xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 08aa6271c86a561b66c6dd91f9a54fa2f12bc859)
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 
38 #include "mlx5.h"
39 #include "mlx5_utils.h"
40 #include "mlx5_rxtx.h"
41 #include "mlx5_autoconf.h"
42 #include "mlx5_defs.h"
43 #include "mlx5_glue.h"
44 #include "mlx5_mr.h"
45 
46 /* Device parameter to enable RX completion queue compression. */
47 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
48 
49 /* Device parameter to enable Multi-Packet Rx queue. */
50 #define MLX5_RX_MPRQ_EN "mprq_en"
51 
52 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
53 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
54 
55 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
56 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
57 
58 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
59 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
60 
61 /* Device parameter to configure inline send. */
62 #define MLX5_TXQ_INLINE "txq_inline"
63 
64 /*
65  * Device parameter to configure the number of TX queues threshold for
66  * enabling inline send.
67  */
68 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
69 
70 /* Device parameter to enable multi-packet send WQEs. */
71 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
72 
73 /* Device parameter to include 2 dsegs in the title WQEBB. */
74 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
75 
76 /* Device parameter to limit the size of inlining packet. */
77 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
78 
79 /* Device parameter to enable hardware Tx vector. */
80 #define MLX5_TX_VEC_EN "tx_vec_en"
81 
82 /* Device parameter to enable hardware Rx vector. */
83 #define MLX5_RX_VEC_EN "rx_vec_en"
84 
85 /* Allow L3 VXLAN flow creation. */
86 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
87 
88 /* Activate Netlink support in VF mode. */
89 #define MLX5_VF_NL_EN "vf_nl_en"
90 
91 #ifndef HAVE_IBV_MLX5_MOD_MPW
92 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
93 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
94 #endif
95 
96 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
97 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
98 #endif
99 
100 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
101 
102 /* Shared memory between primary and secondary processes. */
103 struct mlx5_shared_data *mlx5_shared_data;
104 
105 /* Spinlock for mlx5_shared_data allocation. */
106 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
107 
108 /** Driver-specific log messages type. */
109 int mlx5_logtype;
110 
111 /**
112  * Prepare shared data between primary and secondary process.
113  */
114 static void
115 mlx5_prepare_shared_data(void)
116 {
117 	const struct rte_memzone *mz;
118 
119 	rte_spinlock_lock(&mlx5_shared_data_lock);
120 	if (mlx5_shared_data == NULL) {
121 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
122 			/* Allocate shared memory. */
123 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
124 						 sizeof(*mlx5_shared_data),
125 						 SOCKET_ID_ANY, 0);
126 		} else {
127 			/* Lookup allocated shared memory. */
128 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
129 		}
130 		if (mz == NULL)
131 			rte_panic("Cannot allocate mlx5 shared data\n");
132 		mlx5_shared_data = mz->addr;
133 		/* Initialize shared data. */
134 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
135 			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
136 			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
137 		}
138 	}
139 	rte_spinlock_unlock(&mlx5_shared_data_lock);
140 }
141 
142 /**
143  * Retrieve integer value from environment variable.
144  *
145  * @param[in] name
146  *   Environment variable name.
147  *
148  * @return
149  *   Integer value, 0 if the variable is not set.
150  */
151 int
152 mlx5_getenv_int(const char *name)
153 {
154 	const char *val = getenv(name);
155 
156 	if (val == NULL)
157 		return 0;
158 	return atoi(val);
159 }
160 
161 /**
162  * Verbs callback to allocate a memory. This function should allocate the space
163  * according to the size provided residing inside a huge page.
164  * Please note that all allocation must respect the alignment from libmlx5
165  * (i.e. currently sysconf(_SC_PAGESIZE)).
166  *
167  * @param[in] size
168  *   The size in bytes of the memory to allocate.
169  * @param[in] data
170  *   A pointer to the callback data.
171  *
172  * @return
173  *   Allocated buffer, NULL otherwise and rte_errno is set.
174  */
175 static void *
176 mlx5_alloc_verbs_buf(size_t size, void *data)
177 {
178 	struct priv *priv = data;
179 	void *ret;
180 	size_t alignment = sysconf(_SC_PAGESIZE);
181 	unsigned int socket = SOCKET_ID_ANY;
182 
183 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
184 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
185 
186 		socket = ctrl->socket;
187 	} else if (priv->verbs_alloc_ctx.type ==
188 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
189 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
190 
191 		socket = ctrl->socket;
192 	}
193 	assert(data != NULL);
194 	ret = rte_malloc_socket(__func__, size, alignment, socket);
195 	if (!ret && size)
196 		rte_errno = ENOMEM;
197 	return ret;
198 }
199 
200 /**
201  * Verbs callback to free a memory.
202  *
203  * @param[in] ptr
204  *   A pointer to the memory to free.
205  * @param[in] data
206  *   A pointer to the callback data.
207  */
208 static void
209 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
210 {
211 	assert(data != NULL);
212 	rte_free(ptr);
213 }
214 
215 /**
216  * DPDK callback to close the device.
217  *
218  * Destroy all queues and objects, free memory.
219  *
220  * @param dev
221  *   Pointer to Ethernet device structure.
222  */
223 static void
224 mlx5_dev_close(struct rte_eth_dev *dev)
225 {
226 	struct priv *priv = dev->data->dev_private;
227 	unsigned int i;
228 	int ret;
229 
230 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
231 		dev->data->port_id,
232 		((priv->ctx != NULL) ? priv->ctx->device->name : ""));
233 	/* In case mlx5_dev_stop() has not been called. */
234 	mlx5_dev_interrupt_handler_uninstall(dev);
235 	mlx5_traffic_disable(dev);
236 	/* Prevent crashes when queues are still in use. */
237 	dev->rx_pkt_burst = removed_rx_burst;
238 	dev->tx_pkt_burst = removed_tx_burst;
239 	if (priv->rxqs != NULL) {
240 		/* XXX race condition if mlx5_rx_burst() is still running. */
241 		usleep(1000);
242 		for (i = 0; (i != priv->rxqs_n); ++i)
243 			mlx5_rxq_release(dev, i);
244 		priv->rxqs_n = 0;
245 		priv->rxqs = NULL;
246 	}
247 	if (priv->txqs != NULL) {
248 		/* XXX race condition if mlx5_tx_burst() is still running. */
249 		usleep(1000);
250 		for (i = 0; (i != priv->txqs_n); ++i)
251 			mlx5_txq_release(dev, i);
252 		priv->txqs_n = 0;
253 		priv->txqs = NULL;
254 	}
255 	mlx5_flow_delete_drop_queue(dev);
256 	mlx5_mprq_free_mp(dev);
257 	mlx5_mr_release(dev);
258 	if (priv->pd != NULL) {
259 		assert(priv->ctx != NULL);
260 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
261 		claim_zero(mlx5_glue->close_device(priv->ctx));
262 	} else
263 		assert(priv->ctx == NULL);
264 	if (priv->rss_conf.rss_key != NULL)
265 		rte_free(priv->rss_conf.rss_key);
266 	if (priv->reta_idx != NULL)
267 		rte_free(priv->reta_idx);
268 	if (priv->primary_socket)
269 		mlx5_socket_uninit(dev);
270 	if (priv->config.vf)
271 		mlx5_nl_mac_addr_flush(dev);
272 	if (priv->nl_socket >= 0)
273 		close(priv->nl_socket);
274 	ret = mlx5_hrxq_ibv_verify(dev);
275 	if (ret)
276 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
277 			dev->data->port_id);
278 	ret = mlx5_ind_table_ibv_verify(dev);
279 	if (ret)
280 		DRV_LOG(WARNING, "port %u some indirection table still remain",
281 			dev->data->port_id);
282 	ret = mlx5_rxq_ibv_verify(dev);
283 	if (ret)
284 		DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
285 			dev->data->port_id);
286 	ret = mlx5_rxq_verify(dev);
287 	if (ret)
288 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
289 			dev->data->port_id);
290 	ret = mlx5_txq_ibv_verify(dev);
291 	if (ret)
292 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
293 			dev->data->port_id);
294 	ret = mlx5_txq_verify(dev);
295 	if (ret)
296 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
297 			dev->data->port_id);
298 	ret = mlx5_flow_verify(dev);
299 	if (ret)
300 		DRV_LOG(WARNING, "port %u some flows still remain",
301 			dev->data->port_id);
302 	memset(priv, 0, sizeof(*priv));
303 }
304 
305 const struct eth_dev_ops mlx5_dev_ops = {
306 	.dev_configure = mlx5_dev_configure,
307 	.dev_start = mlx5_dev_start,
308 	.dev_stop = mlx5_dev_stop,
309 	.dev_set_link_down = mlx5_set_link_down,
310 	.dev_set_link_up = mlx5_set_link_up,
311 	.dev_close = mlx5_dev_close,
312 	.promiscuous_enable = mlx5_promiscuous_enable,
313 	.promiscuous_disable = mlx5_promiscuous_disable,
314 	.allmulticast_enable = mlx5_allmulticast_enable,
315 	.allmulticast_disable = mlx5_allmulticast_disable,
316 	.link_update = mlx5_link_update,
317 	.stats_get = mlx5_stats_get,
318 	.stats_reset = mlx5_stats_reset,
319 	.xstats_get = mlx5_xstats_get,
320 	.xstats_reset = mlx5_xstats_reset,
321 	.xstats_get_names = mlx5_xstats_get_names,
322 	.dev_infos_get = mlx5_dev_infos_get,
323 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
324 	.vlan_filter_set = mlx5_vlan_filter_set,
325 	.rx_queue_setup = mlx5_rx_queue_setup,
326 	.tx_queue_setup = mlx5_tx_queue_setup,
327 	.rx_queue_release = mlx5_rx_queue_release,
328 	.tx_queue_release = mlx5_tx_queue_release,
329 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
330 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
331 	.mac_addr_remove = mlx5_mac_addr_remove,
332 	.mac_addr_add = mlx5_mac_addr_add,
333 	.mac_addr_set = mlx5_mac_addr_set,
334 	.set_mc_addr_list = mlx5_set_mc_addr_list,
335 	.mtu_set = mlx5_dev_set_mtu,
336 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
337 	.vlan_offload_set = mlx5_vlan_offload_set,
338 	.reta_update = mlx5_dev_rss_reta_update,
339 	.reta_query = mlx5_dev_rss_reta_query,
340 	.rss_hash_update = mlx5_rss_hash_update,
341 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
342 	.filter_ctrl = mlx5_dev_filter_ctrl,
343 	.rx_descriptor_status = mlx5_rx_descriptor_status,
344 	.tx_descriptor_status = mlx5_tx_descriptor_status,
345 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
346 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
347 	.is_removed = mlx5_is_removed,
348 };
349 
350 static const struct eth_dev_ops mlx5_dev_sec_ops = {
351 	.stats_get = mlx5_stats_get,
352 	.stats_reset = mlx5_stats_reset,
353 	.xstats_get = mlx5_xstats_get,
354 	.xstats_reset = mlx5_xstats_reset,
355 	.xstats_get_names = mlx5_xstats_get_names,
356 	.dev_infos_get = mlx5_dev_infos_get,
357 	.rx_descriptor_status = mlx5_rx_descriptor_status,
358 	.tx_descriptor_status = mlx5_tx_descriptor_status,
359 };
360 
361 /* Available operators in flow isolated mode. */
362 const struct eth_dev_ops mlx5_dev_ops_isolate = {
363 	.dev_configure = mlx5_dev_configure,
364 	.dev_start = mlx5_dev_start,
365 	.dev_stop = mlx5_dev_stop,
366 	.dev_set_link_down = mlx5_set_link_down,
367 	.dev_set_link_up = mlx5_set_link_up,
368 	.dev_close = mlx5_dev_close,
369 	.link_update = mlx5_link_update,
370 	.stats_get = mlx5_stats_get,
371 	.stats_reset = mlx5_stats_reset,
372 	.xstats_get = mlx5_xstats_get,
373 	.xstats_reset = mlx5_xstats_reset,
374 	.xstats_get_names = mlx5_xstats_get_names,
375 	.dev_infos_get = mlx5_dev_infos_get,
376 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
377 	.vlan_filter_set = mlx5_vlan_filter_set,
378 	.rx_queue_setup = mlx5_rx_queue_setup,
379 	.tx_queue_setup = mlx5_tx_queue_setup,
380 	.rx_queue_release = mlx5_rx_queue_release,
381 	.tx_queue_release = mlx5_tx_queue_release,
382 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
383 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
384 	.mac_addr_remove = mlx5_mac_addr_remove,
385 	.mac_addr_add = mlx5_mac_addr_add,
386 	.mac_addr_set = mlx5_mac_addr_set,
387 	.set_mc_addr_list = mlx5_set_mc_addr_list,
388 	.mtu_set = mlx5_dev_set_mtu,
389 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
390 	.vlan_offload_set = mlx5_vlan_offload_set,
391 	.filter_ctrl = mlx5_dev_filter_ctrl,
392 	.rx_descriptor_status = mlx5_rx_descriptor_status,
393 	.tx_descriptor_status = mlx5_tx_descriptor_status,
394 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
395 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
396 	.is_removed = mlx5_is_removed,
397 };
398 
399 static struct {
400 	struct rte_pci_addr pci_addr; /* associated PCI address */
401 	uint32_t ports; /* physical ports bitfield. */
402 } mlx5_dev[32];
403 
404 /**
405  * Get device index in mlx5_dev[] from PCI bus address.
406  *
407  * @param[in] pci_addr
408  *   PCI bus address to look for.
409  *
410  * @return
411  *   mlx5_dev[] index on success, -1 on failure.
412  */
413 static int
414 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
415 {
416 	unsigned int i;
417 	int ret = -1;
418 
419 	assert(pci_addr != NULL);
420 	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
421 		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
422 		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
423 		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
424 		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
425 			return i;
426 		if ((mlx5_dev[i].ports == 0) && (ret == -1))
427 			ret = i;
428 	}
429 	return ret;
430 }
431 
432 /**
433  * Verify and store value for device argument.
434  *
435  * @param[in] key
436  *   Key argument to verify.
437  * @param[in] val
438  *   Value associated with key.
439  * @param opaque
440  *   User data.
441  *
442  * @return
443  *   0 on success, a negative errno value otherwise and rte_errno is set.
444  */
445 static int
446 mlx5_args_check(const char *key, const char *val, void *opaque)
447 {
448 	struct mlx5_dev_config *config = opaque;
449 	unsigned long tmp;
450 
451 	errno = 0;
452 	tmp = strtoul(val, NULL, 0);
453 	if (errno) {
454 		rte_errno = errno;
455 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
456 		return -rte_errno;
457 	}
458 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
459 		config->cqe_comp = !!tmp;
460 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
461 		config->mprq.enabled = !!tmp;
462 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
463 		config->mprq.stride_num_n = tmp;
464 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
465 		config->mprq.max_memcpy_len = tmp;
466 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
467 		config->mprq.min_rxqs_num = tmp;
468 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
469 		config->txq_inline = tmp;
470 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
471 		config->txqs_inline = tmp;
472 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
473 		config->mps = !!tmp ? config->mps : 0;
474 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
475 		config->mpw_hdr_dseg = !!tmp;
476 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
477 		config->inline_max_packet_sz = tmp;
478 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
479 		config->tx_vec_en = !!tmp;
480 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
481 		config->rx_vec_en = !!tmp;
482 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
483 		config->l3_vxlan_en = !!tmp;
484 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
485 		config->vf_nl_en = !!tmp;
486 	} else {
487 		DRV_LOG(WARNING, "%s: unknown parameter", key);
488 		rte_errno = EINVAL;
489 		return -rte_errno;
490 	}
491 	return 0;
492 }
493 
494 /**
495  * Parse device parameters.
496  *
497  * @param config
498  *   Pointer to device configuration structure.
499  * @param devargs
500  *   Device arguments structure.
501  *
502  * @return
503  *   0 on success, a negative errno value otherwise and rte_errno is set.
504  */
505 static int
506 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
507 {
508 	const char **params = (const char *[]){
509 		MLX5_RXQ_CQE_COMP_EN,
510 		MLX5_RX_MPRQ_EN,
511 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
512 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
513 		MLX5_RXQS_MIN_MPRQ,
514 		MLX5_TXQ_INLINE,
515 		MLX5_TXQS_MIN_INLINE,
516 		MLX5_TXQ_MPW_EN,
517 		MLX5_TXQ_MPW_HDR_DSEG_EN,
518 		MLX5_TXQ_MAX_INLINE_LEN,
519 		MLX5_TX_VEC_EN,
520 		MLX5_RX_VEC_EN,
521 		MLX5_L3_VXLAN_EN,
522 		MLX5_VF_NL_EN,
523 		NULL,
524 	};
525 	struct rte_kvargs *kvlist;
526 	int ret = 0;
527 	int i;
528 
529 	if (devargs == NULL)
530 		return 0;
531 	/* Following UGLY cast is done to pass checkpatch. */
532 	kvlist = rte_kvargs_parse(devargs->args, params);
533 	if (kvlist == NULL)
534 		return 0;
535 	/* Process parameters. */
536 	for (i = 0; (params[i] != NULL); ++i) {
537 		if (rte_kvargs_count(kvlist, params[i])) {
538 			ret = rte_kvargs_process(kvlist, params[i],
539 						 mlx5_args_check, config);
540 			if (ret) {
541 				rte_errno = EINVAL;
542 				rte_kvargs_free(kvlist);
543 				return -rte_errno;
544 			}
545 		}
546 	}
547 	rte_kvargs_free(kvlist);
548 	return 0;
549 }
550 
551 static struct rte_pci_driver mlx5_driver;
552 
553 /*
554  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
555  * local resource used by both primary and secondary to avoid duplicate
556  * reservation.
557  * The space has to be available on both primary and secondary process,
558  * TXQ UAR maps to this area using fixed mmap w/o double check.
559  */
560 static void *uar_base;
561 
562 static int
563 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
564 		const struct rte_memseg *ms, void *arg)
565 {
566 	void **addr = arg;
567 
568 	if (*addr == NULL)
569 		*addr = ms->addr;
570 	else
571 		*addr = RTE_MIN(*addr, ms->addr);
572 
573 	return 0;
574 }
575 
576 /**
577  * Reserve UAR address space for primary process.
578  *
579  * @param[in] dev
580  *   Pointer to Ethernet device.
581  *
582  * @return
583  *   0 on success, a negative errno value otherwise and rte_errno is set.
584  */
585 static int
586 mlx5_uar_init_primary(struct rte_eth_dev *dev)
587 {
588 	struct priv *priv = dev->data->dev_private;
589 	void *addr = (void *)0;
590 
591 	if (uar_base) { /* UAR address space mapped. */
592 		priv->uar_base = uar_base;
593 		return 0;
594 	}
595 	/* find out lower bound of hugepage segments */
596 	rte_memseg_walk(find_lower_va_bound, &addr);
597 
598 	/* keep distance to hugepages to minimize potential conflicts. */
599 	addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
600 	/* anonymous mmap, no real memory consumption. */
601 	addr = mmap(addr, MLX5_UAR_SIZE,
602 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
603 	if (addr == MAP_FAILED) {
604 		DRV_LOG(ERR,
605 			"port %u failed to reserve UAR address space, please"
606 			" adjust MLX5_UAR_SIZE or try --base-virtaddr",
607 			dev->data->port_id);
608 		rte_errno = ENOMEM;
609 		return -rte_errno;
610 	}
611 	/* Accept either same addr or a new addr returned from mmap if target
612 	 * range occupied.
613 	 */
614 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
615 		dev->data->port_id, addr);
616 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
617 	uar_base = addr; /* process local, don't reserve again. */
618 	return 0;
619 }
620 
621 /**
622  * Reserve UAR address space for secondary process, align with
623  * primary process.
624  *
625  * @param[in] dev
626  *   Pointer to Ethernet device.
627  *
628  * @return
629  *   0 on success, a negative errno value otherwise and rte_errno is set.
630  */
631 static int
632 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
633 {
634 	struct priv *priv = dev->data->dev_private;
635 	void *addr;
636 
637 	assert(priv->uar_base);
638 	if (uar_base) { /* already reserved. */
639 		assert(uar_base == priv->uar_base);
640 		return 0;
641 	}
642 	/* anonymous mmap, no real memory consumption. */
643 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
644 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
645 	if (addr == MAP_FAILED) {
646 		DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
647 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
648 		rte_errno = ENXIO;
649 		return -rte_errno;
650 	}
651 	if (priv->uar_base != addr) {
652 		DRV_LOG(ERR,
653 			"port %u UAR address %p size %llu occupied, please"
654 			" adjust MLX5_UAR_OFFSET or try EAL parameter"
655 			" --base-virtaddr",
656 			dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
657 		rte_errno = ENXIO;
658 		return -rte_errno;
659 	}
660 	uar_base = addr; /* process local, don't reserve again */
661 	DRV_LOG(INFO, "port %u reserved UAR address space: %p",
662 		dev->data->port_id, addr);
663 	return 0;
664 }
665 
666 /**
667  * DPDK callback to register a PCI device.
668  *
669  * This function creates an Ethernet device for each port of a given
670  * PCI device.
671  *
672  * @param[in] pci_drv
673  *   PCI driver structure (mlx5_driver).
674  * @param[in] pci_dev
675  *   PCI device information.
676  *
677  * @return
678  *   0 on success, a negative errno value otherwise and rte_errno is set.
679  */
680 static int
681 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
682 	       struct rte_pci_device *pci_dev)
683 {
684 	struct ibv_device **list = NULL;
685 	struct ibv_device *ibv_dev;
686 	int err = 0;
687 	struct ibv_context *attr_ctx = NULL;
688 	struct ibv_device_attr_ex device_attr;
689 	unsigned int vf = 0;
690 	unsigned int mps;
691 	unsigned int cqe_comp;
692 	unsigned int tunnel_en = 0;
693 	unsigned int mpls_en = 0;
694 	unsigned int swp = 0;
695 	unsigned int verb_priorities = 0;
696 	unsigned int mprq = 0;
697 	unsigned int mprq_min_stride_size_n = 0;
698 	unsigned int mprq_max_stride_size_n = 0;
699 	unsigned int mprq_min_stride_num_n = 0;
700 	unsigned int mprq_max_stride_num_n = 0;
701 	int idx;
702 	int i;
703 	struct mlx5dv_context attrs_out = {0};
704 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
705 	struct ibv_counter_set_description cs_desc;
706 #endif
707 
708 	/* Prepare shared data between primary and secondary process. */
709 	mlx5_prepare_shared_data();
710 	assert(pci_drv == &mlx5_driver);
711 	/* Get mlx5_dev[] index. */
712 	idx = mlx5_dev_idx(&pci_dev->addr);
713 	if (idx == -1) {
714 		DRV_LOG(ERR, "this driver cannot support any more adapters");
715 		err = ENOMEM;
716 		goto error;
717 	}
718 	DRV_LOG(DEBUG, "using driver device index %d", idx);
719 	/* Save PCI address. */
720 	mlx5_dev[idx].pci_addr = pci_dev->addr;
721 	list = mlx5_glue->get_device_list(&i);
722 	if (list == NULL) {
723 		assert(errno);
724 		err = errno;
725 		if (errno == ENOSYS)
726 			DRV_LOG(ERR,
727 				"cannot list devices, is ib_uverbs loaded?");
728 		goto error;
729 	}
730 	assert(i >= 0);
731 	/*
732 	 * For each listed device, check related sysfs entry against
733 	 * the provided PCI ID.
734 	 */
735 	while (i != 0) {
736 		struct rte_pci_addr pci_addr;
737 
738 		--i;
739 		DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
740 		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
741 			continue;
742 		if ((pci_dev->addr.domain != pci_addr.domain) ||
743 		    (pci_dev->addr.bus != pci_addr.bus) ||
744 		    (pci_dev->addr.devid != pci_addr.devid) ||
745 		    (pci_dev->addr.function != pci_addr.function))
746 			continue;
747 		DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
748 			list[i]->name);
749 		vf = ((pci_dev->id.device_id ==
750 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
751 		      (pci_dev->id.device_id ==
752 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
753 		      (pci_dev->id.device_id ==
754 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
755 		      (pci_dev->id.device_id ==
756 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
757 		attr_ctx = mlx5_glue->open_device(list[i]);
758 		rte_errno = errno;
759 		err = rte_errno;
760 		break;
761 	}
762 	if (attr_ctx == NULL) {
763 		switch (err) {
764 		case 0:
765 			DRV_LOG(ERR,
766 				"cannot access device, is mlx5_ib loaded?");
767 			err = ENODEV;
768 			break;
769 		case EINVAL:
770 			DRV_LOG(ERR,
771 				"cannot use device, are drivers up to date?");
772 			break;
773 		}
774 		goto error;
775 	}
776 	ibv_dev = list[i];
777 	DRV_LOG(DEBUG, "device opened");
778 #ifdef HAVE_IBV_MLX5_MOD_SWP
779 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
780 #endif
781 	/*
782 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
783 	 * as all ConnectX-5 devices.
784 	 */
785 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
786 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
787 #endif
788 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
789 	attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
790 #endif
791 	mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
792 	if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
793 		if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
794 			DRV_LOG(DEBUG, "enhanced MPW is supported");
795 			mps = MLX5_MPW_ENHANCED;
796 		} else {
797 			DRV_LOG(DEBUG, "MPW is supported");
798 			mps = MLX5_MPW;
799 		}
800 	} else {
801 		DRV_LOG(DEBUG, "MPW isn't supported");
802 		mps = MLX5_MPW_DISABLED;
803 	}
804 #ifdef HAVE_IBV_MLX5_MOD_SWP
805 	if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
806 		swp = attrs_out.sw_parsing_caps.sw_parsing_offloads;
807 	DRV_LOG(DEBUG, "SWP support: %u", swp);
808 #endif
809 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
810 	if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
811 		struct mlx5dv_striding_rq_caps mprq_caps =
812 			attrs_out.striding_rq_caps;
813 
814 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
815 			mprq_caps.min_single_stride_log_num_of_bytes);
816 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
817 			mprq_caps.max_single_stride_log_num_of_bytes);
818 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
819 			mprq_caps.min_single_wqe_log_num_of_strides);
820 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
821 			mprq_caps.max_single_wqe_log_num_of_strides);
822 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
823 			mprq_caps.supported_qpts);
824 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
825 		mprq = 1;
826 		mprq_min_stride_size_n =
827 			mprq_caps.min_single_stride_log_num_of_bytes;
828 		mprq_max_stride_size_n =
829 			mprq_caps.max_single_stride_log_num_of_bytes;
830 		mprq_min_stride_num_n =
831 			mprq_caps.min_single_wqe_log_num_of_strides;
832 		mprq_max_stride_num_n =
833 			mprq_caps.max_single_wqe_log_num_of_strides;
834 	}
835 #endif
836 	if (RTE_CACHE_LINE_SIZE == 128 &&
837 	    !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
838 		cqe_comp = 0;
839 	else
840 		cqe_comp = 1;
841 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
842 	if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
843 		tunnel_en = ((attrs_out.tunnel_offloads_caps &
844 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
845 			     (attrs_out.tunnel_offloads_caps &
846 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
847 	}
848 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
849 		tunnel_en ? "" : "not ");
850 #else
851 	DRV_LOG(WARNING,
852 		"tunnel offloading disabled due to old OFED/rdma-core version");
853 #endif
854 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
855 	mpls_en = ((attrs_out.tunnel_offloads_caps &
856 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
857 		   (attrs_out.tunnel_offloads_caps &
858 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
859 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
860 		mpls_en ? "" : "not ");
861 #else
862 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
863 		" old OFED/rdma-core version or firmware configuration");
864 #endif
865 	err = mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr);
866 	if (err) {
867 		DEBUG("ibv_query_device_ex() failed");
868 		goto error;
869 	}
870 	DRV_LOG(INFO, "%u port(s) detected",
871 		device_attr.orig_attr.phys_port_cnt);
872 	for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
873 		char name[RTE_ETH_NAME_MAX_LEN];
874 		int len;
875 		uint32_t port = i + 1; /* ports are indexed from one */
876 		uint32_t test = (1 << i);
877 		struct ibv_context *ctx = NULL;
878 		struct ibv_port_attr port_attr;
879 		struct ibv_pd *pd = NULL;
880 		struct priv *priv = NULL;
881 		struct rte_eth_dev *eth_dev = NULL;
882 		struct ibv_device_attr_ex device_attr_ex;
883 		struct ether_addr mac;
884 		struct mlx5_dev_config config = {
885 			.cqe_comp = cqe_comp,
886 			.mps = mps,
887 			.tunnel_en = tunnel_en,
888 			.mpls_en = mpls_en,
889 			.tx_vec_en = 1,
890 			.rx_vec_en = 1,
891 			.mpw_hdr_dseg = 0,
892 			.txq_inline = MLX5_ARG_UNSET,
893 			.txqs_inline = MLX5_ARG_UNSET,
894 			.inline_max_packet_sz = MLX5_ARG_UNSET,
895 			.vf_nl_en = 1,
896 			.swp = !!swp,
897 			.mprq = {
898 				.enabled = 0, /* Disabled by default. */
899 				.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
900 							mprq_min_stride_num_n),
901 				.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
902 				.min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
903 			},
904 		};
905 
906 		len = snprintf(name, sizeof(name), PCI_PRI_FMT,
907 			 pci_dev->addr.domain, pci_dev->addr.bus,
908 			 pci_dev->addr.devid, pci_dev->addr.function);
909 		if (device_attr.orig_attr.phys_port_cnt > 1)
910 			snprintf(name + len, sizeof(name), " port %u", i);
911 		mlx5_dev[idx].ports |= test;
912 		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
913 			eth_dev = rte_eth_dev_attach_secondary(name);
914 			if (eth_dev == NULL) {
915 				DRV_LOG(ERR, "can not attach rte ethdev");
916 				rte_errno = ENOMEM;
917 				err = rte_errno;
918 				goto error;
919 			}
920 			eth_dev->device = &pci_dev->device;
921 			eth_dev->dev_ops = &mlx5_dev_sec_ops;
922 			err = mlx5_uar_init_secondary(eth_dev);
923 			if (err) {
924 				err = rte_errno;
925 				goto error;
926 			}
927 			/* Receive command fd from primary process */
928 			err = mlx5_socket_connect(eth_dev);
929 			if (err < 0) {
930 				err = rte_errno;
931 				goto error;
932 			}
933 			/* Remap UAR for Tx queues. */
934 			err = mlx5_tx_uar_remap(eth_dev, err);
935 			if (err) {
936 				err = rte_errno;
937 				goto error;
938 			}
939 			/*
940 			 * Ethdev pointer is still required as input since
941 			 * the primary device is not accessible from the
942 			 * secondary process.
943 			 */
944 			eth_dev->rx_pkt_burst =
945 				mlx5_select_rx_function(eth_dev);
946 			eth_dev->tx_pkt_burst =
947 				mlx5_select_tx_function(eth_dev);
948 			rte_eth_dev_probing_finish(eth_dev);
949 			continue;
950 		}
951 		DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
952 		ctx = mlx5_glue->open_device(ibv_dev);
953 		if (ctx == NULL) {
954 			err = ENODEV;
955 			goto port_error;
956 		}
957 		/* Check port status. */
958 		err = mlx5_glue->query_port(ctx, port, &port_attr);
959 		if (err) {
960 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
961 			goto port_error;
962 		}
963 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
964 			DRV_LOG(ERR,
965 				"port %d is not configured in Ethernet mode",
966 				port);
967 			err = EINVAL;
968 			goto port_error;
969 		}
970 		if (port_attr.state != IBV_PORT_ACTIVE)
971 			DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
972 				port,
973 				mlx5_glue->port_state_str(port_attr.state),
974 				port_attr.state);
975 		/* Allocate protection domain. */
976 		pd = mlx5_glue->alloc_pd(ctx);
977 		if (pd == NULL) {
978 			DRV_LOG(ERR, "PD allocation failure");
979 			err = ENOMEM;
980 			goto port_error;
981 		}
982 		mlx5_dev[idx].ports |= test;
983 		/* from rte_ethdev.c */
984 		priv = rte_zmalloc("ethdev private structure",
985 				   sizeof(*priv),
986 				   RTE_CACHE_LINE_SIZE);
987 		if (priv == NULL) {
988 			DRV_LOG(ERR, "priv allocation failure");
989 			err = ENOMEM;
990 			goto port_error;
991 		}
992 		priv->ctx = ctx;
993 		strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
994 			sizeof(priv->ibdev_path));
995 		priv->device_attr = device_attr;
996 		priv->port = port;
997 		priv->pd = pd;
998 		priv->mtu = ETHER_MTU;
999 		err = mlx5_args(&config, pci_dev->device.devargs);
1000 		if (err) {
1001 			DRV_LOG(ERR, "failed to process device arguments: %s",
1002 				strerror(err));
1003 			err = rte_errno;
1004 			goto port_error;
1005 		}
1006 		err = mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex);
1007 		if (err) {
1008 			DRV_LOG(ERR, "ibv_query_device_ex() failed");
1009 			goto port_error;
1010 		}
1011 		config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
1012 				    IBV_DEVICE_RAW_IP_CSUM);
1013 		DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1014 			(config.hw_csum ? "" : "not "));
1015 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1016 		config.flow_counter_en = !!(device_attr.max_counter_sets);
1017 		mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
1018 		DRV_LOG(DEBUG,
1019 			"counter type = %d, num of cs = %ld, attributes = %d",
1020 			cs_desc.counter_type, cs_desc.num_of_cs,
1021 			cs_desc.attributes);
1022 #endif
1023 		config.ind_table_max_size =
1024 			device_attr_ex.rss_caps.max_rwq_indirection_table_size;
1025 		/* Remove this check once DPDK supports larger/variable
1026 		 * indirection tables. */
1027 		if (config.ind_table_max_size >
1028 				(unsigned int)ETH_RSS_RETA_SIZE_512)
1029 			config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1030 		DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1031 			config.ind_table_max_size);
1032 		config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
1033 					 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1034 		DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1035 			(config.hw_vlan_strip ? "" : "not "));
1036 
1037 		config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
1038 					 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1039 		DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1040 			(config.hw_fcs_strip ? "" : "not "));
1041 
1042 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
1043 		config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
1044 #endif
1045 		DRV_LOG(DEBUG,
1046 			"hardware Rx end alignment padding is %ssupported",
1047 			(config.hw_padding ? "" : "not "));
1048 		config.vf = vf;
1049 		config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
1050 			      (device_attr_ex.tso_caps.supported_qpts &
1051 			      (1 << IBV_QPT_RAW_PACKET)));
1052 		if (config.tso)
1053 			config.tso_max_payload_sz =
1054 					device_attr_ex.tso_caps.max_tso;
1055 		if (config.mps && !mps) {
1056 			DRV_LOG(ERR,
1057 				"multi-packet send not supported on this device"
1058 				" (" MLX5_TXQ_MPW_EN ")");
1059 			err = ENOTSUP;
1060 			goto port_error;
1061 		}
1062 		DRV_LOG(INFO, "%s MPS is %s",
1063 			config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1064 			config.mps != MLX5_MPW_DISABLED ? "enabled" :
1065 			"disabled");
1066 		if (config.cqe_comp && !cqe_comp) {
1067 			DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1068 			config.cqe_comp = 0;
1069 		}
1070 		config.mprq.enabled = config.mprq.enabled && mprq;
1071 		if (config.mprq.enabled) {
1072 			if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1073 			    config.mprq.stride_num_n < mprq_min_stride_num_n) {
1074 				config.mprq.stride_num_n =
1075 					RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1076 						mprq_min_stride_num_n);
1077 				DRV_LOG(WARNING,
1078 					"the number of strides"
1079 					" for Multi-Packet RQ is out of range,"
1080 					" setting default value (%u)",
1081 					1 << config.mprq.stride_num_n);
1082 			}
1083 			config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1084 			config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1085 		}
1086 		eth_dev = rte_eth_dev_allocate(name);
1087 		if (eth_dev == NULL) {
1088 			DRV_LOG(ERR, "can not allocate rte ethdev");
1089 			err = ENOMEM;
1090 			goto port_error;
1091 		}
1092 		eth_dev->data->dev_private = priv;
1093 		priv->dev_data = eth_dev->data;
1094 		eth_dev->data->mac_addrs = priv->mac;
1095 		eth_dev->device = &pci_dev->device;
1096 		rte_eth_copy_pci_info(eth_dev, pci_dev);
1097 		eth_dev->device->driver = &mlx5_driver.driver;
1098 		err = mlx5_uar_init_primary(eth_dev);
1099 		if (err) {
1100 			err = rte_errno;
1101 			goto port_error;
1102 		}
1103 		/* Configure the first MAC address by default. */
1104 		if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1105 			DRV_LOG(ERR,
1106 				"port %u cannot get MAC address, is mlx5_en"
1107 				" loaded? (errno: %s)",
1108 				eth_dev->data->port_id, strerror(errno));
1109 			err = ENODEV;
1110 			goto port_error;
1111 		}
1112 		DRV_LOG(INFO,
1113 			"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1114 			eth_dev->data->port_id,
1115 			mac.addr_bytes[0], mac.addr_bytes[1],
1116 			mac.addr_bytes[2], mac.addr_bytes[3],
1117 			mac.addr_bytes[4], mac.addr_bytes[5]);
1118 #ifndef NDEBUG
1119 		{
1120 			char ifname[IF_NAMESIZE];
1121 
1122 			if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1123 				DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1124 					eth_dev->data->port_id, ifname);
1125 			else
1126 				DRV_LOG(DEBUG, "port %u ifname is unknown",
1127 					eth_dev->data->port_id);
1128 		}
1129 #endif
1130 		/* Get actual MTU if possible. */
1131 		err = mlx5_get_mtu(eth_dev, &priv->mtu);
1132 		if (err) {
1133 			err = rte_errno;
1134 			goto port_error;
1135 		}
1136 		DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1137 			priv->mtu);
1138 		/*
1139 		 * Initialize burst functions to prevent crashes before link-up.
1140 		 */
1141 		eth_dev->rx_pkt_burst = removed_rx_burst;
1142 		eth_dev->tx_pkt_burst = removed_tx_burst;
1143 		eth_dev->dev_ops = &mlx5_dev_ops;
1144 		/* Register MAC address. */
1145 		claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1146 		priv->nl_socket = -1;
1147 		priv->nl_sn = 0;
1148 		if (vf && config.vf_nl_en) {
1149 			priv->nl_socket = mlx5_nl_init(RTMGRP_LINK);
1150 			if (priv->nl_socket < 0)
1151 				priv->nl_socket = -1;
1152 			mlx5_nl_mac_addr_sync(eth_dev);
1153 		}
1154 		TAILQ_INIT(&priv->flows);
1155 		TAILQ_INIT(&priv->ctrl_flows);
1156 		/* Hint libmlx5 to use PMD allocator for data plane resources */
1157 		struct mlx5dv_ctx_allocators alctr = {
1158 			.alloc = &mlx5_alloc_verbs_buf,
1159 			.free = &mlx5_free_verbs_buf,
1160 			.data = priv,
1161 		};
1162 		mlx5_glue->dv_set_context_attr(ctx,
1163 					       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1164 					       (void *)((uintptr_t)&alctr));
1165 		/* Bring Ethernet device up. */
1166 		DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1167 			eth_dev->data->port_id);
1168 		mlx5_set_link_up(eth_dev);
1169 		/*
1170 		 * Even though the interrupt handler is not installed yet,
1171 		 * interrupts will still trigger on the asyn_fd from
1172 		 * Verbs context returned by ibv_open_device().
1173 		 */
1174 		mlx5_link_update(eth_dev, 0);
1175 		/* Store device configuration on private structure. */
1176 		priv->config = config;
1177 		/* Create drop queue. */
1178 		err = mlx5_flow_create_drop_queue(eth_dev);
1179 		if (err) {
1180 			DRV_LOG(ERR, "port %u drop queue allocation failed: %s",
1181 				eth_dev->data->port_id, strerror(rte_errno));
1182 			err = rte_errno;
1183 			goto port_error;
1184 		}
1185 		/* Supported Verbs flow priority number detection. */
1186 		if (verb_priorities == 0)
1187 			verb_priorities = mlx5_get_max_verbs_prio(eth_dev);
1188 		if (verb_priorities < MLX5_VERBS_FLOW_PRIO_8) {
1189 			DRV_LOG(ERR, "port %u wrong Verbs flow priorities: %u",
1190 				eth_dev->data->port_id, verb_priorities);
1191 			goto port_error;
1192 		}
1193 		priv->config.max_verbs_prio = verb_priorities;
1194 		rte_eth_dev_probing_finish(eth_dev);
1195 		continue;
1196 port_error:
1197 		if (priv)
1198 			rte_free(priv);
1199 		if (pd)
1200 			claim_zero(mlx5_glue->dealloc_pd(pd));
1201 		if (ctx)
1202 			claim_zero(mlx5_glue->close_device(ctx));
1203 		if (eth_dev && rte_eal_process_type() == RTE_PROC_PRIMARY)
1204 			rte_eth_dev_release_port(eth_dev);
1205 		break;
1206 	}
1207 	/*
1208 	 * XXX if something went wrong in the loop above, there is a resource
1209 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1210 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
1211 	 * way to enumerate the registered ethdevs to free the previous ones.
1212 	 */
1213 	/* no port found, complain */
1214 	if (!mlx5_dev[idx].ports) {
1215 		rte_errno = ENODEV;
1216 		err = rte_errno;
1217 	}
1218 error:
1219 	if (attr_ctx)
1220 		claim_zero(mlx5_glue->close_device(attr_ctx));
1221 	if (list)
1222 		mlx5_glue->free_device_list(list);
1223 	if (err) {
1224 		rte_errno = err;
1225 		return -rte_errno;
1226 	}
1227 	return 0;
1228 }
1229 
1230 static const struct rte_pci_id mlx5_pci_id_map[] = {
1231 	{
1232 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1233 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1234 	},
1235 	{
1236 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1237 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1238 	},
1239 	{
1240 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1241 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1242 	},
1243 	{
1244 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1245 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1246 	},
1247 	{
1248 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1249 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1250 	},
1251 	{
1252 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1253 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1254 	},
1255 	{
1256 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1257 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1258 	},
1259 	{
1260 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1261 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1262 	},
1263 	{
1264 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1265 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1266 	},
1267 	{
1268 		.vendor_id = 0
1269 	}
1270 };
1271 
1272 static struct rte_pci_driver mlx5_driver = {
1273 	.driver = {
1274 		.name = MLX5_DRIVER_NAME
1275 	},
1276 	.id_table = mlx5_pci_id_map,
1277 	.probe = mlx5_pci_probe,
1278 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1279 };
1280 
1281 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1282 
1283 /**
1284  * Suffix RTE_EAL_PMD_PATH with "-glue".
1285  *
1286  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1287  * suffixing its last component.
1288  *
1289  * @param buf[out]
1290  *   Output buffer, should be large enough otherwise NULL is returned.
1291  * @param size
1292  *   Size of @p out.
1293  *
1294  * @return
1295  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1296  */
1297 static char *
1298 mlx5_glue_path(char *buf, size_t size)
1299 {
1300 	static const char *const bad[] = { "/", ".", "..", NULL };
1301 	const char *path = RTE_EAL_PMD_PATH;
1302 	size_t len = strlen(path);
1303 	size_t off;
1304 	int i;
1305 
1306 	while (len && path[len - 1] == '/')
1307 		--len;
1308 	for (off = len; off && path[off - 1] != '/'; --off)
1309 		;
1310 	for (i = 0; bad[i]; ++i)
1311 		if (!strncmp(path + off, bad[i], (int)(len - off)))
1312 			goto error;
1313 	i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1314 	if (i == -1 || (size_t)i >= size)
1315 		goto error;
1316 	return buf;
1317 error:
1318 	DRV_LOG(ERR,
1319 		"unable to append \"-glue\" to last component of"
1320 		" RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1321 		" please re-configure DPDK");
1322 	return NULL;
1323 }
1324 
1325 /**
1326  * Initialization routine for run-time dependency on rdma-core.
1327  */
1328 static int
1329 mlx5_glue_init(void)
1330 {
1331 	char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1332 	const char *path[] = {
1333 		/*
1334 		 * A basic security check is necessary before trusting
1335 		 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1336 		 */
1337 		(geteuid() == getuid() && getegid() == getgid() ?
1338 		 getenv("MLX5_GLUE_PATH") : NULL),
1339 		/*
1340 		 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1341 		 * variant, otherwise let dlopen() look up libraries on its
1342 		 * own.
1343 		 */
1344 		(*RTE_EAL_PMD_PATH ?
1345 		 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1346 	};
1347 	unsigned int i = 0;
1348 	void *handle = NULL;
1349 	void **sym;
1350 	const char *dlmsg;
1351 
1352 	while (!handle && i != RTE_DIM(path)) {
1353 		const char *end;
1354 		size_t len;
1355 		int ret;
1356 
1357 		if (!path[i]) {
1358 			++i;
1359 			continue;
1360 		}
1361 		end = strpbrk(path[i], ":;");
1362 		if (!end)
1363 			end = path[i] + strlen(path[i]);
1364 		len = end - path[i];
1365 		ret = 0;
1366 		do {
1367 			char name[ret + 1];
1368 
1369 			ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1370 				       (int)len, path[i],
1371 				       (!len || *(end - 1) == '/') ? "" : "/");
1372 			if (ret == -1)
1373 				break;
1374 			if (sizeof(name) != (size_t)ret + 1)
1375 				continue;
1376 			DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1377 				name);
1378 			handle = dlopen(name, RTLD_LAZY);
1379 			break;
1380 		} while (1);
1381 		path[i] = end + 1;
1382 		if (!*end)
1383 			++i;
1384 	}
1385 	if (!handle) {
1386 		rte_errno = EINVAL;
1387 		dlmsg = dlerror();
1388 		if (dlmsg)
1389 			DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1390 		goto glue_error;
1391 	}
1392 	sym = dlsym(handle, "mlx5_glue");
1393 	if (!sym || !*sym) {
1394 		rte_errno = EINVAL;
1395 		dlmsg = dlerror();
1396 		if (dlmsg)
1397 			DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1398 		goto glue_error;
1399 	}
1400 	mlx5_glue = *sym;
1401 	return 0;
1402 glue_error:
1403 	if (handle)
1404 		dlclose(handle);
1405 	DRV_LOG(WARNING,
1406 		"cannot initialize PMD due to missing run-time dependency on"
1407 		" rdma-core libraries (libibverbs, libmlx5)");
1408 	return -rte_errno;
1409 }
1410 
1411 #endif
1412 
1413 /**
1414  * Driver initialization routine.
1415  */
1416 RTE_INIT(rte_mlx5_pmd_init);
1417 static void
1418 rte_mlx5_pmd_init(void)
1419 {
1420 	/* Build the static tables for Verbs conversion. */
1421 	mlx5_set_ptype_table();
1422 	mlx5_set_cksum_table();
1423 	mlx5_set_swp_types_table();
1424 	/*
1425 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1426 	 * huge pages. Calling ibv_fork_init() during init allows
1427 	 * applications to use fork() safely for purposes other than
1428 	 * using this PMD, which is not supported in forked processes.
1429 	 */
1430 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1431 	/* Match the size of Rx completion entry to the size of a cacheline. */
1432 	if (RTE_CACHE_LINE_SIZE == 128)
1433 		setenv("MLX5_CQE_SIZE", "128", 0);
1434 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1435 	if (mlx5_glue_init())
1436 		return;
1437 	assert(mlx5_glue);
1438 #endif
1439 #ifndef NDEBUG
1440 	/* Glue structure must not contain any NULL pointers. */
1441 	{
1442 		unsigned int i;
1443 
1444 		for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1445 			assert(((const void *const *)mlx5_glue)[i]);
1446 	}
1447 #endif
1448 	if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1449 		DRV_LOG(ERR,
1450 			"rdma-core glue \"%s\" mismatch: \"%s\" is required",
1451 			mlx5_glue->version, MLX5_GLUE_VERSION);
1452 		return;
1453 	}
1454 	mlx5_glue->fork_init();
1455 	rte_pci_register(&mlx5_driver);
1456 	rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1457 					mlx5_mr_mem_event_cb, NULL);
1458 }
1459 
1460 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1461 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1462 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1463 
1464 /** Initialize driver log type. */
1465 RTE_INIT(vdev_netvsc_init_log)
1466 {
1467 	mlx5_logtype = rte_log_register("pmd.net.mlx5");
1468 	if (mlx5_logtype >= 0)
1469 		rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1470 }
1471