xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 56f08e1671f99a15cf6b00027c2f7d81d69c4f5f)
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] priv
99  *   Pointer to private structure.
100  * @param[out] ifname
101  *   Interface name output buffer.
102  *
103  * @return
104  *   0 on success, -1 on failure and errno is set.
105  */
106 int
107 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
108 {
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 			return -1;
121 	}
122 	while ((dent = readdir(dir)) != NULL) {
123 		char *name = dent->d_name;
124 		FILE *file;
125 		unsigned int dev_port;
126 		int r;
127 
128 		if ((name[0] == '.') &&
129 		    ((name[1] == '\0') ||
130 		     ((name[1] == '.') && (name[2] == '\0'))))
131 			continue;
132 
133 		MKSTR(path, "%s/device/net/%s/%s",
134 		      priv->ibdev_path, name,
135 		      (dev_type ? "dev_id" : "dev_port"));
136 
137 		file = fopen(path, "rb");
138 		if (file == NULL) {
139 			if (errno != ENOENT)
140 				continue;
141 			/*
142 			 * Switch to dev_id when dev_port does not exist as
143 			 * is the case with Linux kernel versions < 3.15.
144 			 */
145 try_dev_id:
146 			match[0] = '\0';
147 			if (dev_type)
148 				break;
149 			dev_type = 1;
150 			dev_port_prev = ~0u;
151 			rewinddir(dir);
152 			continue;
153 		}
154 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
155 		fclose(file);
156 		if (r != 1)
157 			continue;
158 		/*
159 		 * Switch to dev_id when dev_port returns the same value for
160 		 * all ports. May happen when using a MOFED release older than
161 		 * 3.0 with a Linux kernel >= 3.15.
162 		 */
163 		if (dev_port == dev_port_prev)
164 			goto try_dev_id;
165 		dev_port_prev = dev_port;
166 		if (dev_port == (priv->port - 1u))
167 			snprintf(match, sizeof(match), "%s", name);
168 	}
169 	closedir(dir);
170 	if (match[0] == '\0')
171 		return -1;
172 	strncpy(*ifname, match, sizeof(*ifname));
173 	return 0;
174 }
175 
176 /**
177  * Perform ifreq ioctl() on associated Ethernet device.
178  *
179  * @param[in] priv
180  *   Pointer to private structure.
181  * @param req
182  *   Request number to pass to ioctl().
183  * @param[out] ifr
184  *   Interface request structure output buffer.
185  *
186  * @return
187  *   0 on success, -1 on failure and errno is set.
188  */
189 int
190 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
191 {
192 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
193 	int ret = -1;
194 
195 	if (sock == -1)
196 		return ret;
197 	if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
198 		ret = ioctl(sock, req, ifr);
199 	close(sock);
200 	return ret;
201 }
202 
203 /**
204  * Get device MTU.
205  *
206  * @param priv
207  *   Pointer to private structure.
208  * @param[out] mtu
209  *   MTU value output buffer.
210  *
211  * @return
212  *   0 on success, -1 on failure and errno is set.
213  */
214 int
215 priv_get_mtu(struct priv *priv, uint16_t *mtu)
216 {
217 	struct ifreq request;
218 	int ret = priv_ifreq(priv, SIOCGIFMTU, &request);
219 
220 	if (ret)
221 		return ret;
222 	*mtu = request.ifr_mtu;
223 	return 0;
224 }
225 
226 /**
227  * Set device MTU.
228  *
229  * @param priv
230  *   Pointer to private structure.
231  * @param mtu
232  *   MTU value to set.
233  *
234  * @return
235  *   0 on success, -1 on failure and errno is set.
236  */
237 static int
238 priv_set_mtu(struct priv *priv, uint16_t mtu)
239 {
240 	struct ifreq request = { .ifr_mtu = mtu, };
241 
242 	return priv_ifreq(priv, SIOCSIFMTU, &request);
243 }
244 
245 /**
246  * Set device flags.
247  *
248  * @param priv
249  *   Pointer to private structure.
250  * @param keep
251  *   Bitmask for flags that must remain untouched.
252  * @param flags
253  *   Bitmask for flags to modify.
254  *
255  * @return
256  *   0 on success, -1 on failure and errno is set.
257  */
258 int
259 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
260 {
261 	struct ifreq request;
262 	int ret = priv_ifreq(priv, SIOCGIFFLAGS, &request);
263 
264 	if (ret)
265 		return ret;
266 	request.ifr_flags &= keep;
267 	request.ifr_flags |= flags & ~keep;
268 	return priv_ifreq(priv, SIOCSIFFLAGS, &request);
269 }
270 
271 /**
272  * Ethernet device configuration.
273  *
274  * Prepare the driver for a given number of TX and RX queues.
275  *
276  * @param dev
277  *   Pointer to Ethernet device structure.
278  *
279  * @return
280  *   0 on success, errno value on failure.
281  */
282 static int
283 dev_configure(struct rte_eth_dev *dev)
284 {
285 	struct priv *priv = dev->data->dev_private;
286 	unsigned int rxqs_n = dev->data->nb_rx_queues;
287 	unsigned int txqs_n = dev->data->nb_tx_queues;
288 	unsigned int i;
289 	unsigned int j;
290 	unsigned int reta_idx_n;
291 	const uint8_t use_app_rss_key =
292 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
293 	uint64_t supp_tx_offloads = mlx5_priv_get_tx_port_offloads(priv);
294 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
295 	uint64_t supp_rx_offloads =
296 		(mlx5_priv_get_rx_port_offloads(priv) |
297 		 mlx5_priv_get_rx_queue_offloads(priv));
298 	uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
299 
300 	if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
301 		ERROR("Some Tx offloads are not supported "
302 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
303 		      tx_offloads, supp_tx_offloads);
304 		return ENOTSUP;
305 	}
306 	if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
307 		ERROR("Some Rx offloads are not supported "
308 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
309 		      rx_offloads, supp_rx_offloads);
310 		return ENOTSUP;
311 	}
312 	if (use_app_rss_key &&
313 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
314 	     rss_hash_default_key_len)) {
315 		/* MLX5 RSS only support 40bytes key. */
316 		return EINVAL;
317 	}
318 	priv->rss_conf.rss_key =
319 		rte_realloc(priv->rss_conf.rss_key,
320 			    rss_hash_default_key_len, 0);
321 	if (!priv->rss_conf.rss_key) {
322 		ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
323 		return ENOMEM;
324 	}
325 	memcpy(priv->rss_conf.rss_key,
326 	       use_app_rss_key ?
327 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
328 	       rss_hash_default_key,
329 	       rss_hash_default_key_len);
330 	priv->rss_conf.rss_key_len = rss_hash_default_key_len;
331 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
332 	priv->rxqs = (void *)dev->data->rx_queues;
333 	priv->txqs = (void *)dev->data->tx_queues;
334 	if (txqs_n != priv->txqs_n) {
335 		INFO("%p: TX queues number update: %u -> %u",
336 		     (void *)dev, priv->txqs_n, txqs_n);
337 		priv->txqs_n = txqs_n;
338 	}
339 	if (rxqs_n > priv->config.ind_table_max_size) {
340 		ERROR("cannot handle this many RX queues (%u)", rxqs_n);
341 		return EINVAL;
342 	}
343 	if (rxqs_n == priv->rxqs_n)
344 		return 0;
345 	INFO("%p: RX queues number update: %u -> %u",
346 	     (void *)dev, priv->rxqs_n, rxqs_n);
347 	priv->rxqs_n = rxqs_n;
348 	/* If the requested number of RX queues is not a power of two, use the
349 	 * maximum indirection table size for better balancing.
350 	 * The result is always rounded to the next power of two. */
351 	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
352 				     priv->config.ind_table_max_size :
353 				     rxqs_n));
354 	if (priv_rss_reta_index_resize(priv, reta_idx_n))
355 		return ENOMEM;
356 	/* When the number of RX queues is not a power of two, the remaining
357 	 * table entries are padded with reused WQs and hashes are not spread
358 	 * uniformly. */
359 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
360 		(*priv->reta_idx)[i] = j;
361 		if (++j == rxqs_n)
362 			j = 0;
363 	}
364 	return 0;
365 }
366 
367 /**
368  * DPDK callback for Ethernet device configuration.
369  *
370  * @param dev
371  *   Pointer to Ethernet device structure.
372  *
373  * @return
374  *   0 on success, negative errno value on failure.
375  */
376 int
377 mlx5_dev_configure(struct rte_eth_dev *dev)
378 {
379 	struct priv *priv = dev->data->dev_private;
380 	int ret;
381 
382 	priv_lock(priv);
383 	ret = dev_configure(dev);
384 	assert(ret >= 0);
385 	priv_unlock(priv);
386 	return -ret;
387 }
388 
389 /**
390  * DPDK callback to get information about the device.
391  *
392  * @param dev
393  *   Pointer to Ethernet device structure.
394  * @param[out] info
395  *   Info structure output buffer.
396  */
397 void
398 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
399 {
400 	struct priv *priv = dev->data->dev_private;
401 	struct mlx5_dev_config *config = &priv->config;
402 	unsigned int max;
403 	char ifname[IF_NAMESIZE];
404 
405 	info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
406 
407 	priv_lock(priv);
408 	/* FIXME: we should ask the device for these values. */
409 	info->min_rx_bufsize = 32;
410 	info->max_rx_pktlen = 65536;
411 	/*
412 	 * Since we need one CQ per QP, the limit is the minimum number
413 	 * between the two values.
414 	 */
415 	max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
416 		      priv->device_attr.orig_attr.max_qp);
417 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
418 	if (max >= 65535)
419 		max = 65535;
420 	info->max_rx_queues = max;
421 	info->max_tx_queues = max;
422 	info->max_mac_addrs = RTE_DIM(priv->mac);
423 	info->rx_queue_offload_capa =
424 		mlx5_priv_get_rx_queue_offloads(priv);
425 	info->rx_offload_capa = (mlx5_priv_get_rx_port_offloads(priv) |
426 				 info->rx_queue_offload_capa);
427 	info->tx_offload_capa = mlx5_priv_get_tx_port_offloads(priv);
428 	if (priv_get_ifname(priv, &ifname) == 0)
429 		info->if_index = if_nametoindex(ifname);
430 	info->reta_size = priv->reta_idx_n ?
431 		priv->reta_idx_n : config->ind_table_max_size;
432 	info->hash_key_size = priv->rss_conf.rss_key_len;
433 	info->speed_capa = priv->link_speed_capa;
434 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
435 	priv_unlock(priv);
436 }
437 
438 const uint32_t *
439 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
440 {
441 	static const uint32_t ptypes[] = {
442 		/* refers to rxq_cq_to_pkt_type() */
443 		RTE_PTYPE_L2_ETHER,
444 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
445 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
446 		RTE_PTYPE_L4_NONFRAG,
447 		RTE_PTYPE_L4_FRAG,
448 		RTE_PTYPE_L4_TCP,
449 		RTE_PTYPE_L4_UDP,
450 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
451 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
452 		RTE_PTYPE_INNER_L4_NONFRAG,
453 		RTE_PTYPE_INNER_L4_FRAG,
454 		RTE_PTYPE_INNER_L4_TCP,
455 		RTE_PTYPE_INNER_L4_UDP,
456 		RTE_PTYPE_UNKNOWN
457 	};
458 
459 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
460 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
461 		return ptypes;
462 	return NULL;
463 }
464 
465 /**
466  * DPDK callback to retrieve physical link information.
467  *
468  * @param dev
469  *   Pointer to Ethernet device structure.
470  */
471 static int
472 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev)
473 {
474 	struct priv *priv = dev->data->dev_private;
475 	struct ethtool_cmd edata = {
476 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
477 	};
478 	struct ifreq ifr;
479 	struct rte_eth_link dev_link;
480 	int link_speed = 0;
481 
482 	/* priv_lock() is not taken to allow concurrent calls. */
483 
484 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
485 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
486 		return -1;
487 	}
488 	memset(&dev_link, 0, sizeof(dev_link));
489 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
490 				(ifr.ifr_flags & IFF_RUNNING));
491 	ifr.ifr_data = (void *)&edata;
492 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
493 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
494 		     strerror(errno));
495 		return -1;
496 	}
497 	link_speed = ethtool_cmd_speed(&edata);
498 	if (link_speed == -1)
499 		dev_link.link_speed = 0;
500 	else
501 		dev_link.link_speed = link_speed;
502 	priv->link_speed_capa = 0;
503 	if (edata.supported & SUPPORTED_Autoneg)
504 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
505 	if (edata.supported & (SUPPORTED_1000baseT_Full |
506 			       SUPPORTED_1000baseKX_Full))
507 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
508 	if (edata.supported & SUPPORTED_10000baseKR_Full)
509 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
510 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
511 			       SUPPORTED_40000baseCR4_Full |
512 			       SUPPORTED_40000baseSR4_Full |
513 			       SUPPORTED_40000baseLR4_Full))
514 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
515 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
516 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
517 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
518 			ETH_LINK_SPEED_FIXED);
519 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
520 		/* Link status changed. */
521 		dev->data->dev_link = dev_link;
522 		return 0;
523 	}
524 	/* Link status is still the same. */
525 	return -1;
526 }
527 
528 /**
529  * Retrieve physical link information (unlocked version using new ioctl).
530  *
531  * @param dev
532  *   Pointer to Ethernet device structure.
533  */
534 static int
535 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev)
536 {
537 	struct priv *priv = dev->data->dev_private;
538 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
539 	struct ifreq ifr;
540 	struct rte_eth_link dev_link;
541 	uint64_t sc;
542 
543 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
544 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
545 		return -1;
546 	}
547 	memset(&dev_link, 0, sizeof(dev_link));
548 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
549 				(ifr.ifr_flags & IFF_RUNNING));
550 	ifr.ifr_data = (void *)&gcmd;
551 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
552 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
553 		      strerror(errno));
554 		return -1;
555 	}
556 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
557 
558 	alignas(struct ethtool_link_settings)
559 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
560 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
561 	struct ethtool_link_settings *ecmd = (void *)data;
562 
563 	*ecmd = gcmd;
564 	ifr.ifr_data = (void *)ecmd;
565 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
566 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
567 		      strerror(errno));
568 		return -1;
569 	}
570 	dev_link.link_speed = ecmd->speed;
571 	sc = ecmd->link_mode_masks[0] |
572 		((uint64_t)ecmd->link_mode_masks[1] << 32);
573 	priv->link_speed_capa = 0;
574 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
575 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
576 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
577 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
578 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
579 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
580 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
581 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
582 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
583 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
584 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
585 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
586 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
587 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
588 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
589 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
590 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
591 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
592 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
593 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
594 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
595 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
596 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
597 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
598 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
599 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
600 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
601 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
602 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
603 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
604 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
605 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
606 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
607 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
608 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
609 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
610 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
611 				  ETH_LINK_SPEED_FIXED);
612 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
613 		/* Link status changed. */
614 		dev->data->dev_link = dev_link;
615 		return 0;
616 	}
617 	/* Link status is still the same. */
618 	return -1;
619 }
620 
621 /**
622  * Enable receiving and transmitting traffic.
623  *
624  * @param priv
625  *   Pointer to private structure.
626  */
627 static void
628 priv_link_start(struct priv *priv)
629 {
630 	struct rte_eth_dev *dev = priv->dev;
631 	int err;
632 
633 	dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
634 	dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
635 	err = priv_dev_traffic_enable(priv, dev);
636 	if (err)
637 		ERROR("%p: error occurred while configuring control flows: %s",
638 		      (void *)priv, strerror(err));
639 	err = priv_flow_start(priv, &priv->flows);
640 	if (err)
641 		ERROR("%p: error occurred while configuring flows: %s",
642 		      (void *)priv, strerror(err));
643 }
644 
645 /**
646  * Disable receiving and transmitting traffic.
647  *
648  * @param priv
649  *   Pointer to private structure.
650  */
651 static void
652 priv_link_stop(struct priv *priv)
653 {
654 	struct rte_eth_dev *dev = priv->dev;
655 
656 	priv_flow_stop(priv, &priv->flows);
657 	priv_dev_traffic_disable(priv, dev);
658 	dev->rx_pkt_burst = removed_rx_burst;
659 	dev->tx_pkt_burst = removed_tx_burst;
660 }
661 
662 /**
663  * Retrieve physical link information and update rx/tx_pkt_burst callbacks
664  * accordingly.
665  *
666  * @param priv
667  *   Pointer to private structure.
668  * @param wait_to_complete
669  *   Wait for request completion (ignored).
670  */
671 int
672 priv_link_update(struct priv *priv, int wait_to_complete __rte_unused)
673 {
674 	struct rte_eth_dev *dev = priv->dev;
675 	struct utsname utsname;
676 	int ver[3];
677 	int ret;
678 	struct rte_eth_link dev_link = dev->data->dev_link;
679 
680 	if (uname(&utsname) == -1 ||
681 	    sscanf(utsname.release, "%d.%d.%d",
682 		   &ver[0], &ver[1], &ver[2]) != 3 ||
683 	    KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
684 		ret = mlx5_link_update_unlocked_gset(dev);
685 	else
686 		ret = mlx5_link_update_unlocked_gs(dev);
687 	/* If lsc interrupt is disabled, should always be ready for traffic. */
688 	if (!dev->data->dev_conf.intr_conf.lsc) {
689 		priv_link_start(priv);
690 		return ret;
691 	}
692 	/* Re-select burst callbacks only if link status has been changed. */
693 	if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
694 		if (dev->data->dev_link.link_status == ETH_LINK_UP)
695 			priv_link_start(priv);
696 		else
697 			priv_link_stop(priv);
698 	}
699 	return ret;
700 }
701 
702 /**
703  * Querying the link status till it changes to the desired state.
704  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
705  *
706  * @param priv
707  *   Pointer to private structure.
708  * @param status
709  *   Link desired status.
710  *
711  * @return
712  *   0 on success, negative errno value on failure.
713  */
714 int
715 priv_force_link_status_change(struct priv *priv, int status)
716 {
717 	int try = 0;
718 
719 	while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
720 		priv_link_update(priv, 0);
721 		if (priv->dev->data->dev_link.link_status == status)
722 			return 0;
723 		try++;
724 		sleep(1);
725 	}
726 	return -EAGAIN;
727 }
728 
729 /**
730  * DPDK callback to retrieve physical link information.
731  *
732  * @param dev
733  *   Pointer to Ethernet device structure.
734  * @param wait_to_complete
735  *   Wait for request completion (ignored).
736  */
737 int
738 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete __rte_unused)
739 {
740 	struct priv *priv = dev->data->dev_private;
741 	int ret;
742 
743 	priv_lock(priv);
744 	ret = priv_link_update(priv, wait_to_complete);
745 	priv_unlock(priv);
746 	return ret;
747 }
748 
749 /**
750  * DPDK callback to change the MTU.
751  *
752  * @param dev
753  *   Pointer to Ethernet device structure.
754  * @param in_mtu
755  *   New MTU.
756  *
757  * @return
758  *   0 on success, negative errno value on failure.
759  */
760 int
761 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
762 {
763 	struct priv *priv = dev->data->dev_private;
764 	uint16_t kern_mtu;
765 	int ret = 0;
766 
767 	priv_lock(priv);
768 	ret = priv_get_mtu(priv, &kern_mtu);
769 	if (ret)
770 		goto out;
771 	/* Set kernel interface MTU first. */
772 	ret = priv_set_mtu(priv, mtu);
773 	if (ret)
774 		goto out;
775 	ret = priv_get_mtu(priv, &kern_mtu);
776 	if (ret)
777 		goto out;
778 	if (kern_mtu == mtu) {
779 		priv->mtu = mtu;
780 		DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
781 	}
782 	priv_unlock(priv);
783 	return 0;
784 out:
785 	ret = errno;
786 	WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
787 	     strerror(ret));
788 	priv_unlock(priv);
789 	assert(ret >= 0);
790 	return -ret;
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, negative errno value on failure.
803  */
804 int
805 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
806 {
807 	struct priv *priv = dev->data->dev_private;
808 	struct ifreq ifr;
809 	struct ethtool_pauseparam ethpause = {
810 		.cmd = ETHTOOL_GPAUSEPARAM
811 	};
812 	int ret;
813 
814 	ifr.ifr_data = (void *)&ethpause;
815 	priv_lock(priv);
816 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
817 		ret = errno;
818 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
819 		     " failed: %s",
820 		     strerror(ret));
821 		goto out;
822 	}
823 
824 	fc_conf->autoneg = ethpause.autoneg;
825 	if (ethpause.rx_pause && ethpause.tx_pause)
826 		fc_conf->mode = RTE_FC_FULL;
827 	else if (ethpause.rx_pause)
828 		fc_conf->mode = RTE_FC_RX_PAUSE;
829 	else if (ethpause.tx_pause)
830 		fc_conf->mode = RTE_FC_TX_PAUSE;
831 	else
832 		fc_conf->mode = RTE_FC_NONE;
833 	ret = 0;
834 
835 out:
836 	priv_unlock(priv);
837 	assert(ret >= 0);
838 	return -ret;
839 }
840 
841 /**
842  * DPDK callback to modify flow control parameters.
843  *
844  * @param dev
845  *   Pointer to Ethernet device structure.
846  * @param[in] fc_conf
847  *   Flow control parameters.
848  *
849  * @return
850  *   0 on success, negative errno value on failure.
851  */
852 int
853 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
854 {
855 	struct priv *priv = dev->data->dev_private;
856 	struct ifreq ifr;
857 	struct ethtool_pauseparam ethpause = {
858 		.cmd = ETHTOOL_SPAUSEPARAM
859 	};
860 	int ret;
861 
862 	ifr.ifr_data = (void *)&ethpause;
863 	ethpause.autoneg = fc_conf->autoneg;
864 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
865 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
866 		ethpause.rx_pause = 1;
867 	else
868 		ethpause.rx_pause = 0;
869 
870 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
871 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
872 		ethpause.tx_pause = 1;
873 	else
874 		ethpause.tx_pause = 0;
875 
876 	priv_lock(priv);
877 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
878 		ret = errno;
879 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
880 		     " failed: %s",
881 		     strerror(ret));
882 		goto out;
883 	}
884 	ret = 0;
885 
886 out:
887 	priv_unlock(priv);
888 	assert(ret >= 0);
889 	return -ret;
890 }
891 
892 /**
893  * Get PCI information from struct ibv_device.
894  *
895  * @param device
896  *   Pointer to Ethernet device structure.
897  * @param[out] pci_addr
898  *   PCI bus address output buffer.
899  *
900  * @return
901  *   0 on success, -1 on failure and errno is set.
902  */
903 int
904 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
905 			    struct rte_pci_addr *pci_addr)
906 {
907 	FILE *file;
908 	char line[32];
909 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
910 
911 	file = fopen(path, "rb");
912 	if (file == NULL)
913 		return -1;
914 	while (fgets(line, sizeof(line), file) == line) {
915 		size_t len = strlen(line);
916 		int ret;
917 
918 		/* Truncate long lines. */
919 		if (len == (sizeof(line) - 1))
920 			while (line[(len - 1)] != '\n') {
921 				ret = fgetc(file);
922 				if (ret == EOF)
923 					break;
924 				line[(len - 1)] = ret;
925 			}
926 		/* Extract information. */
927 		if (sscanf(line,
928 			   "PCI_SLOT_NAME="
929 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
930 			   &pci_addr->domain,
931 			   &pci_addr->bus,
932 			   &pci_addr->devid,
933 			   &pci_addr->function) == 4) {
934 			ret = 0;
935 			break;
936 		}
937 	}
938 	fclose(file);
939 	return 0;
940 }
941 
942 /**
943  * Update the link status.
944  *
945  * @param priv
946  *   Pointer to private structure.
947  *
948  * @return
949  *   Zero if the callback process can be called immediately.
950  */
951 static int
952 priv_link_status_update(struct priv *priv)
953 {
954 	struct rte_eth_link *link = &priv->dev->data->dev_link;
955 
956 	priv_link_update(priv, 0);
957 	if (((link->link_speed == 0) && link->link_status) ||
958 		((link->link_speed != 0) && !link->link_status)) {
959 		/*
960 		 * Inconsistent status. Event likely occurred before the
961 		 * kernel netdevice exposes the new status.
962 		 */
963 		if (!priv->pending_alarm) {
964 			priv->pending_alarm = 1;
965 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
966 					  mlx5_dev_link_status_handler,
967 					  priv->dev);
968 		}
969 		return 1;
970 	} else if (unlikely(priv->pending_alarm)) {
971 		/* Link interrupt occurred while alarm is already scheduled. */
972 		priv->pending_alarm = 0;
973 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
974 	}
975 	return 0;
976 }
977 
978 /**
979  * Device status handler.
980  *
981  * @param priv
982  *   Pointer to private structure.
983  * @param events
984  *   Pointer to event flags holder.
985  *
986  * @return
987  *   Events bitmap of callback process which can be called immediately.
988  */
989 static uint32_t
990 priv_dev_status_handler(struct priv *priv)
991 {
992 	struct ibv_async_event event;
993 	uint32_t ret = 0;
994 
995 	/* Read all message and acknowledge them. */
996 	for (;;) {
997 		if (mlx5_glue->get_async_event(priv->ctx, &event))
998 			break;
999 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1000 			event.event_type == IBV_EVENT_PORT_ERR) &&
1001 			(priv->dev->data->dev_conf.intr_conf.lsc == 1))
1002 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1003 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1004 			priv->dev->data->dev_conf.intr_conf.rmv == 1)
1005 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1006 		else
1007 			DEBUG("event type %d on port %d not handled",
1008 			      event.event_type, event.element.port_num);
1009 		mlx5_glue->ack_async_event(&event);
1010 	}
1011 	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1012 		if (priv_link_status_update(priv))
1013 			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1014 	return ret;
1015 }
1016 
1017 /**
1018  * Handle delayed link status event.
1019  *
1020  * @param arg
1021  *   Registered argument.
1022  */
1023 void
1024 mlx5_dev_link_status_handler(void *arg)
1025 {
1026 	struct rte_eth_dev *dev = arg;
1027 	struct priv *priv = dev->data->dev_private;
1028 	int ret;
1029 
1030 	while (!priv_trylock(priv)) {
1031 		/* Alarm is being canceled. */
1032 		if (priv->pending_alarm == 0)
1033 			return;
1034 		rte_pause();
1035 	}
1036 	priv->pending_alarm = 0;
1037 	ret = priv_link_status_update(priv);
1038 	priv_unlock(priv);
1039 	if (!ret)
1040 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1041 }
1042 
1043 /**
1044  * Handle interrupts from the NIC.
1045  *
1046  * @param[in] intr_handle
1047  *   Interrupt handler.
1048  * @param cb_arg
1049  *   Callback argument.
1050  */
1051 void
1052 mlx5_dev_interrupt_handler(void *cb_arg)
1053 {
1054 	struct rte_eth_dev *dev = cb_arg;
1055 	struct priv *priv = dev->data->dev_private;
1056 	uint32_t events;
1057 
1058 	priv_lock(priv);
1059 	events = priv_dev_status_handler(priv);
1060 	priv_unlock(priv);
1061 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1062 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1063 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1064 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1065 }
1066 
1067 /**
1068  * Handle interrupts from the socket.
1069  *
1070  * @param cb_arg
1071  *   Callback argument.
1072  */
1073 static void
1074 mlx5_dev_handler_socket(void *cb_arg)
1075 {
1076 	struct rte_eth_dev *dev = cb_arg;
1077 	struct priv *priv = dev->data->dev_private;
1078 
1079 	priv_lock(priv);
1080 	priv_socket_handle(priv);
1081 	priv_unlock(priv);
1082 }
1083 
1084 /**
1085  * Uninstall interrupt handler.
1086  *
1087  * @param priv
1088  *   Pointer to private structure.
1089  * @param dev
1090  *   Pointer to the rte_eth_dev structure.
1091  */
1092 void
1093 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1094 {
1095 	if (dev->data->dev_conf.intr_conf.lsc ||
1096 	    dev->data->dev_conf.intr_conf.rmv)
1097 		rte_intr_callback_unregister(&priv->intr_handle,
1098 					     mlx5_dev_interrupt_handler, dev);
1099 	if (priv->primary_socket)
1100 		rte_intr_callback_unregister(&priv->intr_handle_socket,
1101 					     mlx5_dev_handler_socket, dev);
1102 	if (priv->pending_alarm) {
1103 		priv->pending_alarm = 0;
1104 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1105 	}
1106 	priv->intr_handle.fd = 0;
1107 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1108 	priv->intr_handle_socket.fd = 0;
1109 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1110 }
1111 
1112 /**
1113  * Install interrupt handler.
1114  *
1115  * @param priv
1116  *   Pointer to private structure.
1117  * @param dev
1118  *   Pointer to the rte_eth_dev structure.
1119  */
1120 void
1121 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1122 {
1123 	int rc, flags;
1124 
1125 	assert(priv->ctx->async_fd > 0);
1126 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1127 	rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1128 	if (rc < 0) {
1129 		INFO("failed to change file descriptor async event queue");
1130 		dev->data->dev_conf.intr_conf.lsc = 0;
1131 		dev->data->dev_conf.intr_conf.rmv = 0;
1132 	}
1133 	if (dev->data->dev_conf.intr_conf.lsc ||
1134 	    dev->data->dev_conf.intr_conf.rmv) {
1135 		priv->intr_handle.fd = priv->ctx->async_fd;
1136 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1137 		rte_intr_callback_register(&priv->intr_handle,
1138 					   mlx5_dev_interrupt_handler, dev);
1139 	}
1140 
1141 	rc = priv_socket_init(priv);
1142 	if (!rc && priv->primary_socket) {
1143 		priv->intr_handle_socket.fd = priv->primary_socket;
1144 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1145 		rte_intr_callback_register(&priv->intr_handle_socket,
1146 					   mlx5_dev_handler_socket, dev);
1147 	}
1148 }
1149 
1150 /**
1151  * Change the link state (UP / DOWN).
1152  *
1153  * @param priv
1154  *   Pointer to private data structure.
1155  * @param up
1156  *   Nonzero for link up, otherwise link down.
1157  *
1158  * @return
1159  *   0 on success, errno value on failure.
1160  */
1161 static int
1162 priv_dev_set_link(struct priv *priv, int up)
1163 {
1164 	return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1165 }
1166 
1167 /**
1168  * DPDK callback to bring the link DOWN.
1169  *
1170  * @param dev
1171  *   Pointer to Ethernet device structure.
1172  *
1173  * @return
1174  *   0 on success, errno value on failure.
1175  */
1176 int
1177 mlx5_set_link_down(struct rte_eth_dev *dev)
1178 {
1179 	struct priv *priv = dev->data->dev_private;
1180 	int err;
1181 
1182 	priv_lock(priv);
1183 	err = priv_dev_set_link(priv, 0);
1184 	priv_unlock(priv);
1185 	return err;
1186 }
1187 
1188 /**
1189  * DPDK callback to bring the link UP.
1190  *
1191  * @param dev
1192  *   Pointer to Ethernet device structure.
1193  *
1194  * @return
1195  *   0 on success, errno value on failure.
1196  */
1197 int
1198 mlx5_set_link_up(struct rte_eth_dev *dev)
1199 {
1200 	struct priv *priv = dev->data->dev_private;
1201 	int err;
1202 
1203 	priv_lock(priv);
1204 	err = priv_dev_set_link(priv, 1);
1205 	priv_unlock(priv);
1206 	return err;
1207 }
1208 
1209 /**
1210  * Configure the TX function to use.
1211  *
1212  * @param priv
1213  *   Pointer to private data structure.
1214  * @param dev
1215  *   Pointer to rte_eth_dev structure.
1216  *
1217  * @return
1218  *   Pointer to selected Tx burst function.
1219  */
1220 eth_tx_burst_t
1221 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1222 {
1223 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1224 	struct mlx5_dev_config *config = &priv->config;
1225 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1226 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1227 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1228 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1229 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1230 
1231 	assert(priv != NULL);
1232 	/* Select appropriate TX function. */
1233 	if (vlan_insert || tso)
1234 		return tx_pkt_burst;
1235 	if (config->mps == MLX5_MPW_ENHANCED) {
1236 		if (priv_check_vec_tx_support(priv, dev) > 0) {
1237 			if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1238 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1239 			else
1240 				tx_pkt_burst = mlx5_tx_burst_vec;
1241 			DEBUG("selected Enhanced MPW TX vectorized function");
1242 		} else {
1243 			tx_pkt_burst = mlx5_tx_burst_empw;
1244 			DEBUG("selected Enhanced MPW TX function");
1245 		}
1246 	} else if (config->mps && (config->txq_inline > 0)) {
1247 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1248 		DEBUG("selected MPW inline TX function");
1249 	} else if (config->mps) {
1250 		tx_pkt_burst = mlx5_tx_burst_mpw;
1251 		DEBUG("selected MPW TX function");
1252 	}
1253 	return tx_pkt_burst;
1254 }
1255 
1256 /**
1257  * Configure the RX function to use.
1258  *
1259  * @param priv
1260  *   Pointer to private data structure.
1261  * @param dev
1262  *   Pointer to rte_eth_dev structure.
1263  *
1264  * @return
1265  *   Pointer to selected Rx burst function.
1266  */
1267 eth_rx_burst_t
1268 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1269 {
1270 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1271 
1272 	assert(priv != NULL);
1273 	if (priv_check_vec_rx_support(priv) > 0) {
1274 		rx_pkt_burst = mlx5_rx_burst_vec;
1275 		DEBUG("selected RX vectorized function");
1276 	}
1277 	return rx_pkt_burst;
1278 }
1279 
1280 /**
1281  * Check if mlx5 device was removed.
1282  *
1283  * @param dev
1284  *   Pointer to Ethernet device structure.
1285  *
1286  * @return
1287  *   1 when device is removed, otherwise 0.
1288  */
1289 int
1290 mlx5_is_removed(struct rte_eth_dev *dev)
1291 {
1292 	struct ibv_device_attr device_attr;
1293 	struct priv *priv = dev->data->dev_private;
1294 
1295 	if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1296 		return 1;
1297 	return 0;
1298 }
1299