xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision ccdcba53a3f4d12583c4378e7d8b702ce91e0499)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #define _GNU_SOURCE
7 
8 #include <stddef.h>
9 #include <assert.h>
10 #include <inttypes.h>
11 #include <unistd.h>
12 #include <stdint.h>
13 #include <stdio.h>
14 #include <string.h>
15 #include <stdlib.h>
16 #include <errno.h>
17 #include <dirent.h>
18 #include <net/if.h>
19 #include <sys/ioctl.h>
20 #include <sys/socket.h>
21 #include <netinet/in.h>
22 #include <linux/ethtool.h>
23 #include <linux/sockios.h>
24 #include <fcntl.h>
25 #include <stdalign.h>
26 #include <sys/un.h>
27 #include <time.h>
28 
29 #include <rte_atomic.h>
30 #include <rte_ethdev_driver.h>
31 #include <rte_bus_pci.h>
32 #include <rte_mbuf.h>
33 #include <rte_common.h>
34 #include <rte_interrupts.h>
35 #include <rte_malloc.h>
36 #include <rte_string_fns.h>
37 
38 #include "mlx5.h"
39 #include "mlx5_glue.h"
40 #include "mlx5_rxtx.h"
41 #include "mlx5_utils.h"
42 
43 /* Add defines in case the running kernel is not the same as user headers. */
44 #ifndef ETHTOOL_GLINKSETTINGS
45 struct ethtool_link_settings {
46 	uint32_t cmd;
47 	uint32_t speed;
48 	uint8_t duplex;
49 	uint8_t port;
50 	uint8_t phy_address;
51 	uint8_t autoneg;
52 	uint8_t mdio_support;
53 	uint8_t eth_to_mdix;
54 	uint8_t eth_tp_mdix_ctrl;
55 	int8_t link_mode_masks_nwords;
56 	uint32_t reserved[8];
57 	uint32_t link_mode_masks[];
58 };
59 
60 #define ETHTOOL_GLINKSETTINGS 0x0000004c
61 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
62 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
63 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
64 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
65 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
66 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
67 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
68 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
69 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
70 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
71 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
72 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
73 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
74 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
75 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
76 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
77 #endif
78 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
79 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
80 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
81 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
82 #endif
83 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
84 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
85 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
86 #endif
87 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
88 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
89 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
90 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
91 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
92 #endif
93 
94 /**
95  * Get interface name from private structure.
96  *
97  * @param[in] dev
98  *   Pointer to Ethernet device.
99  * @param[out] ifname
100  *   Interface name output buffer.
101  *
102  * @return
103  *   0 on success, a negative errno value otherwise and rte_errno is set.
104  */
105 int
106 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
107 {
108 	struct priv *priv = dev->data->dev_private;
109 	DIR *dir;
110 	struct dirent *dent;
111 	unsigned int dev_type = 0;
112 	unsigned int dev_port_prev = ~0u;
113 	char match[IF_NAMESIZE] = "";
114 
115 	{
116 		MKSTR(path, "%s/device/net", priv->ibdev_path);
117 
118 		dir = opendir(path);
119 		if (dir == NULL) {
120 			rte_errno = errno;
121 			return -rte_errno;
122 		}
123 	}
124 	while ((dent = readdir(dir)) != NULL) {
125 		char *name = dent->d_name;
126 		FILE *file;
127 		unsigned int dev_port;
128 		int r;
129 
130 		if ((name[0] == '.') &&
131 		    ((name[1] == '\0') ||
132 		     ((name[1] == '.') && (name[2] == '\0'))))
133 			continue;
134 
135 		MKSTR(path, "%s/device/net/%s/%s",
136 		      priv->ibdev_path, name,
137 		      (dev_type ? "dev_id" : "dev_port"));
138 
139 		file = fopen(path, "rb");
140 		if (file == NULL) {
141 			if (errno != ENOENT)
142 				continue;
143 			/*
144 			 * Switch to dev_id when dev_port does not exist as
145 			 * is the case with Linux kernel versions < 3.15.
146 			 */
147 try_dev_id:
148 			match[0] = '\0';
149 			if (dev_type)
150 				break;
151 			dev_type = 1;
152 			dev_port_prev = ~0u;
153 			rewinddir(dir);
154 			continue;
155 		}
156 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
157 		fclose(file);
158 		if (r != 1)
159 			continue;
160 		/*
161 		 * Switch to dev_id when dev_port returns the same value for
162 		 * all ports. May happen when using a MOFED release older than
163 		 * 3.0 with a Linux kernel >= 3.15.
164 		 */
165 		if (dev_port == dev_port_prev)
166 			goto try_dev_id;
167 		dev_port_prev = dev_port;
168 		if (dev_port == (priv->port - 1u))
169 			strlcpy(match, name, sizeof(match));
170 	}
171 	closedir(dir);
172 	if (match[0] == '\0') {
173 		rte_errno = ENOENT;
174 		return -rte_errno;
175 	}
176 	strncpy(*ifname, match, sizeof(*ifname));
177 	return 0;
178 }
179 
180 /**
181  * Get the interface index from device name.
182  *
183  * @param[in] dev
184  *   Pointer to Ethernet device.
185  *
186  * @return
187  *   Interface index on success, a negative errno value otherwise and
188  *   rte_errno is set.
189  */
190 int
191 mlx5_ifindex(const struct rte_eth_dev *dev)
192 {
193 	char ifname[IF_NAMESIZE];
194 	int ret;
195 
196 	ret = mlx5_get_ifname(dev, &ifname);
197 	if (ret)
198 		return ret;
199 	ret = if_nametoindex(ifname);
200 	if (ret == -1) {
201 		rte_errno = errno;
202 		return -rte_errno;
203 	}
204 	return ret;
205 }
206 
207 /**
208  * Perform ifreq ioctl() on associated Ethernet device.
209  *
210  * @param[in] dev
211  *   Pointer to Ethernet device.
212  * @param req
213  *   Request number to pass to ioctl().
214  * @param[out] ifr
215  *   Interface request structure output buffer.
216  *
217  * @return
218  *   0 on success, a negative errno value otherwise and rte_errno is set.
219  */
220 int
221 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
222 {
223 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
224 	int ret = 0;
225 
226 	if (sock == -1) {
227 		rte_errno = errno;
228 		return -rte_errno;
229 	}
230 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
231 	if (ret)
232 		goto error;
233 	ret = ioctl(sock, req, ifr);
234 	if (ret == -1) {
235 		rte_errno = errno;
236 		goto error;
237 	}
238 	close(sock);
239 	return 0;
240 error:
241 	close(sock);
242 	return -rte_errno;
243 }
244 
245 /**
246  * Get device MTU.
247  *
248  * @param dev
249  *   Pointer to Ethernet device.
250  * @param[out] mtu
251  *   MTU value output buffer.
252  *
253  * @return
254  *   0 on success, a negative errno value otherwise and rte_errno is set.
255  */
256 int
257 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
258 {
259 	struct ifreq request;
260 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
261 
262 	if (ret)
263 		return ret;
264 	*mtu = request.ifr_mtu;
265 	return 0;
266 }
267 
268 /**
269  * Set device MTU.
270  *
271  * @param dev
272  *   Pointer to Ethernet device.
273  * @param mtu
274  *   MTU value to set.
275  *
276  * @return
277  *   0 on success, a negative errno value otherwise and rte_errno is set.
278  */
279 static int
280 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
281 {
282 	struct ifreq request = { .ifr_mtu = mtu, };
283 
284 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
285 }
286 
287 /**
288  * Set device flags.
289  *
290  * @param dev
291  *   Pointer to Ethernet device.
292  * @param keep
293  *   Bitmask for flags that must remain untouched.
294  * @param flags
295  *   Bitmask for flags to modify.
296  *
297  * @return
298  *   0 on success, a negative errno value otherwise and rte_errno is set.
299  */
300 int
301 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
302 {
303 	struct ifreq request;
304 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
305 
306 	if (ret)
307 		return ret;
308 	request.ifr_flags &= keep;
309 	request.ifr_flags |= flags & ~keep;
310 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
311 }
312 
313 /**
314  * DPDK callback for Ethernet device configuration.
315  *
316  * @param dev
317  *   Pointer to Ethernet device structure.
318  *
319  * @return
320  *   0 on success, a negative errno value otherwise and rte_errno is set.
321  */
322 int
323 mlx5_dev_configure(struct rte_eth_dev *dev)
324 {
325 	struct priv *priv = dev->data->dev_private;
326 	unsigned int rxqs_n = dev->data->nb_rx_queues;
327 	unsigned int txqs_n = dev->data->nb_tx_queues;
328 	unsigned int i;
329 	unsigned int j;
330 	unsigned int reta_idx_n;
331 	const uint8_t use_app_rss_key =
332 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
333 	uint64_t supp_tx_offloads = mlx5_get_tx_port_offloads(dev);
334 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
335 	uint64_t supp_rx_offloads =
336 		(mlx5_get_rx_port_offloads() |
337 		 mlx5_get_rx_queue_offloads(dev));
338 	uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
339 	int ret = 0;
340 
341 	if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
342 		DRV_LOG(ERR,
343 			"port %u some Tx offloads are not supported requested"
344 			" 0x%" PRIx64 " supported 0x%" PRIx64,
345 			dev->data->port_id, tx_offloads, supp_tx_offloads);
346 		rte_errno = ENOTSUP;
347 		return -rte_errno;
348 	}
349 	if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
350 		DRV_LOG(ERR,
351 			"port %u some Rx offloads are not supported requested"
352 			" 0x%" PRIx64 " supported 0x%" PRIx64,
353 			dev->data->port_id, rx_offloads, supp_rx_offloads);
354 		rte_errno = ENOTSUP;
355 		return -rte_errno;
356 	}
357 	if (use_app_rss_key &&
358 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
359 	     rss_hash_default_key_len)) {
360 		DRV_LOG(ERR, "port %u RSS key len must be %zu Bytes long",
361 			dev->data->port_id, rss_hash_default_key_len);
362 		rte_errno = EINVAL;
363 		return -rte_errno;
364 	}
365 	priv->rss_conf.rss_key =
366 		rte_realloc(priv->rss_conf.rss_key,
367 			    rss_hash_default_key_len, 0);
368 	if (!priv->rss_conf.rss_key) {
369 		DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
370 			dev->data->port_id, rxqs_n);
371 		rte_errno = ENOMEM;
372 		return -rte_errno;
373 	}
374 	memcpy(priv->rss_conf.rss_key,
375 	       use_app_rss_key ?
376 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
377 	       rss_hash_default_key,
378 	       rss_hash_default_key_len);
379 	priv->rss_conf.rss_key_len = rss_hash_default_key_len;
380 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
381 	priv->rxqs = (void *)dev->data->rx_queues;
382 	priv->txqs = (void *)dev->data->tx_queues;
383 	if (txqs_n != priv->txqs_n) {
384 		DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
385 			dev->data->port_id, priv->txqs_n, txqs_n);
386 		priv->txqs_n = txqs_n;
387 	}
388 	if (rxqs_n > priv->config.ind_table_max_size) {
389 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
390 			dev->data->port_id, rxqs_n);
391 		rte_errno = EINVAL;
392 		return -rte_errno;
393 	}
394 	if (rxqs_n == priv->rxqs_n)
395 		return 0;
396 	DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
397 		dev->data->port_id, priv->rxqs_n, rxqs_n);
398 	priv->rxqs_n = rxqs_n;
399 	/* If the requested number of RX queues is not a power of two, use the
400 	 * maximum indirection table size for better balancing.
401 	 * The result is always rounded to the next power of two. */
402 	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
403 				     priv->config.ind_table_max_size :
404 				     rxqs_n));
405 	ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
406 	if (ret)
407 		return ret;
408 	/* When the number of RX queues is not a power of two, the remaining
409 	 * table entries are padded with reused WQs and hashes are not spread
410 	 * uniformly. */
411 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
412 		(*priv->reta_idx)[i] = j;
413 		if (++j == rxqs_n)
414 			j = 0;
415 	}
416 	return 0;
417 }
418 
419 /**
420  * DPDK callback to get information about the device.
421  *
422  * @param dev
423  *   Pointer to Ethernet device structure.
424  * @param[out] info
425  *   Info structure output buffer.
426  */
427 void
428 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
429 {
430 	struct priv *priv = dev->data->dev_private;
431 	struct mlx5_dev_config *config = &priv->config;
432 	unsigned int max;
433 	char ifname[IF_NAMESIZE];
434 
435 	info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
436 	/* FIXME: we should ask the device for these values. */
437 	info->min_rx_bufsize = 32;
438 	info->max_rx_pktlen = 65536;
439 	/*
440 	 * Since we need one CQ per QP, the limit is the minimum number
441 	 * between the two values.
442 	 */
443 	max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
444 		      priv->device_attr.orig_attr.max_qp);
445 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
446 	if (max >= 65535)
447 		max = 65535;
448 	info->max_rx_queues = max;
449 	info->max_tx_queues = max;
450 	info->max_mac_addrs = RTE_DIM(priv->mac);
451 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
452 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
453 				 info->rx_queue_offload_capa);
454 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
455 	if (mlx5_get_ifname(dev, &ifname) == 0)
456 		info->if_index = if_nametoindex(ifname);
457 	info->reta_size = priv->reta_idx_n ?
458 		priv->reta_idx_n : config->ind_table_max_size;
459 	info->hash_key_size = rss_hash_default_key_len;
460 	info->speed_capa = priv->link_speed_capa;
461 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
462 }
463 
464 /**
465  * Get supported packet types.
466  *
467  * @param dev
468  *   Pointer to Ethernet device structure.
469  *
470  * @return
471  *   A pointer to the supported Packet types array.
472  */
473 const uint32_t *
474 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
475 {
476 	static const uint32_t ptypes[] = {
477 		/* refers to rxq_cq_to_pkt_type() */
478 		RTE_PTYPE_L2_ETHER,
479 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
480 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
481 		RTE_PTYPE_L4_NONFRAG,
482 		RTE_PTYPE_L4_FRAG,
483 		RTE_PTYPE_L4_TCP,
484 		RTE_PTYPE_L4_UDP,
485 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
486 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
487 		RTE_PTYPE_INNER_L4_NONFRAG,
488 		RTE_PTYPE_INNER_L4_FRAG,
489 		RTE_PTYPE_INNER_L4_TCP,
490 		RTE_PTYPE_INNER_L4_UDP,
491 		RTE_PTYPE_UNKNOWN
492 	};
493 
494 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
495 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
496 		return ptypes;
497 	return NULL;
498 }
499 
500 /**
501  * DPDK callback to retrieve physical link information.
502  *
503  * @param dev
504  *   Pointer to Ethernet device structure.
505  * @param[out] link
506  *   Storage for current link status.
507  *
508  * @return
509  *   0 on success, a negative errno value otherwise and rte_errno is set.
510  */
511 static int
512 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
513 			       struct rte_eth_link *link)
514 {
515 	struct priv *priv = dev->data->dev_private;
516 	struct ethtool_cmd edata = {
517 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
518 	};
519 	struct ifreq ifr;
520 	struct rte_eth_link dev_link;
521 	int link_speed = 0;
522 	int ret;
523 
524 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
525 	if (ret) {
526 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
527 			dev->data->port_id, strerror(rte_errno));
528 		return ret;
529 	}
530 	memset(&dev_link, 0, sizeof(dev_link));
531 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
532 				(ifr.ifr_flags & IFF_RUNNING));
533 	ifr.ifr_data = (void *)&edata;
534 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
535 	if (ret) {
536 		DRV_LOG(WARNING,
537 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
538 			dev->data->port_id, strerror(rte_errno));
539 		return ret;
540 	}
541 	link_speed = ethtool_cmd_speed(&edata);
542 	if (link_speed == -1)
543 		dev_link.link_speed = 0;
544 	else
545 		dev_link.link_speed = link_speed;
546 	priv->link_speed_capa = 0;
547 	if (edata.supported & SUPPORTED_Autoneg)
548 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
549 	if (edata.supported & (SUPPORTED_1000baseT_Full |
550 			       SUPPORTED_1000baseKX_Full))
551 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
552 	if (edata.supported & SUPPORTED_10000baseKR_Full)
553 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
554 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
555 			       SUPPORTED_40000baseCR4_Full |
556 			       SUPPORTED_40000baseSR4_Full |
557 			       SUPPORTED_40000baseLR4_Full))
558 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
559 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
560 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
561 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
562 			ETH_LINK_SPEED_FIXED);
563 	if ((dev_link.link_speed && !dev_link.link_status) ||
564 	    (!dev_link.link_speed && dev_link.link_status)) {
565 		rte_errno = EAGAIN;
566 		return -rte_errno;
567 	}
568 	*link = dev_link;
569 	return 0;
570 }
571 
572 /**
573  * Retrieve physical link information (unlocked version using new ioctl).
574  *
575  * @param dev
576  *   Pointer to Ethernet device structure.
577  * @param[out] link
578  *   Storage for current link status.
579  *
580  * @return
581  *   0 on success, a negative errno value otherwise and rte_errno is set.
582  */
583 static int
584 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
585 			     struct rte_eth_link *link)
586 
587 {
588 	struct priv *priv = dev->data->dev_private;
589 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
590 	struct ifreq ifr;
591 	struct rte_eth_link dev_link;
592 	uint64_t sc;
593 	int ret;
594 
595 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
596 	if (ret) {
597 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
598 			dev->data->port_id, strerror(rte_errno));
599 		return ret;
600 	}
601 	memset(&dev_link, 0, sizeof(dev_link));
602 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
603 				(ifr.ifr_flags & IFF_RUNNING));
604 	ifr.ifr_data = (void *)&gcmd;
605 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
606 	if (ret) {
607 		DRV_LOG(DEBUG,
608 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
609 			" failed: %s",
610 			dev->data->port_id, strerror(rte_errno));
611 		return ret;
612 	}
613 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
614 
615 	alignas(struct ethtool_link_settings)
616 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
617 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
618 	struct ethtool_link_settings *ecmd = (void *)data;
619 
620 	*ecmd = gcmd;
621 	ifr.ifr_data = (void *)ecmd;
622 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
623 	if (ret) {
624 		DRV_LOG(DEBUG,
625 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
626 			" failed: %s",
627 			dev->data->port_id, strerror(rte_errno));
628 		return ret;
629 	}
630 	dev_link.link_speed = ecmd->speed;
631 	sc = ecmd->link_mode_masks[0] |
632 		((uint64_t)ecmd->link_mode_masks[1] << 32);
633 	priv->link_speed_capa = 0;
634 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
635 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
636 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
637 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
638 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
639 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
640 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
641 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
642 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
643 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
644 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
645 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
646 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
647 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
648 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
649 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
650 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
651 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
652 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
653 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
654 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
655 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
656 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
657 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
658 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
659 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
660 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
661 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
662 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
663 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
664 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
665 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
666 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
667 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
668 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
669 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
670 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
671 				  ETH_LINK_SPEED_FIXED);
672 	if ((dev_link.link_speed && !dev_link.link_status) ||
673 	    (!dev_link.link_speed && dev_link.link_status)) {
674 		rte_errno = EAGAIN;
675 		return -rte_errno;
676 	}
677 	*link = dev_link;
678 	return 0;
679 }
680 
681 /**
682  * DPDK callback to retrieve physical link information.
683  *
684  * @param dev
685  *   Pointer to Ethernet device structure.
686  * @param wait_to_complete
687  *   Wait for request completion.
688  *
689  * @return
690  *   0 if link status was not updated, positive if it was, a negative errno
691  *   value otherwise and rte_errno is set.
692  */
693 int
694 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
695 {
696 	int ret;
697 	struct rte_eth_link dev_link;
698 	time_t start_time = time(NULL);
699 
700 	do {
701 		ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
702 		if (ret)
703 			ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
704 		if (ret == 0)
705 			break;
706 		/* Handle wait to complete situation. */
707 		if (wait_to_complete && ret == -EAGAIN) {
708 			if (abs((int)difftime(time(NULL), start_time)) <
709 			    MLX5_LINK_STATUS_TIMEOUT) {
710 				usleep(0);
711 				continue;
712 			} else {
713 				rte_errno = EBUSY;
714 				return -rte_errno;
715 			}
716 		} else if (ret < 0) {
717 			return ret;
718 		}
719 	} while (wait_to_complete);
720 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
721 		       sizeof(struct rte_eth_link));
722 	dev->data->dev_link = dev_link;
723 	return ret;
724 }
725 
726 /**
727  * DPDK callback to change the MTU.
728  *
729  * @param dev
730  *   Pointer to Ethernet device structure.
731  * @param in_mtu
732  *   New MTU.
733  *
734  * @return
735  *   0 on success, a negative errno value otherwise and rte_errno is set.
736  */
737 int
738 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
739 {
740 	struct priv *priv = dev->data->dev_private;
741 	uint16_t kern_mtu = 0;
742 	int ret;
743 
744 	ret = mlx5_get_mtu(dev, &kern_mtu);
745 	if (ret)
746 		return ret;
747 	/* Set kernel interface MTU first. */
748 	ret = mlx5_set_mtu(dev, mtu);
749 	if (ret)
750 		return ret;
751 	ret = mlx5_get_mtu(dev, &kern_mtu);
752 	if (ret)
753 		return ret;
754 	if (kern_mtu == mtu) {
755 		priv->mtu = mtu;
756 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
757 			dev->data->port_id, mtu);
758 		return 0;
759 	}
760 	rte_errno = EAGAIN;
761 	return -rte_errno;
762 }
763 
764 /**
765  * DPDK callback to get flow control status.
766  *
767  * @param dev
768  *   Pointer to Ethernet device structure.
769  * @param[out] fc_conf
770  *   Flow control output buffer.
771  *
772  * @return
773  *   0 on success, a negative errno value otherwise and rte_errno is set.
774  */
775 int
776 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
777 {
778 	struct ifreq ifr;
779 	struct ethtool_pauseparam ethpause = {
780 		.cmd = ETHTOOL_GPAUSEPARAM
781 	};
782 	int ret;
783 
784 	ifr.ifr_data = (void *)&ethpause;
785 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
786 	if (ret) {
787 		DRV_LOG(WARNING,
788 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
789 			" %s",
790 			dev->data->port_id, strerror(rte_errno));
791 		return ret;
792 	}
793 	fc_conf->autoneg = ethpause.autoneg;
794 	if (ethpause.rx_pause && ethpause.tx_pause)
795 		fc_conf->mode = RTE_FC_FULL;
796 	else if (ethpause.rx_pause)
797 		fc_conf->mode = RTE_FC_RX_PAUSE;
798 	else if (ethpause.tx_pause)
799 		fc_conf->mode = RTE_FC_TX_PAUSE;
800 	else
801 		fc_conf->mode = RTE_FC_NONE;
802 	return 0;
803 }
804 
805 /**
806  * DPDK callback to modify flow control parameters.
807  *
808  * @param dev
809  *   Pointer to Ethernet device structure.
810  * @param[in] fc_conf
811  *   Flow control parameters.
812  *
813  * @return
814  *   0 on success, a negative errno value otherwise and rte_errno is set.
815  */
816 int
817 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
818 {
819 	struct ifreq ifr;
820 	struct ethtool_pauseparam ethpause = {
821 		.cmd = ETHTOOL_SPAUSEPARAM
822 	};
823 	int ret;
824 
825 	ifr.ifr_data = (void *)&ethpause;
826 	ethpause.autoneg = fc_conf->autoneg;
827 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
828 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
829 		ethpause.rx_pause = 1;
830 	else
831 		ethpause.rx_pause = 0;
832 
833 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
834 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
835 		ethpause.tx_pause = 1;
836 	else
837 		ethpause.tx_pause = 0;
838 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
839 	if (ret) {
840 		DRV_LOG(WARNING,
841 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
842 			" failed: %s",
843 			dev->data->port_id, strerror(rte_errno));
844 		return ret;
845 	}
846 	return 0;
847 }
848 
849 /**
850  * Get PCI information from struct ibv_device.
851  *
852  * @param device
853  *   Pointer to Ethernet device structure.
854  * @param[out] pci_addr
855  *   PCI bus address output buffer.
856  *
857  * @return
858  *   0 on success, a negative errno value otherwise and rte_errno is set.
859  */
860 int
861 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
862 			    struct rte_pci_addr *pci_addr)
863 {
864 	FILE *file;
865 	char line[32];
866 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
867 
868 	file = fopen(path, "rb");
869 	if (file == NULL) {
870 		rte_errno = errno;
871 		return -rte_errno;
872 	}
873 	while (fgets(line, sizeof(line), file) == line) {
874 		size_t len = strlen(line);
875 		int ret;
876 
877 		/* Truncate long lines. */
878 		if (len == (sizeof(line) - 1))
879 			while (line[(len - 1)] != '\n') {
880 				ret = fgetc(file);
881 				if (ret == EOF)
882 					break;
883 				line[(len - 1)] = ret;
884 			}
885 		/* Extract information. */
886 		if (sscanf(line,
887 			   "PCI_SLOT_NAME="
888 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
889 			   &pci_addr->domain,
890 			   &pci_addr->bus,
891 			   &pci_addr->devid,
892 			   &pci_addr->function) == 4) {
893 			ret = 0;
894 			break;
895 		}
896 	}
897 	fclose(file);
898 	return 0;
899 }
900 
901 /**
902  * Device status handler.
903  *
904  * @param dev
905  *   Pointer to Ethernet device.
906  * @param events
907  *   Pointer to event flags holder.
908  *
909  * @return
910  *   Events bitmap of callback process which can be called immediately.
911  */
912 static uint32_t
913 mlx5_dev_status_handler(struct rte_eth_dev *dev)
914 {
915 	struct priv *priv = dev->data->dev_private;
916 	struct ibv_async_event event;
917 	uint32_t ret = 0;
918 
919 	if (mlx5_link_update(dev, 0) == -EAGAIN) {
920 		usleep(0);
921 		return 0;
922 	}
923 	/* Read all message and acknowledge them. */
924 	for (;;) {
925 		if (mlx5_glue->get_async_event(priv->ctx, &event))
926 			break;
927 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
928 			event.event_type == IBV_EVENT_PORT_ERR) &&
929 			(dev->data->dev_conf.intr_conf.lsc == 1))
930 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
931 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
932 			dev->data->dev_conf.intr_conf.rmv == 1)
933 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
934 		else
935 			DRV_LOG(DEBUG,
936 				"port %u event type %d on not handled",
937 				dev->data->port_id, event.event_type);
938 		mlx5_glue->ack_async_event(&event);
939 	}
940 	return ret;
941 }
942 
943 /**
944  * Handle interrupts from the NIC.
945  *
946  * @param[in] intr_handle
947  *   Interrupt handler.
948  * @param cb_arg
949  *   Callback argument.
950  */
951 void
952 mlx5_dev_interrupt_handler(void *cb_arg)
953 {
954 	struct rte_eth_dev *dev = cb_arg;
955 	uint32_t events;
956 
957 	events = mlx5_dev_status_handler(dev);
958 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
959 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
960 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
961 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
962 }
963 
964 /**
965  * Handle interrupts from the socket.
966  *
967  * @param cb_arg
968  *   Callback argument.
969  */
970 static void
971 mlx5_dev_handler_socket(void *cb_arg)
972 {
973 	struct rte_eth_dev *dev = cb_arg;
974 
975 	mlx5_socket_handle(dev);
976 }
977 
978 /**
979  * Uninstall interrupt handler.
980  *
981  * @param dev
982  *   Pointer to Ethernet device.
983  */
984 void
985 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
986 {
987 	struct priv *priv = dev->data->dev_private;
988 
989 	if (dev->data->dev_conf.intr_conf.lsc ||
990 	    dev->data->dev_conf.intr_conf.rmv)
991 		rte_intr_callback_unregister(&priv->intr_handle,
992 					     mlx5_dev_interrupt_handler, dev);
993 	if (priv->primary_socket)
994 		rte_intr_callback_unregister(&priv->intr_handle_socket,
995 					     mlx5_dev_handler_socket, dev);
996 	priv->intr_handle.fd = 0;
997 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
998 	priv->intr_handle_socket.fd = 0;
999 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1000 }
1001 
1002 /**
1003  * Install interrupt handler.
1004  *
1005  * @param dev
1006  *   Pointer to Ethernet device.
1007  */
1008 void
1009 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1010 {
1011 	struct priv *priv = dev->data->dev_private;
1012 	int ret;
1013 	int flags;
1014 
1015 	assert(priv->ctx->async_fd > 0);
1016 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1017 	ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1018 	if (ret) {
1019 		DRV_LOG(INFO,
1020 			"port %u failed to change file descriptor async event"
1021 			" queue",
1022 			dev->data->port_id);
1023 		dev->data->dev_conf.intr_conf.lsc = 0;
1024 		dev->data->dev_conf.intr_conf.rmv = 0;
1025 	}
1026 	if (dev->data->dev_conf.intr_conf.lsc ||
1027 	    dev->data->dev_conf.intr_conf.rmv) {
1028 		priv->intr_handle.fd = priv->ctx->async_fd;
1029 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1030 		rte_intr_callback_register(&priv->intr_handle,
1031 					   mlx5_dev_interrupt_handler, dev);
1032 	}
1033 	ret = mlx5_socket_init(dev);
1034 	if (ret)
1035 		DRV_LOG(ERR, "port %u cannot initialise socket: %s",
1036 			dev->data->port_id, strerror(rte_errno));
1037 	else if (priv->primary_socket) {
1038 		priv->intr_handle_socket.fd = priv->primary_socket;
1039 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1040 		rte_intr_callback_register(&priv->intr_handle_socket,
1041 					   mlx5_dev_handler_socket, dev);
1042 	}
1043 }
1044 
1045 /**
1046  * DPDK callback to bring the link DOWN.
1047  *
1048  * @param dev
1049  *   Pointer to Ethernet device structure.
1050  *
1051  * @return
1052  *   0 on success, a negative errno value otherwise and rte_errno is set.
1053  */
1054 int
1055 mlx5_set_link_down(struct rte_eth_dev *dev)
1056 {
1057 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1058 }
1059 
1060 /**
1061  * DPDK callback to bring the link UP.
1062  *
1063  * @param dev
1064  *   Pointer to Ethernet device structure.
1065  *
1066  * @return
1067  *   0 on success, a negative errno value otherwise and rte_errno is set.
1068  */
1069 int
1070 mlx5_set_link_up(struct rte_eth_dev *dev)
1071 {
1072 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1073 }
1074 
1075 /**
1076  * Configure the TX function to use.
1077  *
1078  * @param dev
1079  *   Pointer to private data structure.
1080  *
1081  * @return
1082  *   Pointer to selected Tx burst function.
1083  */
1084 eth_tx_burst_t
1085 mlx5_select_tx_function(struct rte_eth_dev *dev)
1086 {
1087 	struct priv *priv = dev->data->dev_private;
1088 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1089 	struct mlx5_dev_config *config = &priv->config;
1090 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1091 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1092 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1093 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1094 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1095 
1096 	assert(priv != NULL);
1097 	/* Select appropriate TX function. */
1098 	if (vlan_insert || tso)
1099 		return tx_pkt_burst;
1100 	if (config->mps == MLX5_MPW_ENHANCED) {
1101 		if (mlx5_check_vec_tx_support(dev) > 0) {
1102 			if (mlx5_check_raw_vec_tx_support(dev) > 0)
1103 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1104 			else
1105 				tx_pkt_burst = mlx5_tx_burst_vec;
1106 			DRV_LOG(DEBUG,
1107 				"port %u selected enhanced MPW Tx vectorized"
1108 				" function",
1109 				dev->data->port_id);
1110 		} else {
1111 			tx_pkt_burst = mlx5_tx_burst_empw;
1112 			DRV_LOG(DEBUG,
1113 				"port %u selected enhanced MPW Tx function",
1114 				dev->data->port_id);
1115 		}
1116 	} else if (config->mps && (config->txq_inline > 0)) {
1117 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1118 		DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1119 			dev->data->port_id);
1120 	} else if (config->mps) {
1121 		tx_pkt_burst = mlx5_tx_burst_mpw;
1122 		DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1123 			dev->data->port_id);
1124 	}
1125 	return tx_pkt_burst;
1126 }
1127 
1128 /**
1129  * Configure the RX function to use.
1130  *
1131  * @param dev
1132  *   Pointer to private data structure.
1133  *
1134  * @return
1135  *   Pointer to selected Rx burst function.
1136  */
1137 eth_rx_burst_t
1138 mlx5_select_rx_function(struct rte_eth_dev *dev)
1139 {
1140 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1141 
1142 	assert(dev != NULL);
1143 	if (mlx5_check_vec_rx_support(dev) > 0) {
1144 		rx_pkt_burst = mlx5_rx_burst_vec;
1145 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1146 			dev->data->port_id);
1147 	}
1148 	return rx_pkt_burst;
1149 }
1150 
1151 /**
1152  * Check if mlx5 device was removed.
1153  *
1154  * @param dev
1155  *   Pointer to Ethernet device structure.
1156  *
1157  * @return
1158  *   1 when device is removed, otherwise 0.
1159  */
1160 int
1161 mlx5_is_removed(struct rte_eth_dev *dev)
1162 {
1163 	struct ibv_device_attr device_attr;
1164 	struct priv *priv = dev->data->dev_private;
1165 
1166 	if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1167 		return 1;
1168 	return 0;
1169 }
1170