xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 117eaa70584b73eebf6f648cf3ee6f2ab03264a0)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox.
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 
17 /* Verbs header. */
18 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
19 #ifdef PEDANTIC
20 #pragma GCC diagnostic ignored "-Wpedantic"
21 #endif
22 #include <infiniband/verbs.h>
23 #ifdef PEDANTIC
24 #pragma GCC diagnostic error "-Wpedantic"
25 #endif
26 
27 #include <rte_malloc.h>
28 #include <rte_ethdev_driver.h>
29 #include <rte_ethdev_pci.h>
30 #include <rte_pci.h>
31 #include <rte_bus_pci.h>
32 #include <rte_common.h>
33 #include <rte_config.h>
34 #include <rte_eal_memconfig.h>
35 #include <rte_kvargs.h>
36 
37 #include "mlx5.h"
38 #include "mlx5_utils.h"
39 #include "mlx5_rxtx.h"
40 #include "mlx5_autoconf.h"
41 #include "mlx5_defs.h"
42 #include "mlx5_glue.h"
43 
44 /* Device parameter to enable RX completion queue compression. */
45 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
46 
47 /* Device parameter to configure inline send. */
48 #define MLX5_TXQ_INLINE "txq_inline"
49 
50 /*
51  * Device parameter to configure the number of TX queues threshold for
52  * enabling inline send.
53  */
54 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
55 
56 /* Device parameter to enable multi-packet send WQEs. */
57 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
58 
59 /* Device parameter to include 2 dsegs in the title WQEBB. */
60 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
61 
62 /* Device parameter to limit the size of inlining packet. */
63 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
64 
65 /* Device parameter to enable hardware Tx vector. */
66 #define MLX5_TX_VEC_EN "tx_vec_en"
67 
68 /* Device parameter to enable hardware Rx vector. */
69 #define MLX5_RX_VEC_EN "rx_vec_en"
70 
71 #ifndef HAVE_IBV_MLX5_MOD_MPW
72 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
73 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
74 #endif
75 
76 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
77 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
78 #endif
79 
80 /**
81  * Retrieve integer value from environment variable.
82  *
83  * @param[in] name
84  *   Environment variable name.
85  *
86  * @return
87  *   Integer value, 0 if the variable is not set.
88  */
89 int
90 mlx5_getenv_int(const char *name)
91 {
92 	const char *val = getenv(name);
93 
94 	if (val == NULL)
95 		return 0;
96 	return atoi(val);
97 }
98 
99 /**
100  * Verbs callback to allocate a memory. This function should allocate the space
101  * according to the size provided residing inside a huge page.
102  * Please note that all allocation must respect the alignment from libmlx5
103  * (i.e. currently sysconf(_SC_PAGESIZE)).
104  *
105  * @param[in] size
106  *   The size in bytes of the memory to allocate.
107  * @param[in] data
108  *   A pointer to the callback data.
109  *
110  * @return
111  *   a pointer to the allocate space.
112  */
113 static void *
114 mlx5_alloc_verbs_buf(size_t size, void *data)
115 {
116 	struct priv *priv = data;
117 	void *ret;
118 	size_t alignment = sysconf(_SC_PAGESIZE);
119 	unsigned int socket = SOCKET_ID_ANY;
120 
121 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
122 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
123 
124 		socket = ctrl->socket;
125 	} else if (priv->verbs_alloc_ctx.type ==
126 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
127 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
128 
129 		socket = ctrl->socket;
130 	}
131 	assert(data != NULL);
132 	ret = rte_malloc_socket(__func__, size, alignment, socket);
133 	DEBUG("Extern alloc size: %lu, align: %lu: %p", size, alignment, ret);
134 	return ret;
135 }
136 
137 /**
138  * Verbs callback to free a memory.
139  *
140  * @param[in] ptr
141  *   A pointer to the memory to free.
142  * @param[in] data
143  *   A pointer to the callback data.
144  */
145 static void
146 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
147 {
148 	assert(data != NULL);
149 	DEBUG("Extern free request: %p", ptr);
150 	rte_free(ptr);
151 }
152 
153 /**
154  * DPDK callback to close the device.
155  *
156  * Destroy all queues and objects, free memory.
157  *
158  * @param dev
159  *   Pointer to Ethernet device structure.
160  */
161 static void
162 mlx5_dev_close(struct rte_eth_dev *dev)
163 {
164 	struct priv *priv = dev->data->dev_private;
165 	unsigned int i;
166 	int ret;
167 
168 	priv_lock(priv);
169 	DEBUG("%p: closing device \"%s\"",
170 	      (void *)dev,
171 	      ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
172 	/* In case mlx5_dev_stop() has not been called. */
173 	priv_dev_interrupt_handler_uninstall(priv, dev);
174 	priv_dev_traffic_disable(priv, dev);
175 	/* Prevent crashes when queues are still in use. */
176 	dev->rx_pkt_burst = removed_rx_burst;
177 	dev->tx_pkt_burst = removed_tx_burst;
178 	if (priv->rxqs != NULL) {
179 		/* XXX race condition if mlx5_rx_burst() is still running. */
180 		usleep(1000);
181 		for (i = 0; (i != priv->rxqs_n); ++i)
182 			mlx5_priv_rxq_release(priv, i);
183 		priv->rxqs_n = 0;
184 		priv->rxqs = NULL;
185 	}
186 	if (priv->txqs != NULL) {
187 		/* XXX race condition if mlx5_tx_burst() is still running. */
188 		usleep(1000);
189 		for (i = 0; (i != priv->txqs_n); ++i)
190 			mlx5_priv_txq_release(priv, i);
191 		priv->txqs_n = 0;
192 		priv->txqs = NULL;
193 	}
194 	if (priv->pd != NULL) {
195 		assert(priv->ctx != NULL);
196 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
197 		claim_zero(mlx5_glue->close_device(priv->ctx));
198 	} else
199 		assert(priv->ctx == NULL);
200 	if (priv->rss_conf.rss_key != NULL)
201 		rte_free(priv->rss_conf.rss_key);
202 	if (priv->reta_idx != NULL)
203 		rte_free(priv->reta_idx);
204 	priv_socket_uninit(priv);
205 	ret = mlx5_priv_hrxq_ibv_verify(priv);
206 	if (ret)
207 		WARN("%p: some Hash Rx queue still remain", (void *)priv);
208 	ret = mlx5_priv_ind_table_ibv_verify(priv);
209 	if (ret)
210 		WARN("%p: some Indirection table still remain", (void *)priv);
211 	ret = mlx5_priv_rxq_ibv_verify(priv);
212 	if (ret)
213 		WARN("%p: some Verbs Rx queue still remain", (void *)priv);
214 	ret = mlx5_priv_rxq_verify(priv);
215 	if (ret)
216 		WARN("%p: some Rx Queues still remain", (void *)priv);
217 	ret = mlx5_priv_txq_ibv_verify(priv);
218 	if (ret)
219 		WARN("%p: some Verbs Tx queue still remain", (void *)priv);
220 	ret = mlx5_priv_txq_verify(priv);
221 	if (ret)
222 		WARN("%p: some Tx Queues still remain", (void *)priv);
223 	ret = priv_flow_verify(priv);
224 	if (ret)
225 		WARN("%p: some flows still remain", (void *)priv);
226 	ret = priv_mr_verify(priv);
227 	if (ret)
228 		WARN("%p: some Memory Region still remain", (void *)priv);
229 	priv_unlock(priv);
230 	memset(priv, 0, sizeof(*priv));
231 }
232 
233 const struct eth_dev_ops mlx5_dev_ops = {
234 	.dev_configure = mlx5_dev_configure,
235 	.dev_start = mlx5_dev_start,
236 	.dev_stop = mlx5_dev_stop,
237 	.dev_set_link_down = mlx5_set_link_down,
238 	.dev_set_link_up = mlx5_set_link_up,
239 	.dev_close = mlx5_dev_close,
240 	.promiscuous_enable = mlx5_promiscuous_enable,
241 	.promiscuous_disable = mlx5_promiscuous_disable,
242 	.allmulticast_enable = mlx5_allmulticast_enable,
243 	.allmulticast_disable = mlx5_allmulticast_disable,
244 	.link_update = mlx5_link_update,
245 	.stats_get = mlx5_stats_get,
246 	.stats_reset = mlx5_stats_reset,
247 	.xstats_get = mlx5_xstats_get,
248 	.xstats_reset = mlx5_xstats_reset,
249 	.xstats_get_names = mlx5_xstats_get_names,
250 	.dev_infos_get = mlx5_dev_infos_get,
251 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
252 	.vlan_filter_set = mlx5_vlan_filter_set,
253 	.rx_queue_setup = mlx5_rx_queue_setup,
254 	.tx_queue_setup = mlx5_tx_queue_setup,
255 	.rx_queue_release = mlx5_rx_queue_release,
256 	.tx_queue_release = mlx5_tx_queue_release,
257 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
258 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
259 	.mac_addr_remove = mlx5_mac_addr_remove,
260 	.mac_addr_add = mlx5_mac_addr_add,
261 	.mac_addr_set = mlx5_mac_addr_set,
262 	.mtu_set = mlx5_dev_set_mtu,
263 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
264 	.vlan_offload_set = mlx5_vlan_offload_set,
265 	.reta_update = mlx5_dev_rss_reta_update,
266 	.reta_query = mlx5_dev_rss_reta_query,
267 	.rss_hash_update = mlx5_rss_hash_update,
268 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
269 	.filter_ctrl = mlx5_dev_filter_ctrl,
270 	.rx_descriptor_status = mlx5_rx_descriptor_status,
271 	.tx_descriptor_status = mlx5_tx_descriptor_status,
272 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
273 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
274 	.is_removed = mlx5_is_removed,
275 };
276 
277 static const struct eth_dev_ops mlx5_dev_sec_ops = {
278 	.stats_get = mlx5_stats_get,
279 	.stats_reset = mlx5_stats_reset,
280 	.xstats_get = mlx5_xstats_get,
281 	.xstats_reset = mlx5_xstats_reset,
282 	.xstats_get_names = mlx5_xstats_get_names,
283 	.dev_infos_get = mlx5_dev_infos_get,
284 	.rx_descriptor_status = mlx5_rx_descriptor_status,
285 	.tx_descriptor_status = mlx5_tx_descriptor_status,
286 };
287 
288 /* Available operators in flow isolated mode. */
289 const struct eth_dev_ops mlx5_dev_ops_isolate = {
290 	.dev_configure = mlx5_dev_configure,
291 	.dev_start = mlx5_dev_start,
292 	.dev_stop = mlx5_dev_stop,
293 	.dev_set_link_down = mlx5_set_link_down,
294 	.dev_set_link_up = mlx5_set_link_up,
295 	.dev_close = mlx5_dev_close,
296 	.link_update = mlx5_link_update,
297 	.stats_get = mlx5_stats_get,
298 	.stats_reset = mlx5_stats_reset,
299 	.xstats_get = mlx5_xstats_get,
300 	.xstats_reset = mlx5_xstats_reset,
301 	.xstats_get_names = mlx5_xstats_get_names,
302 	.dev_infos_get = mlx5_dev_infos_get,
303 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
304 	.vlan_filter_set = mlx5_vlan_filter_set,
305 	.rx_queue_setup = mlx5_rx_queue_setup,
306 	.tx_queue_setup = mlx5_tx_queue_setup,
307 	.rx_queue_release = mlx5_rx_queue_release,
308 	.tx_queue_release = mlx5_tx_queue_release,
309 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
310 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
311 	.mac_addr_remove = mlx5_mac_addr_remove,
312 	.mac_addr_add = mlx5_mac_addr_add,
313 	.mac_addr_set = mlx5_mac_addr_set,
314 	.mtu_set = mlx5_dev_set_mtu,
315 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
316 	.vlan_offload_set = mlx5_vlan_offload_set,
317 	.filter_ctrl = mlx5_dev_filter_ctrl,
318 	.rx_descriptor_status = mlx5_rx_descriptor_status,
319 	.tx_descriptor_status = mlx5_tx_descriptor_status,
320 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
321 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
322 	.is_removed = mlx5_is_removed,
323 };
324 
325 static struct {
326 	struct rte_pci_addr pci_addr; /* associated PCI address */
327 	uint32_t ports; /* physical ports bitfield. */
328 } mlx5_dev[32];
329 
330 /**
331  * Get device index in mlx5_dev[] from PCI bus address.
332  *
333  * @param[in] pci_addr
334  *   PCI bus address to look for.
335  *
336  * @return
337  *   mlx5_dev[] index on success, -1 on failure.
338  */
339 static int
340 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
341 {
342 	unsigned int i;
343 	int ret = -1;
344 
345 	assert(pci_addr != NULL);
346 	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
347 		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
348 		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
349 		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
350 		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
351 			return i;
352 		if ((mlx5_dev[i].ports == 0) && (ret == -1))
353 			ret = i;
354 	}
355 	return ret;
356 }
357 
358 /**
359  * Verify and store value for device argument.
360  *
361  * @param[in] key
362  *   Key argument to verify.
363  * @param[in] val
364  *   Value associated with key.
365  * @param opaque
366  *   User data.
367  *
368  * @return
369  *   0 on success, negative errno value on failure.
370  */
371 static int
372 mlx5_args_check(const char *key, const char *val, void *opaque)
373 {
374 	struct mlx5_dev_config *config = opaque;
375 	unsigned long tmp;
376 
377 	errno = 0;
378 	tmp = strtoul(val, NULL, 0);
379 	if (errno) {
380 		WARN("%s: \"%s\" is not a valid integer", key, val);
381 		return errno;
382 	}
383 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
384 		config->cqe_comp = !!tmp;
385 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
386 		config->txq_inline = tmp;
387 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
388 		config->txqs_inline = tmp;
389 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
390 		config->mps = !!tmp ? config->mps : 0;
391 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
392 		config->mpw_hdr_dseg = !!tmp;
393 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
394 		config->inline_max_packet_sz = tmp;
395 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
396 		config->tx_vec_en = !!tmp;
397 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
398 		config->rx_vec_en = !!tmp;
399 	} else {
400 		WARN("%s: unknown parameter", key);
401 		return -EINVAL;
402 	}
403 	return 0;
404 }
405 
406 /**
407  * Parse device parameters.
408  *
409  * @param config
410  *   Pointer to device configuration structure.
411  * @param devargs
412  *   Device arguments structure.
413  *
414  * @return
415  *   0 on success, errno value on failure.
416  */
417 static int
418 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
419 {
420 	const char **params = (const char *[]){
421 		MLX5_RXQ_CQE_COMP_EN,
422 		MLX5_TXQ_INLINE,
423 		MLX5_TXQS_MIN_INLINE,
424 		MLX5_TXQ_MPW_EN,
425 		MLX5_TXQ_MPW_HDR_DSEG_EN,
426 		MLX5_TXQ_MAX_INLINE_LEN,
427 		MLX5_TX_VEC_EN,
428 		MLX5_RX_VEC_EN,
429 		NULL,
430 	};
431 	struct rte_kvargs *kvlist;
432 	int ret = 0;
433 	int i;
434 
435 	if (devargs == NULL)
436 		return 0;
437 	/* Following UGLY cast is done to pass checkpatch. */
438 	kvlist = rte_kvargs_parse(devargs->args, params);
439 	if (kvlist == NULL)
440 		return 0;
441 	/* Process parameters. */
442 	for (i = 0; (params[i] != NULL); ++i) {
443 		if (rte_kvargs_count(kvlist, params[i])) {
444 			ret = rte_kvargs_process(kvlist, params[i],
445 						 mlx5_args_check, config);
446 			if (ret != 0) {
447 				rte_kvargs_free(kvlist);
448 				return ret;
449 			}
450 		}
451 	}
452 	rte_kvargs_free(kvlist);
453 	return 0;
454 }
455 
456 static struct rte_pci_driver mlx5_driver;
457 
458 /*
459  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
460  * local resource used by both primary and secondary to avoid duplicate
461  * reservation.
462  * The space has to be available on both primary and secondary process,
463  * TXQ UAR maps to this area using fixed mmap w/o double check.
464  */
465 static void *uar_base;
466 
467 /**
468  * Reserve UAR address space for primary process.
469  *
470  * @param[in] priv
471  *   Pointer to private structure.
472  *
473  * @return
474  *   0 on success, errno value on failure.
475  */
476 static int
477 priv_uar_init_primary(struct priv *priv)
478 {
479 	void *addr = (void *)0;
480 	int i;
481 	const struct rte_mem_config *mcfg;
482 	int ret;
483 
484 	if (uar_base) { /* UAR address space mapped. */
485 		priv->uar_base = uar_base;
486 		return 0;
487 	}
488 	/* find out lower bound of hugepage segments */
489 	mcfg = rte_eal_get_configuration()->mem_config;
490 	for (i = 0; i < RTE_MAX_MEMSEG && mcfg->memseg[i].addr; i++) {
491 		if (addr)
492 			addr = RTE_MIN(addr, mcfg->memseg[i].addr);
493 		else
494 			addr = mcfg->memseg[i].addr;
495 	}
496 	/* keep distance to hugepages to minimize potential conflicts. */
497 	addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
498 	/* anonymous mmap, no real memory consumption. */
499 	addr = mmap(addr, MLX5_UAR_SIZE,
500 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
501 	if (addr == MAP_FAILED) {
502 		ERROR("Failed to reserve UAR address space, please adjust "
503 		      "MLX5_UAR_SIZE or try --base-virtaddr");
504 		ret = ENOMEM;
505 		return ret;
506 	}
507 	/* Accept either same addr or a new addr returned from mmap if target
508 	 * range occupied.
509 	 */
510 	INFO("Reserved UAR address space: %p", addr);
511 	priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
512 	uar_base = addr; /* process local, don't reserve again. */
513 	return 0;
514 }
515 
516 /**
517  * Reserve UAR address space for secondary process, align with
518  * primary process.
519  *
520  * @param[in] priv
521  *   Pointer to private structure.
522  *
523  * @return
524  *   0 on success, errno value on failure.
525  */
526 static int
527 priv_uar_init_secondary(struct priv *priv)
528 {
529 	void *addr;
530 	int ret;
531 
532 	assert(priv->uar_base);
533 	if (uar_base) { /* already reserved. */
534 		assert(uar_base == priv->uar_base);
535 		return 0;
536 	}
537 	/* anonymous mmap, no real memory consumption. */
538 	addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
539 		    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
540 	if (addr == MAP_FAILED) {
541 		ERROR("UAR mmap failed: %p size: %llu",
542 		      priv->uar_base, MLX5_UAR_SIZE);
543 		ret = ENXIO;
544 		return ret;
545 	}
546 	if (priv->uar_base != addr) {
547 		ERROR("UAR address %p size %llu occupied, please adjust "
548 		      "MLX5_UAR_OFFSET or try EAL parameter --base-virtaddr",
549 		      priv->uar_base, MLX5_UAR_SIZE);
550 		ret = ENXIO;
551 		return ret;
552 	}
553 	uar_base = addr; /* process local, don't reserve again */
554 	INFO("Reserved UAR address space: %p", addr);
555 	return 0;
556 }
557 
558 /**
559  * DPDK callback to register a PCI device.
560  *
561  * This function creates an Ethernet device for each port of a given
562  * PCI device.
563  *
564  * @param[in] pci_drv
565  *   PCI driver structure (mlx5_driver).
566  * @param[in] pci_dev
567  *   PCI device information.
568  *
569  * @return
570  *   0 on success, negative errno value on failure.
571  */
572 static int
573 mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
574 {
575 	struct ibv_device **list;
576 	struct ibv_device *ibv_dev;
577 	int err = 0;
578 	struct ibv_context *attr_ctx = NULL;
579 	struct ibv_device_attr_ex device_attr;
580 	unsigned int sriov;
581 	unsigned int mps;
582 	unsigned int cqe_comp;
583 	unsigned int tunnel_en = 0;
584 	int idx;
585 	int i;
586 	struct mlx5dv_context attrs_out;
587 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
588 	struct ibv_counter_set_description cs_desc;
589 #endif
590 
591 	(void)pci_drv;
592 	assert(pci_drv == &mlx5_driver);
593 	/* Get mlx5_dev[] index. */
594 	idx = mlx5_dev_idx(&pci_dev->addr);
595 	if (idx == -1) {
596 		ERROR("this driver cannot support any more adapters");
597 		return -ENOMEM;
598 	}
599 	DEBUG("using driver device index %d", idx);
600 
601 	/* Save PCI address. */
602 	mlx5_dev[idx].pci_addr = pci_dev->addr;
603 	list = mlx5_glue->get_device_list(&i);
604 	if (list == NULL) {
605 		assert(errno);
606 		if (errno == ENOSYS)
607 			ERROR("cannot list devices, is ib_uverbs loaded?");
608 		return -errno;
609 	}
610 	assert(i >= 0);
611 	/*
612 	 * For each listed device, check related sysfs entry against
613 	 * the provided PCI ID.
614 	 */
615 	while (i != 0) {
616 		struct rte_pci_addr pci_addr;
617 
618 		--i;
619 		DEBUG("checking device \"%s\"", list[i]->name);
620 		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
621 			continue;
622 		if ((pci_dev->addr.domain != pci_addr.domain) ||
623 		    (pci_dev->addr.bus != pci_addr.bus) ||
624 		    (pci_dev->addr.devid != pci_addr.devid) ||
625 		    (pci_dev->addr.function != pci_addr.function))
626 			continue;
627 		sriov = ((pci_dev->id.device_id ==
628 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
629 		      (pci_dev->id.device_id ==
630 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
631 		      (pci_dev->id.device_id ==
632 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
633 		      (pci_dev->id.device_id ==
634 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
635 		switch (pci_dev->id.device_id) {
636 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
637 			tunnel_en = 1;
638 			break;
639 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
640 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
641 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
642 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
643 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
644 			tunnel_en = 1;
645 			break;
646 		default:
647 			break;
648 		}
649 		INFO("PCI information matches, using device \"%s\""
650 		     " (SR-IOV: %s)",
651 		     list[i]->name,
652 		     sriov ? "true" : "false");
653 		attr_ctx = mlx5_glue->open_device(list[i]);
654 		err = errno;
655 		break;
656 	}
657 	if (attr_ctx == NULL) {
658 		mlx5_glue->free_device_list(list);
659 		switch (err) {
660 		case 0:
661 			ERROR("cannot access device, is mlx5_ib loaded?");
662 			return -ENODEV;
663 		case EINVAL:
664 			ERROR("cannot use device, are drivers up to date?");
665 			return -EINVAL;
666 		}
667 		assert(err > 0);
668 		return -err;
669 	}
670 	ibv_dev = list[i];
671 
672 	DEBUG("device opened");
673 	/*
674 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
675 	 * as all ConnectX-5 devices.
676 	 */
677 	mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
678 	if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
679 		if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
680 			DEBUG("Enhanced MPW is supported");
681 			mps = MLX5_MPW_ENHANCED;
682 		} else {
683 			DEBUG("MPW is supported");
684 			mps = MLX5_MPW;
685 		}
686 	} else {
687 		DEBUG("MPW isn't supported");
688 		mps = MLX5_MPW_DISABLED;
689 	}
690 	if (RTE_CACHE_LINE_SIZE == 128 &&
691 	    !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
692 		cqe_comp = 0;
693 	else
694 		cqe_comp = 1;
695 	if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr))
696 		goto error;
697 	INFO("%u port(s) detected", device_attr.orig_attr.phys_port_cnt);
698 
699 	for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
700 		char name[RTE_ETH_NAME_MAX_LEN];
701 		int len;
702 		uint32_t port = i + 1; /* ports are indexed from one */
703 		uint32_t test = (1 << i);
704 		struct ibv_context *ctx = NULL;
705 		struct ibv_port_attr port_attr;
706 		struct ibv_pd *pd = NULL;
707 		struct priv *priv = NULL;
708 		struct rte_eth_dev *eth_dev;
709 		struct ibv_device_attr_ex device_attr_ex;
710 		struct ether_addr mac;
711 		uint16_t num_vfs = 0;
712 		struct ibv_device_attr_ex device_attr;
713 		struct mlx5_dev_config config = {
714 			.cqe_comp = cqe_comp,
715 			.mps = mps,
716 			.tunnel_en = tunnel_en,
717 			.tx_vec_en = 1,
718 			.rx_vec_en = 1,
719 			.mpw_hdr_dseg = 0,
720 			.txq_inline = MLX5_ARG_UNSET,
721 			.txqs_inline = MLX5_ARG_UNSET,
722 			.inline_max_packet_sz = MLX5_ARG_UNSET,
723 		};
724 
725 		len = snprintf(name, sizeof(name), PCI_PRI_FMT,
726 			 pci_dev->addr.domain, pci_dev->addr.bus,
727 			 pci_dev->addr.devid, pci_dev->addr.function);
728 		if (device_attr.orig_attr.phys_port_cnt > 1)
729 			snprintf(name + len, sizeof(name), " port %u", i);
730 
731 		mlx5_dev[idx].ports |= test;
732 
733 		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
734 			eth_dev = rte_eth_dev_attach_secondary(name);
735 			if (eth_dev == NULL) {
736 				ERROR("can not attach rte ethdev");
737 				err = ENOMEM;
738 				goto error;
739 			}
740 			eth_dev->device = &pci_dev->device;
741 			eth_dev->dev_ops = &mlx5_dev_sec_ops;
742 			priv = eth_dev->data->dev_private;
743 			err = priv_uar_init_secondary(priv);
744 			if (err < 0) {
745 				err = -err;
746 				goto error;
747 			}
748 			/* Receive command fd from primary process */
749 			err = priv_socket_connect(priv);
750 			if (err < 0) {
751 				err = -err;
752 				goto error;
753 			}
754 			/* Remap UAR for Tx queues. */
755 			err = priv_tx_uar_remap(priv, err);
756 			if (err)
757 				goto error;
758 			/*
759 			 * Ethdev pointer is still required as input since
760 			 * the primary device is not accessible from the
761 			 * secondary process.
762 			 */
763 			eth_dev->rx_pkt_burst =
764 				priv_select_rx_function(priv, eth_dev);
765 			eth_dev->tx_pkt_burst =
766 				priv_select_tx_function(priv, eth_dev);
767 			continue;
768 		}
769 
770 		DEBUG("using port %u (%08" PRIx32 ")", port, test);
771 
772 		ctx = mlx5_glue->open_device(ibv_dev);
773 		if (ctx == NULL) {
774 			err = ENODEV;
775 			goto port_error;
776 		}
777 
778 		mlx5_glue->query_device_ex(ctx, NULL, &device_attr);
779 		/* Check port status. */
780 		err = mlx5_glue->query_port(ctx, port, &port_attr);
781 		if (err) {
782 			ERROR("port query failed: %s", strerror(err));
783 			goto port_error;
784 		}
785 
786 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
787 			ERROR("port %d is not configured in Ethernet mode",
788 			      port);
789 			err = EINVAL;
790 			goto port_error;
791 		}
792 
793 		if (port_attr.state != IBV_PORT_ACTIVE)
794 			DEBUG("port %d is not active: \"%s\" (%d)",
795 			      port, mlx5_glue->port_state_str(port_attr.state),
796 			      port_attr.state);
797 
798 		/* Allocate protection domain. */
799 		pd = mlx5_glue->alloc_pd(ctx);
800 		if (pd == NULL) {
801 			ERROR("PD allocation failure");
802 			err = ENOMEM;
803 			goto port_error;
804 		}
805 
806 		mlx5_dev[idx].ports |= test;
807 
808 		/* from rte_ethdev.c */
809 		priv = rte_zmalloc("ethdev private structure",
810 				   sizeof(*priv),
811 				   RTE_CACHE_LINE_SIZE);
812 		if (priv == NULL) {
813 			ERROR("priv allocation failure");
814 			err = ENOMEM;
815 			goto port_error;
816 		}
817 
818 		priv->ctx = ctx;
819 		strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
820 			sizeof(priv->ibdev_path));
821 		priv->device_attr = device_attr;
822 		priv->port = port;
823 		priv->pd = pd;
824 		priv->mtu = ETHER_MTU;
825 		err = mlx5_args(&config, pci_dev->device.devargs);
826 		if (err) {
827 			ERROR("failed to process device arguments: %s",
828 			      strerror(err));
829 			goto port_error;
830 		}
831 		if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
832 			ERROR("ibv_query_device_ex() failed");
833 			goto port_error;
834 		}
835 
836 		config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
837 				    IBV_DEVICE_RAW_IP_CSUM);
838 		DEBUG("checksum offloading is %ssupported",
839 		      (config.hw_csum ? "" : "not "));
840 
841 #ifdef HAVE_IBV_DEVICE_VXLAN_SUPPORT
842 		config.hw_csum_l2tun =
843 				!!(exp_device_attr.exp_device_cap_flags &
844 				   IBV_DEVICE_VXLAN_SUPPORT);
845 #endif
846 		DEBUG("Rx L2 tunnel checksum offloads are %ssupported",
847 		      (config.hw_csum_l2tun ? "" : "not "));
848 
849 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
850 		config.flow_counter_en = !!(device_attr.max_counter_sets);
851 		mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
852 		DEBUG("counter type = %d, num of cs = %ld, attributes = %d",
853 		      cs_desc.counter_type, cs_desc.num_of_cs,
854 		      cs_desc.attributes);
855 #endif
856 		config.ind_table_max_size =
857 			device_attr_ex.rss_caps.max_rwq_indirection_table_size;
858 		/* Remove this check once DPDK supports larger/variable
859 		 * indirection tables. */
860 		if (config.ind_table_max_size >
861 				(unsigned int)ETH_RSS_RETA_SIZE_512)
862 			config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
863 		DEBUG("maximum RX indirection table size is %u",
864 		      config.ind_table_max_size);
865 		config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
866 					 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
867 		DEBUG("VLAN stripping is %ssupported",
868 		      (config.hw_vlan_strip ? "" : "not "));
869 
870 		config.hw_fcs_strip =
871 				!!(device_attr_ex.orig_attr.device_cap_flags &
872 				IBV_WQ_FLAGS_SCATTER_FCS);
873 		DEBUG("FCS stripping configuration is %ssupported",
874 		      (config.hw_fcs_strip ? "" : "not "));
875 
876 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
877 		config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
878 #endif
879 		DEBUG("hardware RX end alignment padding is %ssupported",
880 		      (config.hw_padding ? "" : "not "));
881 
882 		priv_get_num_vfs(priv, &num_vfs);
883 		config.sriov = (num_vfs || sriov);
884 		config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
885 			      (device_attr_ex.tso_caps.supported_qpts &
886 			      (1 << IBV_QPT_RAW_PACKET)));
887 		if (config.tso)
888 			config.tso_max_payload_sz =
889 					device_attr_ex.tso_caps.max_tso;
890 		if (config.mps && !mps) {
891 			ERROR("multi-packet send not supported on this device"
892 			      " (" MLX5_TXQ_MPW_EN ")");
893 			err = ENOTSUP;
894 			goto port_error;
895 		}
896 		INFO("%sMPS is %s",
897 		     config.mps == MLX5_MPW_ENHANCED ? "Enhanced " : "",
898 		     config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
899 		if (config.cqe_comp && !cqe_comp) {
900 			WARN("Rx CQE compression isn't supported");
901 			config.cqe_comp = 0;
902 		}
903 		err = priv_uar_init_primary(priv);
904 		if (err)
905 			goto port_error;
906 		/* Configure the first MAC address by default. */
907 		if (priv_get_mac(priv, &mac.addr_bytes)) {
908 			ERROR("cannot get MAC address, is mlx5_en loaded?"
909 			      " (errno: %s)", strerror(errno));
910 			err = ENODEV;
911 			goto port_error;
912 		}
913 		INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
914 		     priv->port,
915 		     mac.addr_bytes[0], mac.addr_bytes[1],
916 		     mac.addr_bytes[2], mac.addr_bytes[3],
917 		     mac.addr_bytes[4], mac.addr_bytes[5]);
918 #ifndef NDEBUG
919 		{
920 			char ifname[IF_NAMESIZE];
921 
922 			if (priv_get_ifname(priv, &ifname) == 0)
923 				DEBUG("port %u ifname is \"%s\"",
924 				      priv->port, ifname);
925 			else
926 				DEBUG("port %u ifname is unknown", priv->port);
927 		}
928 #endif
929 		/* Get actual MTU if possible. */
930 		priv_get_mtu(priv, &priv->mtu);
931 		DEBUG("port %u MTU is %u", priv->port, priv->mtu);
932 
933 		eth_dev = rte_eth_dev_allocate(name);
934 		if (eth_dev == NULL) {
935 			ERROR("can not allocate rte ethdev");
936 			err = ENOMEM;
937 			goto port_error;
938 		}
939 		eth_dev->data->dev_private = priv;
940 		eth_dev->data->mac_addrs = priv->mac;
941 		eth_dev->device = &pci_dev->device;
942 		rte_eth_copy_pci_info(eth_dev, pci_dev);
943 		eth_dev->device->driver = &mlx5_driver.driver;
944 		/*
945 		 * Initialize burst functions to prevent crashes before link-up.
946 		 */
947 		eth_dev->rx_pkt_burst = removed_rx_burst;
948 		eth_dev->tx_pkt_burst = removed_tx_burst;
949 		priv->dev = eth_dev;
950 		eth_dev->dev_ops = &mlx5_dev_ops;
951 		/* Register MAC address. */
952 		claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
953 		TAILQ_INIT(&priv->flows);
954 		TAILQ_INIT(&priv->ctrl_flows);
955 
956 		/* Hint libmlx5 to use PMD allocator for data plane resources */
957 		struct mlx5dv_ctx_allocators alctr = {
958 			.alloc = &mlx5_alloc_verbs_buf,
959 			.free = &mlx5_free_verbs_buf,
960 			.data = priv,
961 		};
962 		mlx5_glue->dv_set_context_attr(ctx,
963 					       MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
964 					       (void *)((uintptr_t)&alctr));
965 
966 		/* Bring Ethernet device up. */
967 		DEBUG("forcing Ethernet interface up");
968 		priv_set_flags(priv, ~IFF_UP, IFF_UP);
969 		/* Store device configuration on private structure. */
970 		priv->config = config;
971 		continue;
972 
973 port_error:
974 		if (priv)
975 			rte_free(priv);
976 		if (pd)
977 			claim_zero(mlx5_glue->dealloc_pd(pd));
978 		if (ctx)
979 			claim_zero(mlx5_glue->close_device(ctx));
980 		break;
981 	}
982 
983 	/*
984 	 * XXX if something went wrong in the loop above, there is a resource
985 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
986 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
987 	 * way to enumerate the registered ethdevs to free the previous ones.
988 	 */
989 
990 	/* no port found, complain */
991 	if (!mlx5_dev[idx].ports) {
992 		err = ENODEV;
993 		goto error;
994 	}
995 
996 error:
997 	if (attr_ctx)
998 		claim_zero(mlx5_glue->close_device(attr_ctx));
999 	if (list)
1000 		mlx5_glue->free_device_list(list);
1001 	assert(err >= 0);
1002 	return -err;
1003 }
1004 
1005 static const struct rte_pci_id mlx5_pci_id_map[] = {
1006 	{
1007 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1008 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1009 	},
1010 	{
1011 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1012 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1013 	},
1014 	{
1015 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1016 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1017 	},
1018 	{
1019 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1020 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1021 	},
1022 	{
1023 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1024 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1025 	},
1026 	{
1027 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1028 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1029 	},
1030 	{
1031 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1032 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1033 	},
1034 	{
1035 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1036 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1037 	},
1038 	{
1039 		.vendor_id = 0
1040 	}
1041 };
1042 
1043 static struct rte_pci_driver mlx5_driver = {
1044 	.driver = {
1045 		.name = MLX5_DRIVER_NAME
1046 	},
1047 	.id_table = mlx5_pci_id_map,
1048 	.probe = mlx5_pci_probe,
1049 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1050 };
1051 
1052 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1053 
1054 /**
1055  * Initialization routine for run-time dependency on rdma-core.
1056  */
1057 static int
1058 mlx5_glue_init(void)
1059 {
1060 	void *handle = NULL;
1061 	void **sym;
1062 	const char *dlmsg;
1063 
1064 	handle = dlopen(MLX5_GLUE, RTLD_LAZY);
1065 	if (!handle) {
1066 		rte_errno = EINVAL;
1067 		dlmsg = dlerror();
1068 		if (dlmsg)
1069 			WARN("cannot load glue library: %s", dlmsg);
1070 		goto glue_error;
1071 	}
1072 	sym = dlsym(handle, "mlx5_glue");
1073 	if (!sym || !*sym) {
1074 		rte_errno = EINVAL;
1075 		dlmsg = dlerror();
1076 		if (dlmsg)
1077 			ERROR("cannot resolve glue symbol: %s", dlmsg);
1078 		goto glue_error;
1079 	}
1080 	mlx5_glue = *sym;
1081 	return 0;
1082 glue_error:
1083 	if (handle)
1084 		dlclose(handle);
1085 	WARN("cannot initialize PMD due to missing run-time"
1086 	     " dependency on rdma-core libraries (libibverbs,"
1087 	     " libmlx5)");
1088 	return -rte_errno;
1089 }
1090 
1091 #endif
1092 
1093 /**
1094  * Driver initialization routine.
1095  */
1096 RTE_INIT(rte_mlx5_pmd_init);
1097 static void
1098 rte_mlx5_pmd_init(void)
1099 {
1100 	/* Build the static table for ptype conversion. */
1101 	mlx5_set_ptype_table();
1102 	/*
1103 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1104 	 * huge pages. Calling ibv_fork_init() during init allows
1105 	 * applications to use fork() safely for purposes other than
1106 	 * using this PMD, which is not supported in forked processes.
1107 	 */
1108 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1109 	/* Match the size of Rx completion entry to the size of a cacheline. */
1110 	if (RTE_CACHE_LINE_SIZE == 128)
1111 		setenv("MLX5_CQE_SIZE", "128", 0);
1112 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1113 	if (mlx5_glue_init())
1114 		return;
1115 	assert(mlx5_glue);
1116 #endif
1117 	mlx5_glue->fork_init();
1118 	rte_pci_register(&mlx5_driver);
1119 }
1120 
1121 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1122 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1123 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1124