089 	config.mpls_en = mpls_en;
1090 	/* Check port status. */
1091 	err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1092 	if (err) {
1093 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1094 		goto error;
1095 	}
1096 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1097 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1098 		err = EINVAL;
1099 		goto error;
1100 	}
1101 	if (port_attr.state != IBV_PORT_ACTIVE)
1102 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1103 			mlx5_glue->port_state_str(port_attr.state),
1104 			port_attr.state);
1105 	/* Allocate private eth device data. */
1106 	priv = rte_zmalloc("ethdev private structure",
1107 			   sizeof(*priv),
1108 			   RTE_CACHE_LINE_SIZE);
1109 	if (priv == NULL) {
1110 		DRV_LOG(ERR, "priv allocation failure");
1111 		err = ENOMEM;
1112 		goto error;
1113 	}
1114 	priv->sh = sh;
1115 	priv->ibv_port = spawn->ibv_port;
1116 	priv->mtu = ETHER_MTU;
1117 #ifndef RTE_ARCH_64
1118 	/* Initialize UAR access locks for 32bit implementations. */
1119 	rte_spinlock_init(&priv->uar_lock_cq);
1120 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1121 		rte_spinlock_init(&priv->uar_lock[i]);
1122 #endif
1123 	/* Some internal functions rely on Netlink sockets, open them now. */
1124 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1125 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1126 	priv->nl_sn = 0;
1127 	priv->representor = !!switch_info->representor;
1128 	priv->master = !!switch_info->master;
1129 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1130 	/*
1131 	 * Currently we support single E-Switch per PF configurations
1132 	 * only and vport_id field contains the vport index for
1133 	 * associated VF, which is deduced from representor port name.
1134 	 * For exapmple, let's have the IB device port 10, it has
1135 	 * attached network device eth0, which has port name attribute
1136 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1137 	 * as 3 (2+1). This assigning schema should be changed if the
1138 	 * multiple E-Switch instances per PF configurations or/and PCI
1139 	 * subfunctions are added.
1140 	 */
1141 	priv->vport_id = switch_info->representor ?
1142 			 switch_info->port_name + 1 : -1;
1143 	/* representor_id field keeps the unmodified port/VF index. */
1144 	priv->representor_id = switch_info->representor ?
1145 			       switch_info->port_name : -1;
1146 	/*
1147 	 * Look for sibling devices in order to reuse their switch domain
1148 	 * if any, otherwise allocate one.
1149 	 */
1150 	i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
1151 	if (i > 0) {
1152 		uint16_t port_id[i];
1153 
1154 		i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
1155 		while (i--) {
1156 			const struct mlx5_priv *opriv =
1157 				rte_eth_devices[port_id[i]].data->dev_private;
1158 
1159 			if (!opriv ||
1160 			    opriv->domain_id ==
1161 			    RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1162 				continue;
1163 			priv->domain_id = opriv->domain_id;
1164 			break;
1165 		}
1166 	}
1167 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1168 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1169 		if (err) {
1170 			err = rte_errno;
1171 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1172 				strerror(rte_errno));
1173 			goto error;
1174 		}
1175 		own_domain_id = 1;
1176 	}
1177 	err = mlx5_args(&config, dpdk_dev->devargs);
1178 	if (err) {
1179 		err = rte_errno;
1180 		DRV_LOG(ERR, "failed to process device arguments: %s",
1181 			strerror(rte_errno));
1182 		goto error;
1183 	}
1184 	config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1185 			    IBV_DEVICE_RAW_IP_CSUM);
1186 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1187 		(config.hw_csum ? "" : "not "));
1188 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1189 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1190 	DRV_LOG(DEBUG, "counters are not supported");
1191 #endif
1192 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1193 	if (config.dv_flow_en) {
1194 		DRV_LOG(WARNING, "DV flow is not supported");
1195 		config.dv_flow_en = 0;
1196 	}
1197 #endif
1198 	config.ind_table_max_size =
1199 		sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1200 	/*
1201 	 * Remove this check once DPDK supports larger/variable
1202 	 * indirection tables.
1203 	 */
1204 	if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1205 		config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1206 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1207 		config.ind_table_max_size);
1208 	config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1209 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1210 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1211 		(config.hw_vlan_strip ? "" : "not "));
1212 	config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1213 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1214 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1215 		(config.hw_fcs_strip ? "" : "not "));
1216 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1217 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1218 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1219 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1220 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1221 #endif
1222 	if (config.hw_padding && !hw_padding) {
1223 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1224 		config.hw_padding = 0;
1225 	} else if (config.hw_padding) {
1226 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1227 	}
1228 	config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1229 		      (sh->device_attr.tso_caps.supported_qpts &
1230 		       (1 << IBV_QPT_RAW_PACKET)));
1231 	if (config.tso)
1232 		config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1233 	/*
1234 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1235 	 * by default.
1236 	 */
1237 	if (config.mps == MLX5_ARG_UNSET)
1238 		config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1239 							  MLX5_MPW_DISABLED;
1240 	else
1241 		config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1242 	DRV_LOG(INFO, "%sMPS is %s",
1243 		config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1244 		config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1245 	if (config.cqe_comp && !cqe_comp) {
1246 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1247 		config.cqe_comp = 0;
1248 	}
1249 	if (config.cqe_pad && !cqe_pad) {
1250 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1251 		config.cqe_pad = 0;
1252 	} else if (config.cqe_pad) {
1253 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1254 	}
1255 	if (config.mprq.enabled && mprq) {
1256 		if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1257 		    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1258 			config.mprq.stride_num_n =
1259 				RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1260 					mprq_min_stride_num_n);
1261 			DRV_LOG(WARNING,
1262 				"the number of strides"
1263 				" for Multi-Packet RQ is out of range,"
1264 				" setting default value (%u)",
1265 				1 << config.mprq.stride_num_n);
1266 		}
1267 		config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1268 		config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1269 	} else if (config.mprq.enabled && !mprq) {
1270 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1271 		config.mprq.enabled = 0;
1272 	}
1273 	eth_dev = rte_eth_dev_allocate(name);
1274 	if (eth_dev == NULL) {
1275 		DRV_LOG(ERR, "can not allocate rte ethdev");
1276 		err = ENOMEM;
1277 		goto error;
1278 	}
1279 	/* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1280 	eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1281 	if (priv->representor) {
1282 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1283 		eth_dev->data->representor_id = priv->representor_id;
1284 	}
1285 	eth_dev->data->dev_private = priv;
1286 	priv->dev_data = eth_dev->data;
1287 	eth_dev->data->mac_addrs = priv->mac;
1288 	eth_dev->device = dpdk_dev;
1289 	err = mlx5_uar_init_primary(eth_dev);
1290 	if (err) {
1291 		err = rte_errno;
1292 		goto error;
1293 	}
1294 	/* Configure the first MAC address by default. */
1295 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1296 		DRV_LOG(ERR,
1297 			"port %u cannot get MAC address, is mlx5_en"
1298 			" loaded? (errno: %s)",
1299 			eth_dev->data->port_id, strerror(rte_errno));
1300 		err = ENODEV;
1301 		goto error;
1302 	}
1303 	DRV_LOG(INFO,
1304 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1305 		eth_dev->data->port_id,
1306 		mac.addr_bytes[0], mac.addr_bytes[1],
1307 		mac.addr_bytes[2], mac.addr_bytes[3],
1308 		mac.addr_bytes[4], mac.addr_bytes[5]);
1309 #ifndef NDEBUG
1310 	{
1311 		char ifname[IF_NAMESIZE];
1312 
1313 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1314 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1315 				eth_dev->data->port_id, ifname);
1316 		else
1317 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1318 				eth_dev->data->port_id);
1319 	}
1320 #endif
1321 	/* Get actual MTU if possible. */
1322 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1323 	if (err) {
1324 		err = rte_errno;
1325 		goto error;
1326 	}
1327 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1328 		priv->mtu);
1329 	/* Initialize burst functions to prevent crashes before link-up. */
1330 	eth_dev->rx_pkt_burst = removed_rx_burst;
1331 	eth_dev->tx_pkt_burst = removed_tx_burst;
1332 	eth_dev->dev_ops = &mlx5_dev_ops;
1333 	/* Register MAC address. */
1334 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1335 	if (config.vf && config.vf_nl_en)
1336 		mlx5_nl_mac_addr_sync(eth_dev);
1337 	priv->tcf_context = mlx5_flow_tcf_context_create();
1338 	if (!priv->tcf_context) {
1339 		err = -rte_errno;
1340 		DRV_LOG(WARNING,
1341 			"flow rules relying on switch offloads will not be"
1342 			" supported: cannot open libmnl socket: %s",
1343 			strerror(rte_errno));
1344 	} else {
1345 		struct rte_flow_error error;
1346 		unsigned int ifindex = mlx5_ifindex(eth_dev);
1347 
1348 		if (!ifindex) {
1349 			err = -rte_errno;
1350 			error.message =
1351 				"cannot retrieve network interface index";
1352 		} else {
1353 			err = mlx5_flow_tcf_init(priv->tcf_context,
1354 						 ifindex, &error);
1355 		}
1356 		if (err) {
1357 			DRV_LOG(WARNING,
1358 				"flow rules relying on switch offloads will"
1359 				" not be supported: %s: %s",
1360 				error.message, strerror(rte_errno));
1361 			mlx5_flow_tcf_context_destroy(priv->tcf_context);
1362 			priv->tcf_context = NULL;
1363 		}
1364 	}
1365 	TAILQ_INIT(&priv->flows);
1366 	TAILQ_INIT(&priv->ctrl_flows);
1367 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1368 	struct mlx5dv_ctx_allocators alctr = {
1369 		.alloc = &mlx5_alloc_verbs_buf,
1370 		.free = &mlx5_free_verbs_buf,
1371 		.data = priv,
1372 	};
1373 	mlx5_glue->dv_set_context_attr(sh->ctx,
1374 				       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1375 				       (void *)((uintptr_t)&alctr));
1376 	/* Bring Ethernet device up. */
1377 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1378 		eth_dev->data->port_id);
1379 	mlx5_set_link_up(eth_dev);
1380 	/*
1381 	 * Even though the interrupt handler is not installed yet,
1382 	 * interrupts will still trigger on the asyn_fd from
1383 	 * Verbs context returned by ibv_open_device().
1384 	 */
1385 	mlx5_link_update(eth_dev, 0);
1386 	/* Store device configuration on private structure. */
1387 	priv->config = config;
1388 	/* Supported Verbs flow priority number detection. */
1389 	err = mlx5_flow_discover_priorities(eth_dev);
1390 	if (err < 0) {
1391 		err = -err;
1392 		goto error;
1393 	}
1394 	priv->config.flow_prio = err;
1395 	/*
1396 	 * Once the device is added to the list of memory event
1397 	 * callback, its global MR cache table cannot be expanded
1398 	 * on the fly because of deadlock. If it overflows, lookup
1399 	 * should be done by searching MR list linearly, which is slow.
1400 	 */
1401 	err = mlx5_mr_btree_init(&priv->mr.cache,
1402 				 MLX5_MR_BTREE_CACHE_N * 2,
1403 				 eth_dev->device->numa_node);
1404 	if (err) {
1405 		err = rte_errno;
1406 		goto error;
1407 	}
1408 	/* Add device to memory callback list. */
1409 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1410 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1411 			 priv, mem_event_cb);
1412 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1413 	return eth_dev;
1414 error:
1415 	if (priv) {
1416 		if (priv->nl_socket_route >= 0)
1417 			close(priv->nl_socket_route);
1418 		if (priv->nl_socket_rdma >= 0)
1419 			close(priv->nl_socket_rdma);
1420 		if (priv->tcf_context)
1421 			mlx5_flow_tcf_context_destroy(priv->tcf_context);
1422 		if (own_domain_id)
1423 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1424 		rte_free(priv);
1425 		if (eth_dev != NULL)
1426 			eth_dev->data->dev_private = NULL;
1427 	}
1428 	if (eth_dev != NULL) {
1429 		/* mac_addrs must not be freed alone because part of dev_private */
1430 		eth_dev->data->mac_addrs = NULL;
1431 		rte_eth_dev_release_port(eth_dev);
1432 	}
1433 	if (sh)
1434 		mlx5_free_shared_ibctx(sh);
1435 	assert(err > 0);
1436 	rte_errno = err;
1437 	return NULL;
1438 }
1439 
1440 /**
1441  * Comparison callback to sort device data.
1442  *
1443  * This is meant to be used with qsort().
1444  *
1445  * @param a[in]
1446  *   Pointer to pointer to first data object.
1447  * @param b[in]
1448  *   Pointer to pointer to second data object.
1449  *
1450  * @return
1451  *   0 if both objects are equal, less than 0 if the first argument is less
1452  *   than the second, greater than 0 otherwise.
1453  */
1454 static int
1455 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1456 {
1457 	const struct mlx5_switch_info *si_a =
1458 		&((const struct mlx5_dev_spawn_data *)a)->info;
1459 	const struct mlx5_switch_info *si_b =
1460 		&((const struct mlx5_dev_spawn_data *)b)->info;
1461 	int ret;
1462 
1463 	/* Master device first. */
1464 	ret = si_b->master - si_a->master;
1465 	if (ret)
1466 		return ret;
1467 	/* Then representor devices. */
1468 	ret = si_b->representor - si_a->representor;
1469 	if (ret)
1470 		return ret;
1471 	/* Unidentified devices come last in no specific order. */
1472 	if (!si_a->representor)
1473 		return 0;
1474 	/* Order representors by name. */
1475 	return si_a->port_name - si_b->port_name;
1476 }
1477 
1478 /**
1479  * DPDK callback to register a PCI device.
1480  *
1481  * This function spawns Ethernet devices out of a given PCI device.
1482  *
1483  * @param[in] pci_drv
1484  *   PCI driver structure (mlx5_driver).
1485  * @param[in] pci_dev
1486  *   PCI device information.
1487  *
1488  * @return
1489  *   0 on success, a negative errno value otherwise and rte_errno is set.
1490  */
1491 static int
1492 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1493 	       struct rte_pci_device *pci_dev)
1494 {
1495 	struct ibv_device **ibv_list;
1496 	/*
1497 	 * Number of found IB Devices matching with requested PCI BDF.
1498 	 * nd != 1 means there are multiple IB devices over the same
1499 	 * PCI device and we have representors and master.
1500 	 */
1501 	unsigned int nd = 0;
1502 	/*
1503 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1504 	 * we have the single multiport IB device, and there may be
1505 	 * representors attached to some of found ports.
1506 	 */
1507 	unsigned int np = 0;
1508 	/*
1509 	 * Number of DPDK ethernet devices to Spawn - either over
1510 	 * multiple IB devices or multiple ports of single IB device.
1511 	 * Actually this is the number of iterations to spawn.
1512 	 */
1513 	unsigned int ns = 0;
1514 	struct mlx5_dev_config dev_config;
1515 	int ret;
1516 
1517 	assert(pci_drv == &mlx5_driver);
1518 	errno = 0;
1519 	ibv_list = mlx5_glue->get_device_list(&ret);
1520 	if (!ibv_list) {
1521 		rte_errno = errno ? errno : ENOSYS;
1522 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1523 		return -rte_errno;
1524 	}
1525 	/*
1526 	 * First scan the list of all Infiniband devices to find
1527 	 * matching ones, gathering into the list.
1528 	 */
1529 	struct ibv_device *ibv_match[ret + 1];
1530 	int nl_route = -1;
1531 	int nl_rdma = -1;
1532 	unsigned int i;
1533 
1534 	while (ret-- > 0) {
1535 		struct rte_pci_addr pci_addr;
1536 
1537 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1538 		if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1539 			continue;
1540 		if (pci_dev->addr.domain != pci_addr.domain ||
1541 		    pci_dev->addr.bus != pci_addr.bus ||
1542 		    pci_dev->addr.devid != pci_addr.devid ||
1543 		    pci_dev->addr.function != pci_addr.function)
1544 			continue;
1545 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1546 			ibv_list[ret]->name);
1547 		ibv_match[nd++] = ibv_list[ret];
1548 	}
1549 	ibv_match[nd] = NULL;
1550 	if (!nd) {
1551 		/* No device macthes, just complain and bail out. */
1552 		mlx5_glue->free_device_list(ibv_list);
1553 		DRV_LOG(WARNING,
1554 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1555 			" are kernel drivers loaded?",
1556 			pci_dev->addr.domain, pci_dev->addr.bus,
1557 			pci_dev->addr.devid, pci_dev->addr.function);
1558 		rte_errno = ENOENT;
1559 		ret = -rte_errno;
1560 		return ret;
1561 	}
1562 	nl_route = mlx5_nl_init(NETLINK_ROUTE);
1563 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1564 	if (nd == 1) {
1565 		/*
1566 		 * Found single matching device may have multiple ports.
1567 		 * Each port may be representor, we have to check the port
1568 		 * number and check the representors existence.
1569 		 */
1570 		if (nl_rdma >= 0)
1571 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1572 		if (!np)
1573 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1574 					 " ports number", ibv_match[0]->name);
1575 	}
1576 	/*
1577 	 * Now we can determine the maximal
1578 	 * amount of devices to be spawned.
1579 	 */
1580 	struct mlx5_dev_spawn_data list[np ? np : nd];
1581 
1582 	if (np > 1) {
1583 		/*
1584 		 * Signle IB device with multiple ports found,
1585 		 * it may be E-Switch master device and representors.
1586 		 * We have to perform identification trough the ports.
1587 		 */
1588 		assert(nl_rdma >= 0);
1589 		assert(ns == 0);
1590 		assert(nd == 1);
1591 		for (i = 1; i <= np; ++i) {
1592 			list[ns].max_port = np;
1593 			list[ns].ibv_port = i;
1594 			list[ns].ibv_dev = ibv_match[0];
1595 			list[ns].eth_dev = NULL;
1596 			list[ns].ifindex = mlx5_nl_ifindex
1597 					(nl_rdma, list[ns].ibv_dev->name, i);
1598 			if (!list[ns].ifindex) {
1599 				/*
1600 				 * No network interface index found for the
1601 				 * specified port, it means there is no
1602 				 * representor on this port. It's OK,
1603 				 * there can be disabled ports, for example
1604 				 * if sriov_numvfs < sriov_totalvfs.
1605 				 */
1606 				continue;
1607 			}
1608 			ret = -1;
1609 			if (nl_route >= 0)
1610 				ret = mlx5_nl_switch_info
1611 					       (nl_route,
1612 						list[ns].ifindex,
1613 						&list[ns].info);
1614 			if (ret || (!list[ns].info.representor &&
1615 				    !list[ns].info.master)) {
1616 				/*
1617 				 * We failed to recognize representors with
1618 				 * Netlink, let's try to perform the task
1619 				 * with sysfs.
1620 				 */
1621 				ret =  mlx5_sysfs_switch_info
1622 						(list[ns].ifindex,
1623 						 &list[ns].info);
1624 			}
1625 			if (!ret && (list[ns].info.representor ^
1626 				     list[ns].info.master))
1627 				ns++;
1628 		}
1629 		if (!ns) {
1630 			DRV_LOG(ERR,
1631 				"unable to recognize master/representors"
1632 				" on the IB device with multiple ports");
1633 			rte_errno = ENOENT;
1634 			ret = -rte_errno;
1635 			goto exit;
1636 		}
1637 	} else {
1638 		/*
1639 		 * The existence of several matching entries (nd > 1) means
1640 		 * port representors have been instantiated. No existing Verbs
1641 		 * call nor sysfs entries can tell them apart, this can only
1642 		 * be done through Netlink calls assuming kernel drivers are
1643 		 * recent enough to support them.
1644 		 *
1645 		 * In the event of identification failure through Netlink,
1646 		 * try again through sysfs, then:
1647 		 *
1648 		 * 1. A single IB device matches (nd == 1) with single
1649 		 *    port (np=0/1) and is not a representor, assume
1650 		 *    no switch support.
1651 		 *
1652 		 * 2. Otherwise no safe assumptions can be made;
1653 		 *    complain louder and bail out.
1654 		 */
1655 		np = 1;
1656 		for (i = 0; i != nd; ++i) {
1657 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1658 			list[ns].max_port = 1;
1659 			list[ns].ibv_port = 1;
1660 			list[ns].ibv_dev = ibv_match[i];
1661 			list[ns].eth_dev = NULL;
1662 			list[ns].ifindex = 0;
1663 			if (nl_rdma >= 0)
1664 				list[ns].ifindex = mlx5_nl_ifindex
1665 					(nl_rdma, list[ns].ibv_dev->name, 1);
1666 			if (!list[ns].ifindex) {
1667 				/*
1668 				 * No network interface index found for the
1669 				 * specified device, it means there it is not
1670 				 * a representor/master.
1671 				 */
1672 				continue;
1673 			}
1674 			ret = -1;
1675 			if (nl_route >= 0)
1676 				ret = mlx5_nl_switch_info
1677 					       (nl_route,
1678 						list[ns].ifindex,
1679 						&list[ns].info);
1680 			if (ret || (!list[ns].info.representor &&
1681 				    !list[ns].info.master)) {
1682 				/*
1683 				 * We failed to recognize representors with
1684 				 * Netlink, let's try to perform the task
1685 				 * with sysfs.
1686 				 */
1687 				ret =  mlx5_sysfs_switch_info
1688 						(list[ns].ifindex,
1689 						 &list[ns].info);
1690 			}
1691 			if (!ret && (list[ns].info.representor ^
1692 				     list[ns].info.master)) {
1693 				ns++;
1694 			} else if ((nd == 1) &&
1695 				   !list[ns].info.representor &&
1696 				   !list[ns].info.master) {
1697 				/*
1698 				 * Single IB device with
1699 				 * one physical port and
1700 				 * attached network device.
1701 				 * May be SRIOV is not enabled
1702 				 * or there is no representors.
1703 				 */
1704 				DRV_LOG(INFO, "no E-Switch support detected");
1705 				ns++;
1706 				break;
1707 			}
1708 		}
1709 		if (!ns) {
1710 			DRV_LOG(ERR,
1711 				"unable to recognize master/representors"
1712 				" on the multiple IB devices");
1713 			rte_errno = ENOENT;
1714 			ret = -rte_errno;
1715 			goto exit;
1716 		}
1717 	}
1718 	assert(ns);
1719 	/*
1720 	 * Sort list to probe devices in natural order for users convenience
1721 	 * (i.e. master first, then representors from lowest to highest ID).
1722 	 */
1723 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1724 	/* Default configuration. */
1725 	dev_config = (struct mlx5_dev_config){
1726 		.hw_padding = 0,
1727 		.mps = MLX5_ARG_UNSET,
1728 		.tx_vec_en = 1,
1729 		.rx_vec_en = 1,
1730 		.txq_inline = MLX5_ARG_UNSET,
1731 		.txqs_inline = MLX5_ARG_UNSET,
1732 		.txqs_vec = MLX5_ARG_UNSET,
1733 		.inline_max_packet_sz = MLX5_ARG_UNSET,
1734 		.vf_nl_en = 1,
1735 		.mprq = {
1736 			.enabled = 0, /* Disabled by default. */
1737 			.stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1738 			.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1739 			.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1740 		},
1741 	};
1742 	/* Device specific configuration. */
1743 	switch (pci_dev->id.device_id) {
1744 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1745 		dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1746 		break;
1747 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1748 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1749 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1750 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1751 		dev_config.vf = 1;
1752 		break;
1753 	default:
1754 		break;
1755 	}
1756 	/* Set architecture-dependent default value if unset. */
1757 	if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1758 		dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1759 	for (i = 0; i != ns; ++i) {
1760 		uint32_t restore;
1761 
1762 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1763 						 &list[i],
1764 						 dev_config);
1765 		if (!list[i].eth_dev) {
1766 			if (rte_errno != EBUSY && rte_errno != EEXIST)
1767 				break;
1768 			/* Device is disabled or already spawned. Ignore it. */
1769 			continue;
1770 		}
1771 		restore = list[i].eth_dev->data->dev_flags;
1772 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1773 		/* Restore non-PCI flags cleared by the above call. */
1774 		list[i].eth_dev->data->dev_flags |= restore;
1775 		rte_eth_dev_probing_finish(list[i].eth_dev);
1776 	}
1777 	if (i != ns) {
1778 		DRV_LOG(ERR,
1779 			"probe of PCI device " PCI_PRI_FMT " aborted after"
1780 			" encountering an error: %s",
1781 			pci_dev->addr.domain, pci_dev->addr.bus,
1782 			pci_dev->addr.devid, pci_dev->addr.function,
1783 			strerror(rte_errno));
1784 		ret = -rte_errno;
1785 		/* Roll back. */
1786 		while (i--) {
1787 			if (!list[i].eth_dev)
1788 				continue;
1789 			mlx5_dev_close(list[i].eth_dev);
1790 			/* mac_addrs must not be freed because in dev_private */
1791 			list[i].eth_dev->data->mac_addrs = NULL;
1792 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1793 		}
1794 		/* Restore original error. */
1795 		rte_errno = -ret;
1796 	} else {
1797 		ret = 0;
1798 	}
1799 exit:
1800 	/*
1801 	 * Do the routine cleanup:
1802 	 * - close opened Netlink sockets
1803 	 * - free the Infiniband device list
1804 	 */
1805 	if (nl_rdma >= 0)
1806 		close(nl_rdma);
1807 	if (nl_route >= 0)
1808 		close(nl_route);
1809 	assert(ibv_list);
1810 	mlx5_glue->free_device_list(ibv_list);
1811 	return ret;
1812 }
1813 
1814 /**
1815  * DPDK callback to remove a PCI device.
1816  *
1817  * This function removes all Ethernet devices belong to a given PCI device.
1818  *
1819  * @param[in] pci_dev
1820  *   Pointer to the PCI device.
1821  *
1822  * @return
1823  *   0 on success, the function cannot fail.
1824  */
1825 static int
1826 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1827 {
1828 	uint16_t port_id;
1829 	struct rte_eth_dev *port;
1830 
1831 	for (port_id = 0; port_id < RTE_MAX_ETHPORTS; port_id++) {
1832 		port = &rte_eth_devices[port_id];
1833 		if (port->state != RTE_ETH_DEV_UNUSED &&
1834 				port->device == &pci_dev->device)
1835 			rte_eth_dev_close(port_id);
1836 	}
1837 	return 0;
1838 }
1839 
1840 static const struct rte_pci_id mlx5_pci_id_map[] = {
1841 	{
1842 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1843 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1844 	},
1845 	{
1846 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1847 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1848 	},
1849 	{
1850 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1851 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1852 	},
1853 	{
1854 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1855 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1856 	},
1857 	{
1858 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1859 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1860 	},
1861 	{
1862 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1863 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1864 	},
1865 	{
1866 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1867 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1868 	},
1869 	{
1870 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1871 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1872 	},
1873 	{
1874 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1875 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1876 	},
1877 	{
1878 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1879 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
1880 	},
1881 	{
1882 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1883 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
1884 	},
1885 	{
1886 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1887 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
1888 	},
1889 	{
1890 		.vendor_id = 0
1891 	}
1892 };
1893 
1894 static struct rte_pci_driver mlx5_driver = {
1895 	.driver = {
1896 		.name = MLX5_DRIVER_NAME
1897 	},
1898 	.id_table = mlx5_pci_id_map,
1899 	.probe = mlx5_pci_probe,
1900 	.remove = mlx5_pci_remove,
1901 	.dma_map = mlx5_dma_map,
1902 	.dma_unmap = mlx5_dma_unmap,
1903 	.drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
1904 		      RTE_PCI_DRV_PROBE_AGAIN),
1905 };
1906 
1907 #ifdef RTE_IBVERBS_LINK_DLOPEN
1908 
1909 /**
1910  * Suffix RTE_EAL_PMD_PATH with "-glue".
1911  *
1912  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1913  * suffixing its last component.
1914  *
1915  * @param buf[out]
1916  *   Output buffer, should be large enough otherwise NULL is returned.
1917  * @param size
1918  *   Size of @p out.
1919  *
1920  * @return
1921  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1922  */
1923 static char *
1924 mlx5_glue_path(char *buf, size_t size)
1925 {
1926 	static const char *const bad[] = { "/", ".", "..", NULL };
1927 	const char *path = RTE_EAL_PMD_PATH;
1928 	size_t len = strlen(path);
1929 	size_t off;
1930 	int i;
1931 
1932 	while (len && path[len - 1] == '/')
1933 		--len;
1934 	for (off = len; off && path[off - 1] != '/'; --off)
1935 		;
1936 	for (i = 0; bad[i]; ++i)
1937 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1938 			goto error;
1939 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1940 	if (i == -1 || (size_t)i >= size)
1941 		goto error;
1942 	return buf;
1943 error:
1944 	DRV_LOG(ERR,
1945 		"unable to append \"-glue\" to last component of"
1946 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1947 		" please re-configure DPDK");
1948 	return NULL;
1949 }
1950 
1951 /**
1952  * Initialization routine for run-time dependency on rdma-core.
1953  */
1954 static int
1955 mlx5_glue_init(void)
1956 {
1957 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1958 	const char *path[] = {
1959 		/*
1960 		 * A basic security check is necessary before trusting
1961 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1962 		 */
1963 		(geteuid() == getuid() && getegid() == getgid() ?
1964 		 getenv("MLX5_GLUE_PATH") : NULL),
1965 		/*
1966 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1967 		 * variant, otherwise let dlopen() look up libraries on its
1968 		 * own.
1969 		 */
1970 		(*RTE_EAL_PMD_PATH ?
1971 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1972 	};
1973 	unsigned int i = 0;
1974 	void *handle = NULL;
1975 	void **sym;
1976 	const char *dlmsg;
1977 
1978 	while (!handle && i != RTE_DIM(path)) {
1979 		const char *end;
1980 		size_t len;
1981 		int ret;
1982 
1983 		if (!path[i]) {
1984 			++i;
1985 			continue;
1986 		}
1987 		end = strpbrk(path[i], ":;");
1988 		if (!end)
1989 			end = path[i] + strlen(path[i]);
1990 		len = end - path[i];
1991 		ret = 0;
1992 		do {
1993 			char name[ret + 1];
1994 
1995 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1996 				       (int)len, path[i],
1997 				       (!len || *(end - 1) == '/') ? "" : "/");
1998 			if (ret == -1)
1999 				break;
2000 			if (sizeof(name) != (size_t)ret + 1)
2001 				continue;
2002 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2003 				name);
2004 			handle = dlopen(name, RTLD_LAZY);
2005 			break;
2006 		} while (1);
2007 		path[i] = end + 1;
2008 		if (!*end)
2009 			++i;
2010 	}
2011 	if (!handle) {
2012 		rte_errno = EINVAL;
2013 		dlmsg = dlerror();
2014 		if (dlmsg)
2015 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2016 		goto glue_error;
2017 	}
2018 	sym = dlsym(handle, "mlx5_glue");
2019 	if (!sym || !*sym) {
2020 		rte_errno = EINVAL;
2021 		dlmsg = dlerror();
2022 		if (dlmsg)
2023 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2024 		goto glue_error;
2025 	}
2026 	mlx5_glue = *sym;
2027 	return 0;
2028 glue_error:
2029 	if (handle)
2030 		dlclose(handle);
2031 	DRV_LOG(WARNING,
2032 		"cannot initialize PMD due to missing run-time dependency on"
2033 		" rdma-core libraries (libibverbs, libmlx5)");
2034 	return -rte_errno;
2035 }
2036 
2037 #endif
2038 
2039 /**
2040  * Driver initialization routine.
2041  */
2042 RTE_INIT(rte_mlx5_pmd_init)
2043 {
2044 	/* Initialize driver log type. */
2045 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
2046 	if (mlx5_logtype >= 0)
2047 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2048 
2049 	/* Build the static tables for Verbs conversion. */
2050 	mlx5_set_ptype_table();
2051 	mlx5_set_cksum_table();
2052 	mlx5_set_swp_types_table();
2053 	/*
2054 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2055 	 * huge pages. Calling ibv_fork_init() during init allows
2056 	 * applications to use fork() safely for purposes other than
2057 	 * using this PMD, which is not supported in forked processes.
2058 	 */
2059 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2060 	/* Match the size of Rx completion entry to the size of a cacheline. */
2061 	if (RTE_CACHE_LINE_SIZE == 128)
2062 		setenv("MLX5_CQE_SIZE", "128", 0);
2063 	/*
2064 	 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2065 	 * cleanup all the Verbs resources even when the device was removed.
2066 	 */
2067 	setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2068 #ifdef RTE_IBVERBS_LINK_DLOPEN
2069 	if (mlx5_glue_init())
2070 		return;
2071 	assert(mlx5_glue);
2072 #endif
2073 #ifndef NDEBUG
2074 	/* Glue structure must not contain any NULL pointers. */
2075 	{
2076 		unsigned int i;
2077 
2078 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2079 			assert(((const void *const *)mlx5_glue)[i]);
2080 	}
2081 #endif
2082 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2083 		DRV_LOG(ERR,
2084 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
2085 			mlx5_glue->version, MLX5_GLUE_VERSION);
2086 		return;
2087 	}
2088 	mlx5_glue->fork_init();
2089 	rte_pci_register(&mlx5_driver);
2090 }
2091 
2092 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2093 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2094 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
2095