xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 0f99970b4adc943264df0487904d340124765e68)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox.
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 <sys/utsname.h>
22 #include <netinet/in.h>
23 #include <linux/ethtool.h>
24 #include <linux/sockios.h>
25 #include <linux/version.h>
26 #include <fcntl.h>
27 #include <stdalign.h>
28 #include <sys/un.h>
29 
30 #include <rte_atomic.h>
31 #include <rte_ethdev_driver.h>
32 #include <rte_bus_pci.h>
33 #include <rte_mbuf.h>
34 #include <rte_common.h>
35 #include <rte_interrupts.h>
36 #include <rte_alarm.h>
37 #include <rte_malloc.h>
38 
39 #include "mlx5.h"
40 #include "mlx5_glue.h"
41 #include "mlx5_rxtx.h"
42 #include "mlx5_utils.h"
43 
44 /* Add defines in case the running kernel is not the same as user headers. */
45 #ifndef ETHTOOL_GLINKSETTINGS
46 struct ethtool_link_settings {
47 	uint32_t cmd;
48 	uint32_t speed;
49 	uint8_t duplex;
50 	uint8_t port;
51 	uint8_t phy_address;
52 	uint8_t autoneg;
53 	uint8_t mdio_support;
54 	uint8_t eth_to_mdix;
55 	uint8_t eth_tp_mdix_ctrl;
56 	int8_t link_mode_masks_nwords;
57 	uint32_t reserved[8];
58 	uint32_t link_mode_masks[];
59 };
60 
61 #define ETHTOOL_GLINKSETTINGS 0x0000004c
62 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
63 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
64 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
65 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
66 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
67 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
68 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
69 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
70 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
71 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
72 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
73 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
74 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
75 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
76 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
77 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
78 #endif
79 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
80 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
81 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
82 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
83 #endif
84 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
85 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
86 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
87 #endif
88 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
89 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
90 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
91 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
92 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
93 #endif
94 
95 /**
96  * Get interface name from private structure.
97  *
98  * @param[in] dev
99  *   Pointer to Ethernet device.
100  * @param[out] ifname
101  *   Interface name output buffer.
102  *
103  * @return
104  *   0 on success, a negative errno value otherwise and rte_errno is set.
105  */
106 int
107 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
108 {
109 	struct priv *priv = dev->data->dev_private;
110 	DIR *dir;
111 	struct dirent *dent;
112 	unsigned int dev_type = 0;
113 	unsigned int dev_port_prev = ~0u;
114 	char match[IF_NAMESIZE] = "";
115 
116 	{
117 		MKSTR(path, "%s/device/net", priv->ibdev_path);
118 
119 		dir = opendir(path);
120 		if (dir == NULL) {
121 			rte_errno = errno;
122 			return -rte_errno;
123 		}
124 	}
125 	while ((dent = readdir(dir)) != NULL) {
126 		char *name = dent->d_name;
127 		FILE *file;
128 		unsigned int dev_port;
129 		int r;
130 
131 		if ((name[0] == '.') &&
132 		    ((name[1] == '\0') ||
133 		     ((name[1] == '.') && (name[2] == '\0'))))
134 			continue;
135 
136 		MKSTR(path, "%s/device/net/%s/%s",
137 		      priv->ibdev_path, name,
138 		      (dev_type ? "dev_id" : "dev_port"));
139 
140 		file = fopen(path, "rb");
141 		if (file == NULL) {
142 			if (errno != ENOENT)
143 				continue;
144 			/*
145 			 * Switch to dev_id when dev_port does not exist as
146 			 * is the case with Linux kernel versions < 3.15.
147 			 */
148 try_dev_id:
149 			match[0] = '\0';
150 			if (dev_type)
151 				break;
152 			dev_type = 1;
153 			dev_port_prev = ~0u;
154 			rewinddir(dir);
155 			continue;
156 		}
157 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
158 		fclose(file);
159 		if (r != 1)
160 			continue;
161 		/*
162 		 * Switch to dev_id when dev_port returns the same value for
163 		 * all ports. May happen when using a MOFED release older than
164 		 * 3.0 with a Linux kernel >= 3.15.
165 		 */
166 		if (dev_port == dev_port_prev)
167 			goto try_dev_id;
168 		dev_port_prev = dev_port;
169 		if (dev_port == (priv->port - 1u))
170 			snprintf(match, sizeof(match), "%s", name);
171 	}
172 	closedir(dir);
173 	if (match[0] == '\0') {
174 		rte_errno = ENOENT;
175 		return -rte_errno;
176 	}
177 	strncpy(*ifname, match, sizeof(*ifname));
178 	return 0;
179 }
180 
181 /**
182  * Perform ifreq ioctl() on associated Ethernet device.
183  *
184  * @param[in] dev
185  *   Pointer to Ethernet device.
186  * @param req
187  *   Request number to pass to ioctl().
188  * @param[out] ifr
189  *   Interface request structure output buffer.
190  *
191  * @return
192  *   0 on success, a negative errno value otherwise and rte_errno is set.
193  */
194 int
195 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
196 {
197 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
198 	int ret = 0;
199 
200 	if (sock == -1) {
201 		rte_errno = errno;
202 		return -rte_errno;
203 	}
204 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
205 	if (ret)
206 		goto error;
207 	ret = ioctl(sock, req, ifr);
208 	if (ret == -1) {
209 		rte_errno = errno;
210 		goto error;
211 	}
212 	close(sock);
213 	return 0;
214 error:
215 	close(sock);
216 	return -rte_errno;
217 }
218 
219 /**
220  * Get device MTU.
221  *
222  * @param dev
223  *   Pointer to Ethernet device.
224  * @param[out] mtu
225  *   MTU value output buffer.
226  *
227  * @return
228  *   0 on success, a negative errno value otherwise and rte_errno is set.
229  */
230 int
231 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
232 {
233 	struct ifreq request;
234 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
235 
236 	if (ret)
237 		return ret;
238 	*mtu = request.ifr_mtu;
239 	return 0;
240 }
241 
242 /**
243  * Set device MTU.
244  *
245  * @param dev
246  *   Pointer to Ethernet device.
247  * @param mtu
248  *   MTU value to set.
249  *
250  * @return
251  *   0 on success, a negative errno value otherwise and rte_errno is set.
252  */
253 static int
254 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
255 {
256 	struct ifreq request = { .ifr_mtu = mtu, };
257 
258 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
259 }
260 
261 /**
262  * Set device flags.
263  *
264  * @param dev
265  *   Pointer to Ethernet device.
266  * @param keep
267  *   Bitmask for flags that must remain untouched.
268  * @param flags
269  *   Bitmask for flags to modify.
270  *
271  * @return
272  *   0 on success, a negative errno value otherwise and rte_errno is set.
273  */
274 int
275 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
276 {
277 	struct ifreq request;
278 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
279 
280 	if (ret)
281 		return ret;
282 	request.ifr_flags &= keep;
283 	request.ifr_flags |= flags & ~keep;
284 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
285 }
286 
287 /**
288  * DPDK callback for Ethernet device configuration.
289  *
290  * @param dev
291  *   Pointer to Ethernet device structure.
292  *
293  * @return
294  *   0 on success, a negative errno value otherwise and rte_errno is set.
295  */
296 int
297 mlx5_dev_configure(struct rte_eth_dev *dev)
298 {
299 	struct priv *priv = dev->data->dev_private;
300 	unsigned int rxqs_n = dev->data->nb_rx_queues;
301 	unsigned int txqs_n = dev->data->nb_tx_queues;
302 	unsigned int i;
303 	unsigned int j;
304 	unsigned int reta_idx_n;
305 	const uint8_t use_app_rss_key =
306 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
307 	uint64_t supp_tx_offloads = mlx5_get_tx_port_offloads(dev);
308 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
309 	uint64_t supp_rx_offloads =
310 		(mlx5_get_rx_port_offloads() |
311 		 mlx5_get_rx_queue_offloads(dev));
312 	uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
313 	int ret = 0;
314 
315 	if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
316 		ERROR("port %u some Tx offloads are not supported "
317 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
318 		      dev->data->port_id, tx_offloads, supp_tx_offloads);
319 		rte_errno = ENOTSUP;
320 		return -rte_errno;
321 	}
322 	if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
323 		ERROR("port %u some Rx offloads are not supported "
324 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
325 		      dev->data->port_id, rx_offloads, supp_rx_offloads);
326 		rte_errno = ENOTSUP;
327 		return -rte_errno;
328 	}
329 	if (use_app_rss_key &&
330 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
331 	     rss_hash_default_key_len)) {
332 		/* MLX5 RSS only support 40bytes key. */
333 		rte_errno = EINVAL;
334 		return -rte_errno;
335 	}
336 	priv->rss_conf.rss_key =
337 		rte_realloc(priv->rss_conf.rss_key,
338 			    rss_hash_default_key_len, 0);
339 	if (!priv->rss_conf.rss_key) {
340 		ERROR("port %u cannot allocate RSS hash key memory (%u)",
341 		      dev->data->port_id, rxqs_n);
342 		rte_errno = ENOMEM;
343 		return -rte_errno;
344 	}
345 	memcpy(priv->rss_conf.rss_key,
346 	       use_app_rss_key ?
347 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
348 	       rss_hash_default_key,
349 	       rss_hash_default_key_len);
350 	priv->rss_conf.rss_key_len = rss_hash_default_key_len;
351 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
352 	priv->rxqs = (void *)dev->data->rx_queues;
353 	priv->txqs = (void *)dev->data->tx_queues;
354 	if (txqs_n != priv->txqs_n) {
355 		INFO("port %u Tx queues number update: %u -> %u",
356 		     dev->data->port_id, priv->txqs_n, txqs_n);
357 		priv->txqs_n = txqs_n;
358 	}
359 	if (rxqs_n > priv->config.ind_table_max_size) {
360 		ERROR("port %u cannot handle this many Rx queues (%u)",
361 		      dev->data->port_id, rxqs_n);
362 		rte_errno = EINVAL;
363 		return -rte_errno;
364 	}
365 	if (rxqs_n == priv->rxqs_n)
366 		return 0;
367 	INFO("port %u Rx queues number update: %u -> %u",
368 	     dev->data->port_id, priv->rxqs_n, rxqs_n);
369 	priv->rxqs_n = rxqs_n;
370 	/* If the requested number of RX queues is not a power of two, use the
371 	 * maximum indirection table size for better balancing.
372 	 * The result is always rounded to the next power of two. */
373 	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
374 				     priv->config.ind_table_max_size :
375 				     rxqs_n));
376 	ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
377 	if (ret)
378 		return ret;
379 	/* When the number of RX queues is not a power of two, the remaining
380 	 * table entries are padded with reused WQs and hashes are not spread
381 	 * uniformly. */
382 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
383 		(*priv->reta_idx)[i] = j;
384 		if (++j == rxqs_n)
385 			j = 0;
386 	}
387 	return 0;
388 }
389 
390 /**
391  * DPDK callback to get information about the device.
392  *
393  * @param dev
394  *   Pointer to Ethernet device structure.
395  * @param[out] info
396  *   Info structure output buffer.
397  */
398 void
399 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
400 {
401 	struct priv *priv = dev->data->dev_private;
402 	struct mlx5_dev_config *config = &priv->config;
403 	unsigned int max;
404 	char ifname[IF_NAMESIZE];
405 
406 	info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
407 	/* FIXME: we should ask the device for these values. */
408 	info->min_rx_bufsize = 32;
409 	info->max_rx_pktlen = 65536;
410 	/*
411 	 * Since we need one CQ per QP, the limit is the minimum number
412 	 * between the two values.
413 	 */
414 	max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
415 		      priv->device_attr.orig_attr.max_qp);
416 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
417 	if (max >= 65535)
418 		max = 65535;
419 	info->max_rx_queues = max;
420 	info->max_tx_queues = max;
421 	info->max_mac_addrs = RTE_DIM(priv->mac);
422 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
423 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
424 				 info->rx_queue_offload_capa);
425 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
426 	if (mlx5_get_ifname(dev, &ifname) == 0)
427 		info->if_index = if_nametoindex(ifname);
428 	info->reta_size = priv->reta_idx_n ?
429 		priv->reta_idx_n : config->ind_table_max_size;
430 	info->hash_key_size = priv->rss_conf.rss_key_len;
431 	info->speed_capa = priv->link_speed_capa;
432 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
433 }
434 
435 /**
436  * Get supported packet types.
437  *
438  * @param dev
439  *   Pointer to Ethernet device structure.
440  *
441  * @return
442  *   A pointer to the supported Packet types array.
443  */
444 const uint32_t *
445 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
446 {
447 	static const uint32_t ptypes[] = {
448 		/* refers to rxq_cq_to_pkt_type() */
449 		RTE_PTYPE_L2_ETHER,
450 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
451 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
452 		RTE_PTYPE_L4_NONFRAG,
453 		RTE_PTYPE_L4_FRAG,
454 		RTE_PTYPE_L4_TCP,
455 		RTE_PTYPE_L4_UDP,
456 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
457 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
458 		RTE_PTYPE_INNER_L4_NONFRAG,
459 		RTE_PTYPE_INNER_L4_FRAG,
460 		RTE_PTYPE_INNER_L4_TCP,
461 		RTE_PTYPE_INNER_L4_UDP,
462 		RTE_PTYPE_UNKNOWN
463 	};
464 
465 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
466 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
467 		return ptypes;
468 	return NULL;
469 }
470 
471 /**
472  * DPDK callback to retrieve physical link information.
473  *
474  * @param dev
475  *   Pointer to Ethernet device structure.
476  *
477  * @return
478  *   0 on success, a negative errno value otherwise and rte_errno is set.
479  */
480 static int
481 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev)
482 {
483 	struct priv *priv = dev->data->dev_private;
484 	struct ethtool_cmd edata = {
485 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
486 	};
487 	struct ifreq ifr;
488 	struct rte_eth_link dev_link;
489 	int link_speed = 0;
490 	int ret;
491 
492 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
493 	if (ret) {
494 		WARN("port %u ioctl(SIOCGIFFLAGS) failed: %s",
495 		     dev->data->port_id, strerror(rte_errno));
496 		return ret;
497 	}
498 	memset(&dev_link, 0, sizeof(dev_link));
499 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
500 				(ifr.ifr_flags & IFF_RUNNING));
501 	ifr.ifr_data = (void *)&edata;
502 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
503 	if (ret) {
504 		WARN("port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
505 		     dev->data->port_id, strerror(rte_errno));
506 		return ret;
507 	}
508 	link_speed = ethtool_cmd_speed(&edata);
509 	if (link_speed == -1)
510 		dev_link.link_speed = 0;
511 	else
512 		dev_link.link_speed = link_speed;
513 	priv->link_speed_capa = 0;
514 	if (edata.supported & SUPPORTED_Autoneg)
515 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
516 	if (edata.supported & (SUPPORTED_1000baseT_Full |
517 			       SUPPORTED_1000baseKX_Full))
518 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
519 	if (edata.supported & SUPPORTED_10000baseKR_Full)
520 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
521 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
522 			       SUPPORTED_40000baseCR4_Full |
523 			       SUPPORTED_40000baseSR4_Full |
524 			       SUPPORTED_40000baseLR4_Full))
525 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
526 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
527 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
528 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
529 			ETH_LINK_SPEED_FIXED);
530 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
531 		/* Link status changed. */
532 		dev->data->dev_link = dev_link;
533 		return 0;
534 	}
535 	/* Link status is still the same. */
536 	rte_errno = EAGAIN;
537 	return -rte_errno;
538 }
539 
540 /**
541  * Retrieve physical link information (unlocked version using new ioctl).
542  *
543  * @param dev
544  *   Pointer to Ethernet device structure.
545  *
546  * @return
547  *   0 on success, a negative errno value otherwise and rte_errno is set.
548  */
549 static int
550 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev)
551 {
552 	struct priv *priv = dev->data->dev_private;
553 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
554 	struct ifreq ifr;
555 	struct rte_eth_link dev_link;
556 	uint64_t sc;
557 	int ret;
558 
559 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
560 	if (ret) {
561 		WARN("port %u ioctl(SIOCGIFFLAGS) failed: %s",
562 		     dev->data->port_id, strerror(rte_errno));
563 		return ret;
564 	}
565 	memset(&dev_link, 0, sizeof(dev_link));
566 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
567 				(ifr.ifr_flags & IFF_RUNNING));
568 	ifr.ifr_data = (void *)&gcmd;
569 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
570 	if (ret) {
571 		DEBUG("port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
572 		      " failed: %s", dev->data->port_id, strerror(rte_errno));
573 		return ret;
574 	}
575 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
576 
577 	alignas(struct ethtool_link_settings)
578 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
579 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
580 	struct ethtool_link_settings *ecmd = (void *)data;
581 
582 	*ecmd = gcmd;
583 	ifr.ifr_data = (void *)ecmd;
584 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
585 	if (ret) {
586 		DEBUG("port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
587 		      " failed: %s", dev->data->port_id, strerror(rte_errno));
588 		return ret;
589 	}
590 	dev_link.link_speed = ecmd->speed;
591 	sc = ecmd->link_mode_masks[0] |
592 		((uint64_t)ecmd->link_mode_masks[1] << 32);
593 	priv->link_speed_capa = 0;
594 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
595 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
596 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
597 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
598 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
599 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
600 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
601 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
602 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
603 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
604 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
605 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
606 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
607 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
608 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
609 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
610 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
611 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
612 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
613 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
614 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
615 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
616 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
617 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
618 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
619 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
620 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
621 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
622 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
623 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
624 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
625 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
626 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
627 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
628 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
629 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
630 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
631 				  ETH_LINK_SPEED_FIXED);
632 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
633 		/* Link status changed. */
634 		dev->data->dev_link = dev_link;
635 		return 0;
636 	}
637 	/* Link status is still the same. */
638 	rte_errno = EAGAIN;
639 	return -rte_errno;
640 }
641 
642 /**
643  * Enable receiving and transmitting traffic.
644  *
645  * @param dev
646  *   Pointer to Ethernet device.
647  */
648 static void
649 mlx5_link_start(struct rte_eth_dev *dev)
650 {
651 	struct priv *priv = dev->data->dev_private;
652 	int ret;
653 
654 	dev->tx_pkt_burst = mlx5_select_tx_function(dev);
655 	dev->rx_pkt_burst = mlx5_select_rx_function(dev);
656 	ret = mlx5_traffic_enable(dev);
657 	if (ret) {
658 		ERROR("port %u error occurred while configuring control flows:"
659 		      " %s", dev->data->port_id, strerror(rte_errno));
660 		return;
661 	}
662 	ret = mlx5_flow_start(dev, &priv->flows);
663 	if (ret)
664 		ERROR("port %u error occurred while configuring flows: %s",
665 		       dev->data->port_id, strerror(rte_errno));
666 }
667 
668 /**
669  * Disable receiving and transmitting traffic.
670  *
671  * @param dev
672  *   Pointer to Ethernet device.
673  */
674 static void
675 mlx5_link_stop(struct rte_eth_dev *dev)
676 {
677 	struct priv *priv = dev->data->dev_private;
678 
679 	mlx5_flow_stop(dev, &priv->flows);
680 	mlx5_traffic_disable(dev);
681 	dev->rx_pkt_burst = removed_rx_burst;
682 	dev->tx_pkt_burst = removed_tx_burst;
683 }
684 
685 /**
686  * Querying the link status till it changes to the desired state.
687  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
688  *
689  * @param dev
690  *   Pointer to Ethernet device.
691  * @param status
692  *   Link desired status.
693  *
694  * @return
695  *   0 on success, a negative errno value otherwise and rte_errno is set.
696  */
697 int
698 mlx5_force_link_status_change(struct rte_eth_dev *dev, int status)
699 {
700 	int try = 0;
701 
702 	while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
703 		mlx5_link_update(dev, 0);
704 		if (dev->data->dev_link.link_status == status)
705 			return 0;
706 		try++;
707 		sleep(1);
708 	}
709 	rte_errno = EAGAIN;
710 	return -rte_errno;
711 }
712 
713 /**
714  * DPDK callback to retrieve physical link information.
715  *
716  * @param dev
717  *   Pointer to Ethernet device structure.
718  * @param wait_to_complete
719  *   Wait for request completion (ignored).
720  *
721  * @return
722  *   0 on success, a negative errno value otherwise and rte_errno is set.
723  */
724 int
725 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete __rte_unused)
726 {
727 	struct utsname utsname;
728 	int ver[3];
729 	int ret;
730 	struct rte_eth_link dev_link = dev->data->dev_link;
731 
732 	if (uname(&utsname) == -1 ||
733 	    sscanf(utsname.release, "%d.%d.%d",
734 		   &ver[0], &ver[1], &ver[2]) != 3 ||
735 	    KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
736 		ret = mlx5_link_update_unlocked_gset(dev);
737 	else
738 		ret = mlx5_link_update_unlocked_gs(dev);
739 	if (ret)
740 		return ret;
741 	/* If lsc interrupt is disabled, should always be ready for traffic. */
742 	if (!dev->data->dev_conf.intr_conf.lsc) {
743 		mlx5_link_start(dev);
744 		return 0;
745 	}
746 	/* Re-select burst callbacks only if link status has been changed. */
747 	if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
748 		if (dev->data->dev_link.link_status == ETH_LINK_UP)
749 			mlx5_link_start(dev);
750 		else
751 			mlx5_link_stop(dev);
752 	}
753 	return 0;
754 }
755 
756 /**
757  * DPDK callback to change the MTU.
758  *
759  * @param dev
760  *   Pointer to Ethernet device structure.
761  * @param in_mtu
762  *   New MTU.
763  *
764  * @return
765  *   0 on success, a negative errno value otherwise and rte_errno is set.
766  */
767 int
768 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
769 {
770 	struct priv *priv = dev->data->dev_private;
771 	uint16_t kern_mtu = 0;
772 	int ret;
773 
774 	ret = mlx5_get_mtu(dev, &kern_mtu);
775 	if (ret)
776 		return ret;
777 	/* Set kernel interface MTU first. */
778 	ret = mlx5_set_mtu(dev, mtu);
779 	if (ret)
780 		return ret;
781 	ret = mlx5_get_mtu(dev, &kern_mtu);
782 	if (ret)
783 		return ret;
784 	if (kern_mtu == mtu) {
785 		priv->mtu = mtu;
786 		DEBUG("port %u adapter MTU set to %u", dev->data->port_id, mtu);
787 		return 0;
788 	}
789 	rte_errno = EAGAIN;
790 	return -rte_errno;
791 }
792 
793 /**
794  * DPDK callback to get flow control status.
795  *
796  * @param dev
797  *   Pointer to Ethernet device structure.
798  * @param[out] fc_conf
799  *   Flow control output buffer.
800  *
801  * @return
802  *   0 on success, a negative errno value otherwise and rte_errno is set.
803  */
804 int
805 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
806 {
807 	struct ifreq ifr;
808 	struct ethtool_pauseparam ethpause = {
809 		.cmd = ETHTOOL_GPAUSEPARAM
810 	};
811 	int ret;
812 
813 	ifr.ifr_data = (void *)&ethpause;
814 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
815 	if (ret) {
816 		WARN("port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
817 		     " %s", dev->data->port_id, strerror(rte_errno));
818 		return ret;
819 	}
820 	fc_conf->autoneg = ethpause.autoneg;
821 	if (ethpause.rx_pause && ethpause.tx_pause)
822 		fc_conf->mode = RTE_FC_FULL;
823 	else if (ethpause.rx_pause)
824 		fc_conf->mode = RTE_FC_RX_PAUSE;
825 	else if (ethpause.tx_pause)
826 		fc_conf->mode = RTE_FC_TX_PAUSE;
827 	else
828 		fc_conf->mode = RTE_FC_NONE;
829 	return 0;
830 }
831 
832 /**
833  * DPDK callback to modify flow control parameters.
834  *
835  * @param dev
836  *   Pointer to Ethernet device structure.
837  * @param[in] fc_conf
838  *   Flow control parameters.
839  *
840  * @return
841  *   0 on success, a negative errno value otherwise and rte_errno is set.
842  */
843 int
844 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
845 {
846 	struct ifreq ifr;
847 	struct ethtool_pauseparam ethpause = {
848 		.cmd = ETHTOOL_SPAUSEPARAM
849 	};
850 	int ret;
851 
852 	ifr.ifr_data = (void *)&ethpause;
853 	ethpause.autoneg = fc_conf->autoneg;
854 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
855 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
856 		ethpause.rx_pause = 1;
857 	else
858 		ethpause.rx_pause = 0;
859 
860 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
861 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
862 		ethpause.tx_pause = 1;
863 	else
864 		ethpause.tx_pause = 0;
865 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
866 	if (ret) {
867 		WARN("port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
868 		     " failed: %s", dev->data->port_id, strerror(rte_errno));
869 		return ret;
870 	}
871 	return 0;
872 }
873 
874 /**
875  * Get PCI information from struct ibv_device.
876  *
877  * @param device
878  *   Pointer to Ethernet device structure.
879  * @param[out] pci_addr
880  *   PCI bus address output buffer.
881  *
882  * @return
883  *   0 on success, a negative errno value otherwise and rte_errno is set.
884  */
885 int
886 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
887 			    struct rte_pci_addr *pci_addr)
888 {
889 	FILE *file;
890 	char line[32];
891 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
892 
893 	file = fopen(path, "rb");
894 	if (file == NULL) {
895 		rte_errno = errno;
896 		return -rte_errno;
897 	}
898 	while (fgets(line, sizeof(line), file) == line) {
899 		size_t len = strlen(line);
900 		int ret;
901 
902 		/* Truncate long lines. */
903 		if (len == (sizeof(line) - 1))
904 			while (line[(len - 1)] != '\n') {
905 				ret = fgetc(file);
906 				if (ret == EOF)
907 					break;
908 				line[(len - 1)] = ret;
909 			}
910 		/* Extract information. */
911 		if (sscanf(line,
912 			   "PCI_SLOT_NAME="
913 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
914 			   &pci_addr->domain,
915 			   &pci_addr->bus,
916 			   &pci_addr->devid,
917 			   &pci_addr->function) == 4) {
918 			ret = 0;
919 			break;
920 		}
921 	}
922 	fclose(file);
923 	return 0;
924 }
925 
926 /**
927  * Update the link status.
928  *
929  * @param dev
930  *   Pointer to Ethernet device.
931  *
932  * @return
933  *   Zero if the callback process can be called immediately, negative errno
934  *   value otherwise and rte_errno is set.
935  */
936 static int
937 mlx5_link_status_update(struct rte_eth_dev *dev)
938 {
939 	struct priv *priv = dev->data->dev_private;
940 	struct rte_eth_link *link = &dev->data->dev_link;
941 	int ret;
942 
943 	ret = mlx5_link_update(dev, 0);
944 	if (ret)
945 		return ret;
946 	if (((link->link_speed == 0) && link->link_status) ||
947 		((link->link_speed != 0) && !link->link_status)) {
948 		/*
949 		 * Inconsistent status. Event likely occurred before the
950 		 * kernel netdevice exposes the new status.
951 		 */
952 		if (!priv->pending_alarm) {
953 			priv->pending_alarm = 1;
954 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
955 					  mlx5_dev_link_status_handler,
956 					  priv->dev);
957 		}
958 		return 1;
959 	} else if (unlikely(priv->pending_alarm)) {
960 		/* Link interrupt occurred while alarm is already scheduled. */
961 		priv->pending_alarm = 0;
962 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
963 	}
964 	return 0;
965 }
966 
967 /**
968  * Device status handler.
969  *
970  * @param dev
971  *   Pointer to Ethernet device.
972  * @param events
973  *   Pointer to event flags holder.
974  *
975  * @return
976  *   Events bitmap of callback process which can be called immediately.
977  */
978 static uint32_t
979 mlx5_dev_status_handler(struct rte_eth_dev *dev)
980 {
981 	struct priv *priv = dev->data->dev_private;
982 	struct ibv_async_event event;
983 	uint32_t ret = 0;
984 
985 	/* Read all message and acknowledge them. */
986 	for (;;) {
987 		if (mlx5_glue->get_async_event(priv->ctx, &event))
988 			break;
989 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
990 			event.event_type == IBV_EVENT_PORT_ERR) &&
991 			(dev->data->dev_conf.intr_conf.lsc == 1))
992 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
993 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
994 			dev->data->dev_conf.intr_conf.rmv == 1)
995 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
996 		else
997 			DEBUG("port %u event type %d on not handled",
998 			      dev->data->port_id, event.event_type);
999 		mlx5_glue->ack_async_event(&event);
1000 	}
1001 	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1002 		if (mlx5_link_status_update(dev))
1003 			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1004 	return ret;
1005 }
1006 
1007 /**
1008  * Handle delayed link status event.
1009  *
1010  * @param arg
1011  *   Registered argument.
1012  */
1013 void
1014 mlx5_dev_link_status_handler(void *arg)
1015 {
1016 	struct rte_eth_dev *dev = arg;
1017 	struct priv *priv = dev->data->dev_private;
1018 	int ret;
1019 
1020 	priv->pending_alarm = 0;
1021 	ret = mlx5_link_status_update(dev);
1022 	if (!ret)
1023 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1024 }
1025 
1026 /**
1027  * Handle interrupts from the NIC.
1028  *
1029  * @param[in] intr_handle
1030  *   Interrupt handler.
1031  * @param cb_arg
1032  *   Callback argument.
1033  */
1034 void
1035 mlx5_dev_interrupt_handler(void *cb_arg)
1036 {
1037 	struct rte_eth_dev *dev = cb_arg;
1038 	uint32_t events;
1039 
1040 	events = mlx5_dev_status_handler(dev);
1041 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1042 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1043 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1044 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1045 }
1046 
1047 /**
1048  * Handle interrupts from the socket.
1049  *
1050  * @param cb_arg
1051  *   Callback argument.
1052  */
1053 static void
1054 mlx5_dev_handler_socket(void *cb_arg)
1055 {
1056 	struct rte_eth_dev *dev = cb_arg;
1057 
1058 	mlx5_socket_handle(dev);
1059 }
1060 
1061 /**
1062  * Uninstall interrupt handler.
1063  *
1064  * @param dev
1065  *   Pointer to Ethernet device.
1066  */
1067 void
1068 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1069 {
1070 	struct priv *priv = dev->data->dev_private;
1071 
1072 	if (dev->data->dev_conf.intr_conf.lsc ||
1073 	    dev->data->dev_conf.intr_conf.rmv)
1074 		rte_intr_callback_unregister(&priv->intr_handle,
1075 					     mlx5_dev_interrupt_handler, dev);
1076 	if (priv->primary_socket)
1077 		rte_intr_callback_unregister(&priv->intr_handle_socket,
1078 					     mlx5_dev_handler_socket, dev);
1079 	if (priv->pending_alarm) {
1080 		priv->pending_alarm = 0;
1081 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1082 	}
1083 	priv->intr_handle.fd = 0;
1084 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1085 	priv->intr_handle_socket.fd = 0;
1086 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1087 }
1088 
1089 /**
1090  * Install interrupt handler.
1091  *
1092  * @param dev
1093  *   Pointer to Ethernet device.
1094  */
1095 void
1096 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1097 {
1098 	struct priv *priv = dev->data->dev_private;
1099 	int ret;
1100 	int flags;
1101 
1102 	assert(priv->ctx->async_fd > 0);
1103 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1104 	ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1105 	if (ret) {
1106 		INFO("port %u failed to change file descriptor async event"
1107 		     " queue", dev->data->port_id);
1108 		dev->data->dev_conf.intr_conf.lsc = 0;
1109 		dev->data->dev_conf.intr_conf.rmv = 0;
1110 	}
1111 	if (dev->data->dev_conf.intr_conf.lsc ||
1112 	    dev->data->dev_conf.intr_conf.rmv) {
1113 		priv->intr_handle.fd = priv->ctx->async_fd;
1114 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1115 		rte_intr_callback_register(&priv->intr_handle,
1116 					   mlx5_dev_interrupt_handler, dev);
1117 	}
1118 	ret = mlx5_socket_init(dev);
1119 	if (ret)
1120 		ERROR("port %u cannot initialise socket: %s",
1121 		      dev->data->port_id, strerror(rte_errno));
1122 	else if (priv->primary_socket) {
1123 		priv->intr_handle_socket.fd = priv->primary_socket;
1124 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1125 		rte_intr_callback_register(&priv->intr_handle_socket,
1126 					   mlx5_dev_handler_socket, dev);
1127 	}
1128 }
1129 
1130 /**
1131  * DPDK callback to bring the link DOWN.
1132  *
1133  * @param dev
1134  *   Pointer to Ethernet device structure.
1135  *
1136  * @return
1137  *   0 on success, a negative errno value otherwise and rte_errno is set.
1138  */
1139 int
1140 mlx5_set_link_down(struct rte_eth_dev *dev)
1141 {
1142 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1143 }
1144 
1145 /**
1146  * DPDK callback to bring the link UP.
1147  *
1148  * @param dev
1149  *   Pointer to Ethernet device structure.
1150  *
1151  * @return
1152  *   0 on success, a negative errno value otherwise and rte_errno is set.
1153  */
1154 int
1155 mlx5_set_link_up(struct rte_eth_dev *dev)
1156 {
1157 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1158 }
1159 
1160 /**
1161  * Configure the TX function to use.
1162  *
1163  * @param dev
1164  *   Pointer to private data structure.
1165  *
1166  * @return
1167  *   Pointer to selected Tx burst function.
1168  */
1169 eth_tx_burst_t
1170 mlx5_select_tx_function(struct rte_eth_dev *dev)
1171 {
1172 	struct priv *priv = dev->data->dev_private;
1173 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1174 	struct mlx5_dev_config *config = &priv->config;
1175 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1176 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1177 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1178 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1179 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1180 
1181 	assert(priv != NULL);
1182 	/* Select appropriate TX function. */
1183 	if (vlan_insert || tso)
1184 		return tx_pkt_burst;
1185 	if (config->mps == MLX5_MPW_ENHANCED) {
1186 		if (mlx5_check_vec_tx_support(dev) > 0) {
1187 			if (mlx5_check_raw_vec_tx_support(dev) > 0)
1188 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1189 			else
1190 				tx_pkt_burst = mlx5_tx_burst_vec;
1191 			DEBUG("port %u selected enhanced MPW Tx vectorized"
1192 			      " function", dev->data->port_id);
1193 		} else {
1194 			tx_pkt_burst = mlx5_tx_burst_empw;
1195 			DEBUG("port %u selected enhanced MPW Tx function",
1196 			      dev->data->port_id);
1197 		}
1198 	} else if (config->mps && (config->txq_inline > 0)) {
1199 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1200 		DEBUG("port %u selected MPW inline Tx function",
1201 		      dev->data->port_id);
1202 	} else if (config->mps) {
1203 		tx_pkt_burst = mlx5_tx_burst_mpw;
1204 		DEBUG("port %u selected MPW Tx function", dev->data->port_id);
1205 	}
1206 	return tx_pkt_burst;
1207 }
1208 
1209 /**
1210  * Configure the RX function to use.
1211  *
1212  * @param dev
1213  *   Pointer to private data structure.
1214  *
1215  * @return
1216  *   Pointer to selected Rx burst function.
1217  */
1218 eth_rx_burst_t
1219 mlx5_select_rx_function(struct rte_eth_dev *dev)
1220 {
1221 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1222 
1223 	assert(dev != NULL);
1224 	if (mlx5_check_vec_rx_support(dev) > 0) {
1225 		rx_pkt_burst = mlx5_rx_burst_vec;
1226 		DEBUG("port %u selected Rx vectorized function",
1227 		      dev->data->port_id);
1228 	}
1229 	return rx_pkt_burst;
1230 }
1231 
1232 /**
1233  * Check if mlx5 device was removed.
1234  *
1235  * @param dev
1236  *   Pointer to Ethernet device structure.
1237  *
1238  * @return
1239  *   1 when device is removed, otherwise 0.
1240  */
1241 int
1242 mlx5_is_removed(struct rte_eth_dev *dev)
1243 {
1244 	struct ibv_device_attr device_attr;
1245 	struct priv *priv = dev->data->dev_private;
1246 
1247 	if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1248 		return 1;
1249 	return 0;
1250 }
1251