xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 4e30ead5e7ca886535e2b30632b2948d2aac1681)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2015 6WIND S.A.
5  *   Copyright 2015 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of 6WIND S.A. nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <stddef.h>
35 #include <unistd.h>
36 #include <string.h>
37 #include <assert.h>
38 #include <stdint.h>
39 #include <stdlib.h>
40 #include <errno.h>
41 #include <net/if.h>
42 
43 /* Verbs header. */
44 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
45 #ifdef PEDANTIC
46 #pragma GCC diagnostic ignored "-Wpedantic"
47 #endif
48 #include <infiniband/verbs.h>
49 #ifdef PEDANTIC
50 #pragma GCC diagnostic error "-Wpedantic"
51 #endif
52 
53 /* DPDK headers don't like -pedantic. */
54 #ifdef PEDANTIC
55 #pragma GCC diagnostic ignored "-Wpedantic"
56 #endif
57 #include <rte_malloc.h>
58 #include <rte_ethdev.h>
59 #include <rte_ethdev_pci.h>
60 #include <rte_pci.h>
61 #include <rte_common.h>
62 #include <rte_kvargs.h>
63 #ifdef PEDANTIC
64 #pragma GCC diagnostic error "-Wpedantic"
65 #endif
66 
67 #include "mlx5.h"
68 #include "mlx5_utils.h"
69 #include "mlx5_rxtx.h"
70 #include "mlx5_autoconf.h"
71 #include "mlx5_defs.h"
72 
73 /* Device parameter to enable RX completion queue compression. */
74 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
75 
76 /* Device parameter to configure inline send. */
77 #define MLX5_TXQ_INLINE "txq_inline"
78 
79 /*
80  * Device parameter to configure the number of TX queues threshold for
81  * enabling inline send.
82  */
83 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
84 
85 /* Device parameter to enable multi-packet send WQEs. */
86 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
87 
88 /* Device parameter to include 2 dsegs in the title WQEBB. */
89 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
90 
91 /* Device parameter to limit the size of inlining packet. */
92 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
93 
94 /* Device parameter to enable hardware TSO offload. */
95 #define MLX5_TSO "tso"
96 
97 /* Default PMD specific parameter value. */
98 #define MLX5_ARG_UNSET (-1)
99 
100 struct mlx5_args {
101 	int cqe_comp;
102 	int txq_inline;
103 	int txqs_inline;
104 	int mps;
105 	int mpw_hdr_dseg;
106 	int inline_max_packet_sz;
107 	int tso;
108 };
109 /**
110  * Retrieve integer value from environment variable.
111  *
112  * @param[in] name
113  *   Environment variable name.
114  *
115  * @return
116  *   Integer value, 0 if the variable is not set.
117  */
118 int
119 mlx5_getenv_int(const char *name)
120 {
121 	const char *val = getenv(name);
122 
123 	if (val == NULL)
124 		return 0;
125 	return atoi(val);
126 }
127 
128 /**
129  * DPDK callback to close the device.
130  *
131  * Destroy all queues and objects, free memory.
132  *
133  * @param dev
134  *   Pointer to Ethernet device structure.
135  */
136 static void
137 mlx5_dev_close(struct rte_eth_dev *dev)
138 {
139 	struct priv *priv = mlx5_get_priv(dev);
140 	unsigned int i;
141 
142 	priv_lock(priv);
143 	DEBUG("%p: closing device \"%s\"",
144 	      (void *)dev,
145 	      ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
146 	/* In case mlx5_dev_stop() has not been called. */
147 	priv_dev_interrupt_handler_uninstall(priv, dev);
148 	priv_special_flow_disable_all(priv);
149 	priv_mac_addrs_disable(priv);
150 	priv_destroy_hash_rxqs(priv);
151 
152 	/* Remove flow director elements. */
153 	priv_fdir_disable(priv);
154 	priv_fdir_delete_filters_list(priv);
155 
156 	/* Prevent crashes when queues are still in use. */
157 	dev->rx_pkt_burst = removed_rx_burst;
158 	dev->tx_pkt_burst = removed_tx_burst;
159 	if (priv->rxqs != NULL) {
160 		/* XXX race condition if mlx5_rx_burst() is still running. */
161 		usleep(1000);
162 		for (i = 0; (i != priv->rxqs_n); ++i) {
163 			struct rxq *rxq = (*priv->rxqs)[i];
164 			struct rxq_ctrl *rxq_ctrl;
165 
166 			if (rxq == NULL)
167 				continue;
168 			rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
169 			(*priv->rxqs)[i] = NULL;
170 			rxq_cleanup(rxq_ctrl);
171 			rte_free(rxq_ctrl);
172 		}
173 		priv->rxqs_n = 0;
174 		priv->rxqs = NULL;
175 	}
176 	if (priv->txqs != NULL) {
177 		/* XXX race condition if mlx5_tx_burst() is still running. */
178 		usleep(1000);
179 		for (i = 0; (i != priv->txqs_n); ++i) {
180 			struct txq *txq = (*priv->txqs)[i];
181 			struct txq_ctrl *txq_ctrl;
182 
183 			if (txq == NULL)
184 				continue;
185 			txq_ctrl = container_of(txq, struct txq_ctrl, txq);
186 			(*priv->txqs)[i] = NULL;
187 			txq_cleanup(txq_ctrl);
188 			rte_free(txq_ctrl);
189 		}
190 		priv->txqs_n = 0;
191 		priv->txqs = NULL;
192 	}
193 	if (priv->pd != NULL) {
194 		assert(priv->ctx != NULL);
195 		claim_zero(ibv_dealloc_pd(priv->pd));
196 		claim_zero(ibv_close_device(priv->ctx));
197 	} else
198 		assert(priv->ctx == NULL);
199 	if (priv->rss_conf != NULL) {
200 		for (i = 0; (i != hash_rxq_init_n); ++i)
201 			rte_free((*priv->rss_conf)[i]);
202 		rte_free(priv->rss_conf);
203 	}
204 	if (priv->reta_idx != NULL)
205 		rte_free(priv->reta_idx);
206 	priv_unlock(priv);
207 	memset(priv, 0, sizeof(*priv));
208 }
209 
210 static const struct eth_dev_ops mlx5_dev_ops = {
211 	.dev_configure = mlx5_dev_configure,
212 	.dev_start = mlx5_dev_start,
213 	.dev_stop = mlx5_dev_stop,
214 	.dev_set_link_down = mlx5_set_link_down,
215 	.dev_set_link_up = mlx5_set_link_up,
216 	.dev_close = mlx5_dev_close,
217 	.promiscuous_enable = mlx5_promiscuous_enable,
218 	.promiscuous_disable = mlx5_promiscuous_disable,
219 	.allmulticast_enable = mlx5_allmulticast_enable,
220 	.allmulticast_disable = mlx5_allmulticast_disable,
221 	.link_update = mlx5_link_update,
222 	.stats_get = mlx5_stats_get,
223 	.stats_reset = mlx5_stats_reset,
224 	.xstats_get = mlx5_xstats_get,
225 	.xstats_reset = mlx5_xstats_reset,
226 	.xstats_get_names = mlx5_xstats_get_names,
227 	.dev_infos_get = mlx5_dev_infos_get,
228 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
229 	.vlan_filter_set = mlx5_vlan_filter_set,
230 	.rx_queue_setup = mlx5_rx_queue_setup,
231 	.tx_queue_setup = mlx5_tx_queue_setup,
232 	.rx_queue_release = mlx5_rx_queue_release,
233 	.tx_queue_release = mlx5_tx_queue_release,
234 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
235 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
236 	.mac_addr_remove = mlx5_mac_addr_remove,
237 	.mac_addr_add = mlx5_mac_addr_add,
238 	.mac_addr_set = mlx5_mac_addr_set,
239 	.mtu_set = mlx5_dev_set_mtu,
240 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
241 	.vlan_offload_set = mlx5_vlan_offload_set,
242 	.reta_update = mlx5_dev_rss_reta_update,
243 	.reta_query = mlx5_dev_rss_reta_query,
244 	.rss_hash_update = mlx5_rss_hash_update,
245 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
246 	.filter_ctrl = mlx5_dev_filter_ctrl,
247 	.rx_descriptor_status = mlx5_rx_descriptor_status,
248 	.tx_descriptor_status = mlx5_tx_descriptor_status,
249 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
250 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
251 };
252 
253 static struct {
254 	struct rte_pci_addr pci_addr; /* associated PCI address */
255 	uint32_t ports; /* physical ports bitfield. */
256 } mlx5_dev[32];
257 
258 /**
259  * Get device index in mlx5_dev[] from PCI bus address.
260  *
261  * @param[in] pci_addr
262  *   PCI bus address to look for.
263  *
264  * @return
265  *   mlx5_dev[] index on success, -1 on failure.
266  */
267 static int
268 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
269 {
270 	unsigned int i;
271 	int ret = -1;
272 
273 	assert(pci_addr != NULL);
274 	for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
275 		if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
276 		    (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
277 		    (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
278 		    (mlx5_dev[i].pci_addr.function == pci_addr->function))
279 			return i;
280 		if ((mlx5_dev[i].ports == 0) && (ret == -1))
281 			ret = i;
282 	}
283 	return ret;
284 }
285 
286 /**
287  * Verify and store value for device argument.
288  *
289  * @param[in] key
290  *   Key argument to verify.
291  * @param[in] val
292  *   Value associated with key.
293  * @param opaque
294  *   User data.
295  *
296  * @return
297  *   0 on success, negative errno value on failure.
298  */
299 static int
300 mlx5_args_check(const char *key, const char *val, void *opaque)
301 {
302 	struct mlx5_args *args = opaque;
303 	unsigned long tmp;
304 
305 	errno = 0;
306 	tmp = strtoul(val, NULL, 0);
307 	if (errno) {
308 		WARN("%s: \"%s\" is not a valid integer", key, val);
309 		return errno;
310 	}
311 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
312 		args->cqe_comp = !!tmp;
313 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
314 		args->txq_inline = tmp;
315 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
316 		args->txqs_inline = tmp;
317 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
318 		args->mps = !!tmp;
319 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
320 		args->mpw_hdr_dseg = !!tmp;
321 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
322 		args->inline_max_packet_sz = tmp;
323 	} else if (strcmp(MLX5_TSO, key) == 0) {
324 		args->tso = !!tmp;
325 	} else {
326 		WARN("%s: unknown parameter", key);
327 		return -EINVAL;
328 	}
329 	return 0;
330 }
331 
332 /**
333  * Parse device parameters.
334  *
335  * @param priv
336  *   Pointer to private structure.
337  * @param devargs
338  *   Device arguments structure.
339  *
340  * @return
341  *   0 on success, errno value on failure.
342  */
343 static int
344 mlx5_args(struct mlx5_args *args, struct rte_devargs *devargs)
345 {
346 	const char **params = (const char *[]){
347 		MLX5_RXQ_CQE_COMP_EN,
348 		MLX5_TXQ_INLINE,
349 		MLX5_TXQS_MIN_INLINE,
350 		MLX5_TXQ_MPW_EN,
351 		MLX5_TXQ_MPW_HDR_DSEG_EN,
352 		MLX5_TXQ_MAX_INLINE_LEN,
353 		MLX5_TSO,
354 		NULL,
355 	};
356 	struct rte_kvargs *kvlist;
357 	int ret = 0;
358 	int i;
359 
360 	if (devargs == NULL)
361 		return 0;
362 	/* Following UGLY cast is done to pass checkpatch. */
363 	kvlist = rte_kvargs_parse(devargs->args, params);
364 	if (kvlist == NULL)
365 		return 0;
366 	/* Process parameters. */
367 	for (i = 0; (params[i] != NULL); ++i) {
368 		if (rte_kvargs_count(kvlist, params[i])) {
369 			ret = rte_kvargs_process(kvlist, params[i],
370 						 mlx5_args_check, args);
371 			if (ret != 0) {
372 				rte_kvargs_free(kvlist);
373 				return ret;
374 			}
375 		}
376 	}
377 	rte_kvargs_free(kvlist);
378 	return 0;
379 }
380 
381 static struct rte_pci_driver mlx5_driver;
382 
383 /**
384  * Assign parameters from args into priv, only non default
385  * values are considered.
386  *
387  * @param[out] priv
388  *   Pointer to private structure.
389  * @param[in] args
390  *   Pointer to args values.
391  */
392 static void
393 mlx5_args_assign(struct priv *priv, struct mlx5_args *args)
394 {
395 	if (args->cqe_comp != MLX5_ARG_UNSET)
396 		priv->cqe_comp = args->cqe_comp;
397 	if (args->txq_inline != MLX5_ARG_UNSET)
398 		priv->txq_inline = args->txq_inline;
399 	if (args->txqs_inline != MLX5_ARG_UNSET)
400 		priv->txqs_inline = args->txqs_inline;
401 	if (args->mps != MLX5_ARG_UNSET)
402 		priv->mps = args->mps ? priv->mps : 0;
403 	if (args->mpw_hdr_dseg != MLX5_ARG_UNSET)
404 		priv->mpw_hdr_dseg = args->mpw_hdr_dseg;
405 	if (args->inline_max_packet_sz != MLX5_ARG_UNSET)
406 		priv->inline_max_packet_sz = args->inline_max_packet_sz;
407 	if (args->tso != MLX5_ARG_UNSET)
408 		priv->tso = args->tso;
409 }
410 
411 /**
412  * DPDK callback to register a PCI device.
413  *
414  * This function creates an Ethernet device for each port of a given
415  * PCI device.
416  *
417  * @param[in] pci_drv
418  *   PCI driver structure (mlx5_driver).
419  * @param[in] pci_dev
420  *   PCI device information.
421  *
422  * @return
423  *   0 on success, negative errno value on failure.
424  */
425 static int
426 mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
427 {
428 	struct ibv_device **list;
429 	struct ibv_device *ibv_dev;
430 	int err = 0;
431 	struct ibv_context *attr_ctx = NULL;
432 	struct ibv_device_attr device_attr;
433 	unsigned int sriov;
434 	unsigned int mps;
435 	unsigned int tunnel_en;
436 	int idx;
437 	int i;
438 
439 	(void)pci_drv;
440 	assert(pci_drv == &mlx5_driver);
441 	/* Get mlx5_dev[] index. */
442 	idx = mlx5_dev_idx(&pci_dev->addr);
443 	if (idx == -1) {
444 		ERROR("this driver cannot support any more adapters");
445 		return -ENOMEM;
446 	}
447 	DEBUG("using driver device index %d", idx);
448 
449 	/* Save PCI address. */
450 	mlx5_dev[idx].pci_addr = pci_dev->addr;
451 	list = ibv_get_device_list(&i);
452 	if (list == NULL) {
453 		assert(errno);
454 		if (errno == ENOSYS)
455 			ERROR("cannot list devices, is ib_uverbs loaded?");
456 		return -errno;
457 	}
458 	assert(i >= 0);
459 	/*
460 	 * For each listed device, check related sysfs entry against
461 	 * the provided PCI ID.
462 	 */
463 	while (i != 0) {
464 		struct rte_pci_addr pci_addr;
465 
466 		--i;
467 		DEBUG("checking device \"%s\"", list[i]->name);
468 		if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
469 			continue;
470 		if ((pci_dev->addr.domain != pci_addr.domain) ||
471 		    (pci_dev->addr.bus != pci_addr.bus) ||
472 		    (pci_dev->addr.devid != pci_addr.devid) ||
473 		    (pci_dev->addr.function != pci_addr.function))
474 			continue;
475 		sriov = ((pci_dev->id.device_id ==
476 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
477 		      (pci_dev->id.device_id ==
478 		       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
479 		      (pci_dev->id.device_id ==
480 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
481 		      (pci_dev->id.device_id ==
482 		       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
483 		/*
484 		 * Multi-packet send is supported by ConnectX-4 Lx PF as well
485 		 * as all ConnectX-5 devices.
486 		 */
487 		switch (pci_dev->id.device_id) {
488 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
489 			tunnel_en = 1;
490 			mps = MLX5_MPW_DISABLED;
491 			break;
492 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
493 			mps = MLX5_MPW;
494 			break;
495 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
496 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
497 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
498 		case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
499 			tunnel_en = 1;
500 			mps = MLX5_MPW_ENHANCED;
501 			break;
502 		default:
503 			mps = MLX5_MPW_DISABLED;
504 		}
505 		INFO("PCI information matches, using device \"%s\""
506 		     " (SR-IOV: %s, %sMPS: %s)",
507 		     list[i]->name,
508 		     sriov ? "true" : "false",
509 		     mps == MLX5_MPW_ENHANCED ? "Enhanced " : "",
510 		     mps != MLX5_MPW_DISABLED ? "true" : "false");
511 		attr_ctx = ibv_open_device(list[i]);
512 		err = errno;
513 		break;
514 	}
515 	if (attr_ctx == NULL) {
516 		ibv_free_device_list(list);
517 		switch (err) {
518 		case 0:
519 			ERROR("cannot access device, is mlx5_ib loaded?");
520 			return -ENODEV;
521 		case EINVAL:
522 			ERROR("cannot use device, are drivers up to date?");
523 			return -EINVAL;
524 		}
525 		assert(err > 0);
526 		return -err;
527 	}
528 	ibv_dev = list[i];
529 
530 	DEBUG("device opened");
531 	if (ibv_query_device(attr_ctx, &device_attr))
532 		goto error;
533 	INFO("%u port(s) detected", device_attr.phys_port_cnt);
534 
535 	for (i = 0; i < device_attr.phys_port_cnt; i++) {
536 		uint32_t port = i + 1; /* ports are indexed from one */
537 		uint32_t test = (1 << i);
538 		struct ibv_context *ctx = NULL;
539 		struct ibv_port_attr port_attr;
540 		struct ibv_pd *pd = NULL;
541 		struct priv *priv = NULL;
542 		struct rte_eth_dev *eth_dev;
543 		struct ibv_exp_device_attr exp_device_attr;
544 		struct ether_addr mac;
545 		uint16_t num_vfs = 0;
546 		struct mlx5_args args = {
547 			.cqe_comp = MLX5_ARG_UNSET,
548 			.txq_inline = MLX5_ARG_UNSET,
549 			.txqs_inline = MLX5_ARG_UNSET,
550 			.mps = MLX5_ARG_UNSET,
551 			.mpw_hdr_dseg = MLX5_ARG_UNSET,
552 			.inline_max_packet_sz = MLX5_ARG_UNSET,
553 			.tso = MLX5_ARG_UNSET,
554 		};
555 
556 		exp_device_attr.comp_mask =
557 			IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
558 			IBV_EXP_DEVICE_ATTR_RX_HASH |
559 			IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
560 			IBV_EXP_DEVICE_ATTR_RX_PAD_END_ALIGN |
561 			IBV_EXP_DEVICE_ATTR_TSO_CAPS |
562 			0;
563 
564 		DEBUG("using port %u (%08" PRIx32 ")", port, test);
565 
566 		ctx = ibv_open_device(ibv_dev);
567 		if (ctx == NULL)
568 			goto port_error;
569 
570 		/* Check port status. */
571 		err = ibv_query_port(ctx, port, &port_attr);
572 		if (err) {
573 			ERROR("port query failed: %s", strerror(err));
574 			goto port_error;
575 		}
576 
577 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
578 			ERROR("port %d is not configured in Ethernet mode",
579 			      port);
580 			goto port_error;
581 		}
582 
583 		if (port_attr.state != IBV_PORT_ACTIVE)
584 			DEBUG("port %d is not active: \"%s\" (%d)",
585 			      port, ibv_port_state_str(port_attr.state),
586 			      port_attr.state);
587 
588 		/* Allocate protection domain. */
589 		pd = ibv_alloc_pd(ctx);
590 		if (pd == NULL) {
591 			ERROR("PD allocation failure");
592 			err = ENOMEM;
593 			goto port_error;
594 		}
595 
596 		mlx5_dev[idx].ports |= test;
597 
598 		/* from rte_ethdev.c */
599 		priv = rte_zmalloc("ethdev private structure",
600 				   sizeof(*priv),
601 				   RTE_CACHE_LINE_SIZE);
602 		if (priv == NULL) {
603 			ERROR("priv allocation failure");
604 			err = ENOMEM;
605 			goto port_error;
606 		}
607 
608 		priv->ctx = ctx;
609 		priv->device_attr = device_attr;
610 		priv->port = port;
611 		priv->pd = pd;
612 		priv->mtu = ETHER_MTU;
613 		priv->mps = mps; /* Enable MPW by default if supported. */
614 		priv->cqe_comp = 1; /* Enable compression by default. */
615 		priv->tunnel_en = tunnel_en;
616 		err = mlx5_args(&args, pci_dev->device.devargs);
617 		if (err) {
618 			ERROR("failed to process device arguments: %s",
619 			      strerror(err));
620 			goto port_error;
621 		}
622 		mlx5_args_assign(priv, &args);
623 		if (ibv_exp_query_device(ctx, &exp_device_attr)) {
624 			ERROR("ibv_exp_query_device() failed");
625 			goto port_error;
626 		}
627 
628 		priv->hw_csum =
629 			((exp_device_attr.exp_device_cap_flags &
630 			  IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
631 			 (exp_device_attr.exp_device_cap_flags &
632 			  IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
633 		DEBUG("checksum offloading is %ssupported",
634 		      (priv->hw_csum ? "" : "not "));
635 
636 		priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
637 					 IBV_EXP_DEVICE_VXLAN_SUPPORT);
638 		DEBUG("L2 tunnel checksum offloads are %ssupported",
639 		      (priv->hw_csum_l2tun ? "" : "not "));
640 
641 		priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
642 		/* Remove this check once DPDK supports larger/variable
643 		 * indirection tables. */
644 		if (priv->ind_table_max_size >
645 				(unsigned int)ETH_RSS_RETA_SIZE_512)
646 			priv->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
647 		DEBUG("maximum RX indirection table size is %u",
648 		      priv->ind_table_max_size);
649 		priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
650 					 IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
651 		DEBUG("VLAN stripping is %ssupported",
652 		      (priv->hw_vlan_strip ? "" : "not "));
653 
654 		priv->hw_fcs_strip = !!(exp_device_attr.exp_device_cap_flags &
655 					IBV_EXP_DEVICE_SCATTER_FCS);
656 		DEBUG("FCS stripping configuration is %ssupported",
657 		      (priv->hw_fcs_strip ? "" : "not "));
658 
659 		priv->hw_padding = !!exp_device_attr.rx_pad_end_addr_align;
660 		DEBUG("hardware RX end alignment padding is %ssupported",
661 		      (priv->hw_padding ? "" : "not "));
662 
663 		priv_get_num_vfs(priv, &num_vfs);
664 		priv->sriov = (num_vfs || sriov);
665 		priv->tso = ((priv->tso) &&
666 			    (exp_device_attr.tso_caps.max_tso > 0) &&
667 			    (exp_device_attr.tso_caps.supported_qpts &
668 			    (1 << IBV_QPT_RAW_ETH)));
669 		if (priv->tso)
670 			priv->max_tso_payload_sz =
671 				exp_device_attr.tso_caps.max_tso;
672 		if (priv->mps && !mps) {
673 			ERROR("multi-packet send not supported on this device"
674 			      " (" MLX5_TXQ_MPW_EN ")");
675 			err = ENOTSUP;
676 			goto port_error;
677 		} else if (priv->mps && priv->tso) {
678 			WARN("multi-packet send not supported in conjunction "
679 			      "with TSO. MPS disabled");
680 			priv->mps = 0;
681 		}
682 		INFO("%sMPS is %s",
683 		     priv->mps == MLX5_MPW_ENHANCED ? "Enhanced " : "",
684 		     priv->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
685 		/* Set default values for Enhanced MPW, a.k.a MPWv2. */
686 		if (priv->mps == MLX5_MPW_ENHANCED) {
687 			if (args.txqs_inline == MLX5_ARG_UNSET)
688 				priv->txqs_inline = MLX5_EMPW_MIN_TXQS;
689 			if (args.inline_max_packet_sz == MLX5_ARG_UNSET)
690 				priv->inline_max_packet_sz =
691 					MLX5_EMPW_MAX_INLINE_LEN;
692 			if (args.txq_inline == MLX5_ARG_UNSET)
693 				priv->txq_inline = MLX5_WQE_SIZE_MAX -
694 						   MLX5_WQE_SIZE;
695 		}
696 		/* Allocate and register default RSS hash keys. */
697 		priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
698 					    sizeof((*priv->rss_conf)[0]), 0);
699 		if (priv->rss_conf == NULL) {
700 			err = ENOMEM;
701 			goto port_error;
702 		}
703 		err = rss_hash_rss_conf_new_key(priv,
704 						rss_hash_default_key,
705 						rss_hash_default_key_len,
706 						ETH_RSS_PROTO_MASK);
707 		if (err)
708 			goto port_error;
709 		/* Configure the first MAC address by default. */
710 		if (priv_get_mac(priv, &mac.addr_bytes)) {
711 			ERROR("cannot get MAC address, is mlx5_en loaded?"
712 			      " (errno: %s)", strerror(errno));
713 			goto port_error;
714 		}
715 		INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
716 		     priv->port,
717 		     mac.addr_bytes[0], mac.addr_bytes[1],
718 		     mac.addr_bytes[2], mac.addr_bytes[3],
719 		     mac.addr_bytes[4], mac.addr_bytes[5]);
720 		/* Register MAC address. */
721 		claim_zero(priv_mac_addr_add(priv, 0,
722 					     (const uint8_t (*)[ETHER_ADDR_LEN])
723 					     mac.addr_bytes));
724 		/* Initialize FD filters list. */
725 		err = fdir_init_filters_list(priv);
726 		if (err)
727 			goto port_error;
728 #ifndef NDEBUG
729 		{
730 			char ifname[IF_NAMESIZE];
731 
732 			if (priv_get_ifname(priv, &ifname) == 0)
733 				DEBUG("port %u ifname is \"%s\"",
734 				      priv->port, ifname);
735 			else
736 				DEBUG("port %u ifname is unknown", priv->port);
737 		}
738 #endif
739 		/* Get actual MTU if possible. */
740 		priv_get_mtu(priv, &priv->mtu);
741 		DEBUG("port %u MTU is %u", priv->port, priv->mtu);
742 
743 		/* from rte_ethdev.c */
744 		{
745 			char name[RTE_ETH_NAME_MAX_LEN];
746 
747 			snprintf(name, sizeof(name), "%s port %u",
748 				 ibv_get_device_name(ibv_dev), port);
749 			eth_dev = rte_eth_dev_allocate(name);
750 		}
751 		if (eth_dev == NULL) {
752 			ERROR("can not allocate rte ethdev");
753 			err = ENOMEM;
754 			goto port_error;
755 		}
756 
757 		/* Secondary processes have to use local storage for their
758 		 * private data as well as a copy of eth_dev->data, but this
759 		 * pointer must not be modified before burst functions are
760 		 * actually called. */
761 		if (mlx5_is_secondary()) {
762 			struct mlx5_secondary_data *sd =
763 				&mlx5_secondary_data[eth_dev->data->port_id];
764 			sd->primary_priv = eth_dev->data->dev_private;
765 			if (sd->primary_priv == NULL) {
766 				ERROR("no private data for port %u",
767 						eth_dev->data->port_id);
768 				err = EINVAL;
769 				goto port_error;
770 			}
771 			sd->shared_dev_data = eth_dev->data;
772 			rte_spinlock_init(&sd->lock);
773 			memcpy(sd->data.name, sd->shared_dev_data->name,
774 				   sizeof(sd->data.name));
775 			sd->data.dev_private = priv;
776 			sd->data.rx_mbuf_alloc_failed = 0;
777 			sd->data.mtu = ETHER_MTU;
778 			sd->data.port_id = sd->shared_dev_data->port_id;
779 			sd->data.mac_addrs = priv->mac;
780 			eth_dev->tx_pkt_burst = mlx5_tx_burst_secondary_setup;
781 			eth_dev->rx_pkt_burst = mlx5_rx_burst_secondary_setup;
782 		} else {
783 			eth_dev->data->dev_private = priv;
784 			eth_dev->data->mac_addrs = priv->mac;
785 		}
786 
787 		eth_dev->device = &pci_dev->device;
788 		rte_eth_copy_pci_info(eth_dev, pci_dev);
789 		eth_dev->device->driver = &mlx5_driver.driver;
790 		priv->dev = eth_dev;
791 		eth_dev->dev_ops = &mlx5_dev_ops;
792 
793 		/* Bring Ethernet device up. */
794 		DEBUG("forcing Ethernet interface up");
795 		priv_set_flags(priv, ~IFF_UP, IFF_UP);
796 		mlx5_link_update(priv->dev, 1);
797 		continue;
798 
799 port_error:
800 		if (priv) {
801 			rte_free(priv->rss_conf);
802 			rte_free(priv);
803 		}
804 		if (pd)
805 			claim_zero(ibv_dealloc_pd(pd));
806 		if (ctx)
807 			claim_zero(ibv_close_device(ctx));
808 		break;
809 	}
810 
811 	/*
812 	 * XXX if something went wrong in the loop above, there is a resource
813 	 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
814 	 * long as the dpdk does not provide a way to deallocate a ethdev and a
815 	 * way to enumerate the registered ethdevs to free the previous ones.
816 	 */
817 
818 	/* no port found, complain */
819 	if (!mlx5_dev[idx].ports) {
820 		err = ENODEV;
821 		goto error;
822 	}
823 
824 error:
825 	if (attr_ctx)
826 		claim_zero(ibv_close_device(attr_ctx));
827 	if (list)
828 		ibv_free_device_list(list);
829 	assert(err >= 0);
830 	return -err;
831 }
832 
833 static const struct rte_pci_id mlx5_pci_id_map[] = {
834 	{
835 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
836 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
837 	},
838 	{
839 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
840 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
841 	},
842 	{
843 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
844 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
845 	},
846 	{
847 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
848 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
849 	},
850 	{
851 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
852 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
853 	},
854 	{
855 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
856 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
857 	},
858 	{
859 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
860 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
861 	},
862 	{
863 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
864 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
865 	},
866 	{
867 		.vendor_id = 0
868 	}
869 };
870 
871 static struct rte_pci_driver mlx5_driver = {
872 	.driver = {
873 		.name = MLX5_DRIVER_NAME
874 	},
875 	.id_table = mlx5_pci_id_map,
876 	.probe = mlx5_pci_probe,
877 	.drv_flags = RTE_PCI_DRV_INTR_LSC,
878 };
879 
880 /**
881  * Driver initialization routine.
882  */
883 RTE_INIT(rte_mlx5_pmd_init);
884 static void
885 rte_mlx5_pmd_init(void)
886 {
887 	/*
888 	 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
889 	 * huge pages. Calling ibv_fork_init() during init allows
890 	 * applications to use fork() safely for purposes other than
891 	 * using this PMD, which is not supported in forked processes.
892 	 */
893 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
894 	ibv_fork_init();
895 	rte_eal_pci_register(&mlx5_driver);
896 }
897 
898 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
899 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
900 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
901