xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 42c3576d44dd48602aba0820b98c07e2c4278c0e)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2015 6WIND S.A.
5  *   Copyright 2015 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of 6WIND S.A. nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #define _GNU_SOURCE
35 
36 #include <stddef.h>
37 #include <assert.h>
38 #include <unistd.h>
39 #include <stdint.h>
40 #include <stdio.h>
41 #include <string.h>
42 #include <stdlib.h>
43 #include <errno.h>
44 #include <dirent.h>
45 #include <net/if.h>
46 #include <sys/ioctl.h>
47 #include <sys/socket.h>
48 #include <sys/utsname.h>
49 #include <netinet/in.h>
50 #include <linux/ethtool.h>
51 #include <linux/sockios.h>
52 #include <linux/version.h>
53 #include <fcntl.h>
54 #include <stdalign.h>
55 #include <sys/un.h>
56 
57 #include <rte_atomic.h>
58 #include <rte_ethdev_driver.h>
59 #include <rte_bus_pci.h>
60 #include <rte_mbuf.h>
61 #include <rte_common.h>
62 #include <rte_interrupts.h>
63 #include <rte_alarm.h>
64 #include <rte_malloc.h>
65 
66 #include "mlx5.h"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_utils.h"
69 
70 /* Add defines in case the running kernel is not the same as user headers. */
71 #ifndef ETHTOOL_GLINKSETTINGS
72 struct ethtool_link_settings {
73 	uint32_t cmd;
74 	uint32_t speed;
75 	uint8_t duplex;
76 	uint8_t port;
77 	uint8_t phy_address;
78 	uint8_t autoneg;
79 	uint8_t mdio_support;
80 	uint8_t eth_to_mdix;
81 	uint8_t eth_tp_mdix_ctrl;
82 	int8_t link_mode_masks_nwords;
83 	uint32_t reserved[8];
84 	uint32_t link_mode_masks[];
85 };
86 
87 #define ETHTOOL_GLINKSETTINGS 0x0000004c
88 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
89 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
90 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
91 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
92 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
93 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
94 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
95 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
96 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
97 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
98 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
99 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
100 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
101 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
102 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
103 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
104 #endif
105 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
106 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
107 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
108 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
109 #endif
110 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
111 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
112 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
113 #endif
114 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
115 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
116 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
117 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
118 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
119 #endif
120 
121 /**
122  * Get interface name from private structure.
123  *
124  * @param[in] priv
125  *   Pointer to private structure.
126  * @param[out] ifname
127  *   Interface name output buffer.
128  *
129  * @return
130  *   0 on success, -1 on failure and errno is set.
131  */
132 int
133 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
134 {
135 	DIR *dir;
136 	struct dirent *dent;
137 	unsigned int dev_type = 0;
138 	unsigned int dev_port_prev = ~0u;
139 	char match[IF_NAMESIZE] = "";
140 
141 	{
142 		MKSTR(path, "%s/device/net", priv->ibdev_path);
143 
144 		dir = opendir(path);
145 		if (dir == NULL)
146 			return -1;
147 	}
148 	while ((dent = readdir(dir)) != NULL) {
149 		char *name = dent->d_name;
150 		FILE *file;
151 		unsigned int dev_port;
152 		int r;
153 
154 		if ((name[0] == '.') &&
155 		    ((name[1] == '\0') ||
156 		     ((name[1] == '.') && (name[2] == '\0'))))
157 			continue;
158 
159 		MKSTR(path, "%s/device/net/%s/%s",
160 		      priv->ibdev_path, name,
161 		      (dev_type ? "dev_id" : "dev_port"));
162 
163 		file = fopen(path, "rb");
164 		if (file == NULL) {
165 			if (errno != ENOENT)
166 				continue;
167 			/*
168 			 * Switch to dev_id when dev_port does not exist as
169 			 * is the case with Linux kernel versions < 3.15.
170 			 */
171 try_dev_id:
172 			match[0] = '\0';
173 			if (dev_type)
174 				break;
175 			dev_type = 1;
176 			dev_port_prev = ~0u;
177 			rewinddir(dir);
178 			continue;
179 		}
180 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
181 		fclose(file);
182 		if (r != 1)
183 			continue;
184 		/*
185 		 * Switch to dev_id when dev_port returns the same value for
186 		 * all ports. May happen when using a MOFED release older than
187 		 * 3.0 with a Linux kernel >= 3.15.
188 		 */
189 		if (dev_port == dev_port_prev)
190 			goto try_dev_id;
191 		dev_port_prev = dev_port;
192 		if (dev_port == (priv->port - 1u))
193 			snprintf(match, sizeof(match), "%s", name);
194 	}
195 	closedir(dir);
196 	if (match[0] == '\0')
197 		return -1;
198 	strncpy(*ifname, match, sizeof(*ifname));
199 	return 0;
200 }
201 
202 /**
203  * Check if the counter is located on ib counters file.
204  *
205  * @param[in] cntr
206  *   Counter name.
207  *
208  * @return
209  *   1 if counter is located on ib counters file , 0 otherwise.
210  */
211 int
212 priv_is_ib_cntr(const char *cntr)
213 {
214 	if (!strcmp(cntr, "out_of_buffer"))
215 		return 1;
216 	return 0;
217 }
218 
219 /**
220  * Read from sysfs entry.
221  *
222  * @param[in] priv
223  *   Pointer to private structure.
224  * @param[in] entry
225  *   Entry name relative to sysfs path.
226  * @param[out] buf
227  *   Data output buffer.
228  * @param size
229  *   Buffer size.
230  *
231  * @return
232  *   0 on success, -1 on failure and errno is set.
233  */
234 static int
235 priv_sysfs_read(const struct priv *priv, const char *entry,
236 		char *buf, size_t size)
237 {
238 	char ifname[IF_NAMESIZE];
239 	FILE *file;
240 	int ret;
241 	int err;
242 
243 	if (priv_get_ifname(priv, &ifname))
244 		return -1;
245 
246 	if (priv_is_ib_cntr(entry)) {
247 		MKSTR(path, "%s/ports/1/hw_counters/%s",
248 		      priv->ibdev_path, entry);
249 		file = fopen(path, "rb");
250 	} else {
251 		MKSTR(path, "%s/device/net/%s/%s",
252 		      priv->ibdev_path, ifname, entry);
253 		file = fopen(path, "rb");
254 	}
255 	if (file == NULL)
256 		return -1;
257 	ret = fread(buf, 1, size, file);
258 	err = errno;
259 	if (((size_t)ret < size) && (ferror(file)))
260 		ret = -1;
261 	else
262 		ret = size;
263 	fclose(file);
264 	errno = err;
265 	return ret;
266 }
267 
268 /**
269  * Write to sysfs entry.
270  *
271  * @param[in] priv
272  *   Pointer to private structure.
273  * @param[in] entry
274  *   Entry name relative to sysfs path.
275  * @param[in] buf
276  *   Data buffer.
277  * @param size
278  *   Buffer size.
279  *
280  * @return
281  *   0 on success, -1 on failure and errno is set.
282  */
283 static int
284 priv_sysfs_write(const struct priv *priv, const char *entry,
285 		 char *buf, size_t size)
286 {
287 	char ifname[IF_NAMESIZE];
288 	FILE *file;
289 	int ret;
290 	int err;
291 
292 	if (priv_get_ifname(priv, &ifname))
293 		return -1;
294 
295 	MKSTR(path, "%s/device/net/%s/%s", priv->ibdev_path, ifname, entry);
296 
297 	file = fopen(path, "wb");
298 	if (file == NULL)
299 		return -1;
300 	ret = fwrite(buf, 1, size, file);
301 	err = errno;
302 	if (((size_t)ret < size) || (ferror(file)))
303 		ret = -1;
304 	else
305 		ret = size;
306 	fclose(file);
307 	errno = err;
308 	return ret;
309 }
310 
311 /**
312  * Get unsigned long sysfs property.
313  *
314  * @param priv
315  *   Pointer to private structure.
316  * @param[in] name
317  *   Entry name relative to sysfs path.
318  * @param[out] value
319  *   Value output buffer.
320  *
321  * @return
322  *   0 on success, -1 on failure and errno is set.
323  */
324 static int
325 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
326 {
327 	int ret;
328 	unsigned long value_ret;
329 	char value_str[32];
330 
331 	ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
332 	if (ret == -1) {
333 		DEBUG("cannot read %s value from sysfs: %s",
334 		      name, strerror(errno));
335 		return -1;
336 	}
337 	value_str[ret] = '\0';
338 	errno = 0;
339 	value_ret = strtoul(value_str, NULL, 0);
340 	if (errno) {
341 		DEBUG("invalid %s value `%s': %s", name, value_str,
342 		      strerror(errno));
343 		return -1;
344 	}
345 	*value = value_ret;
346 	return 0;
347 }
348 
349 /**
350  * Set unsigned long sysfs property.
351  *
352  * @param priv
353  *   Pointer to private structure.
354  * @param[in] name
355  *   Entry name relative to sysfs path.
356  * @param value
357  *   Value to set.
358  *
359  * @return
360  *   0 on success, -1 on failure and errno is set.
361  */
362 static int
363 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
364 {
365 	int ret;
366 	MKSTR(value_str, "%lu", value);
367 
368 	ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
369 	if (ret == -1) {
370 		DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
371 		      name, value_str, value, strerror(errno));
372 		return -1;
373 	}
374 	return 0;
375 }
376 
377 /**
378  * Perform ifreq ioctl() on associated Ethernet device.
379  *
380  * @param[in] priv
381  *   Pointer to private structure.
382  * @param req
383  *   Request number to pass to ioctl().
384  * @param[out] ifr
385  *   Interface request structure output buffer.
386  *
387  * @return
388  *   0 on success, -1 on failure and errno is set.
389  */
390 int
391 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
392 {
393 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
394 	int ret = -1;
395 
396 	if (sock == -1)
397 		return ret;
398 	if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
399 		ret = ioctl(sock, req, ifr);
400 	close(sock);
401 	return ret;
402 }
403 
404 /**
405  * Return the number of active VFs for the current device.
406  *
407  * @param[in] priv
408  *   Pointer to private structure.
409  * @param[out] num_vfs
410  *   Number of active VFs.
411  *
412  * @return
413  *   0 on success, -1 on failure and errno is set.
414  */
415 int
416 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
417 {
418 	/* The sysfs entry name depends on the operating system. */
419 	const char **name = (const char *[]){
420 		"device/sriov_numvfs",
421 		"device/mlx5_num_vfs",
422 		NULL,
423 	};
424 	int ret;
425 
426 	do {
427 		unsigned long ulong_num_vfs;
428 
429 		ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
430 		if (!ret)
431 			*num_vfs = ulong_num_vfs;
432 	} while (*(++name) && ret);
433 	return ret;
434 }
435 
436 /**
437  * Get device MTU.
438  *
439  * @param priv
440  *   Pointer to private structure.
441  * @param[out] mtu
442  *   MTU value output buffer.
443  *
444  * @return
445  *   0 on success, -1 on failure and errno is set.
446  */
447 int
448 priv_get_mtu(struct priv *priv, uint16_t *mtu)
449 {
450 	unsigned long ulong_mtu;
451 
452 	if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
453 		return -1;
454 	*mtu = ulong_mtu;
455 	return 0;
456 }
457 
458 /**
459  * Read device counter from sysfs.
460  *
461  * @param priv
462  *   Pointer to private structure.
463  * @param name
464  *   Counter name.
465  * @param[out] cntr
466  *   Counter output buffer.
467  *
468  * @return
469  *   0 on success, -1 on failure and errno is set.
470  */
471 int
472 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr)
473 {
474 	unsigned long ulong_ctr;
475 
476 	if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1)
477 		return -1;
478 	*cntr = ulong_ctr;
479 	return 0;
480 }
481 
482 /**
483  * Set device MTU.
484  *
485  * @param priv
486  *   Pointer to private structure.
487  * @param mtu
488  *   MTU value to set.
489  *
490  * @return
491  *   0 on success, -1 on failure and errno is set.
492  */
493 static int
494 priv_set_mtu(struct priv *priv, uint16_t mtu)
495 {
496 	uint16_t new_mtu;
497 
498 	if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
499 	    priv_get_mtu(priv, &new_mtu))
500 		return -1;
501 	if (new_mtu == mtu)
502 		return 0;
503 	errno = EINVAL;
504 	return -1;
505 }
506 
507 /**
508  * Set device flags.
509  *
510  * @param priv
511  *   Pointer to private structure.
512  * @param keep
513  *   Bitmask for flags that must remain untouched.
514  * @param flags
515  *   Bitmask for flags to modify.
516  *
517  * @return
518  *   0 on success, -1 on failure and errno is set.
519  */
520 int
521 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
522 {
523 	unsigned long tmp;
524 
525 	if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
526 		return -1;
527 	tmp &= keep;
528 	tmp |= (flags & (~keep));
529 	return priv_set_sysfs_ulong(priv, "flags", tmp);
530 }
531 
532 /**
533  * Ethernet device configuration.
534  *
535  * Prepare the driver for a given number of TX and RX queues.
536  *
537  * @param dev
538  *   Pointer to Ethernet device structure.
539  *
540  * @return
541  *   0 on success, errno value on failure.
542  */
543 static int
544 dev_configure(struct rte_eth_dev *dev)
545 {
546 	struct priv *priv = dev->data->dev_private;
547 	unsigned int rxqs_n = dev->data->nb_rx_queues;
548 	unsigned int txqs_n = dev->data->nb_tx_queues;
549 	unsigned int i;
550 	unsigned int j;
551 	unsigned int reta_idx_n;
552 	const uint8_t use_app_rss_key =
553 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
554 	uint64_t supp_tx_offloads = mlx5_priv_get_tx_port_offloads(priv);
555 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
556 	uint64_t supp_rx_offloads =
557 		(mlx5_priv_get_rx_port_offloads(priv) |
558 		 mlx5_priv_get_rx_queue_offloads(priv));
559 	uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
560 
561 	if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
562 		ERROR("Some Tx offloads are not supported "
563 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
564 		      tx_offloads, supp_tx_offloads);
565 		return ENOTSUP;
566 	}
567 	if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
568 		ERROR("Some Rx offloads are not supported "
569 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
570 		      rx_offloads, supp_rx_offloads);
571 		return ENOTSUP;
572 	}
573 	if (use_app_rss_key &&
574 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
575 	     rss_hash_default_key_len)) {
576 		/* MLX5 RSS only support 40bytes key. */
577 		return EINVAL;
578 	}
579 	priv->rss_conf.rss_key =
580 		rte_realloc(priv->rss_conf.rss_key,
581 			    rss_hash_default_key_len, 0);
582 	if (!priv->rss_conf.rss_key) {
583 		ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
584 		return ENOMEM;
585 	}
586 	memcpy(priv->rss_conf.rss_key,
587 	       use_app_rss_key ?
588 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
589 	       rss_hash_default_key,
590 	       rss_hash_default_key_len);
591 	priv->rss_conf.rss_key_len = rss_hash_default_key_len;
592 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
593 	priv->rxqs = (void *)dev->data->rx_queues;
594 	priv->txqs = (void *)dev->data->tx_queues;
595 	if (txqs_n != priv->txqs_n) {
596 		INFO("%p: TX queues number update: %u -> %u",
597 		     (void *)dev, priv->txqs_n, txqs_n);
598 		priv->txqs_n = txqs_n;
599 	}
600 	if (rxqs_n > priv->config.ind_table_max_size) {
601 		ERROR("cannot handle this many RX queues (%u)", rxqs_n);
602 		return EINVAL;
603 	}
604 	if (rxqs_n == priv->rxqs_n)
605 		return 0;
606 	INFO("%p: RX queues number update: %u -> %u",
607 	     (void *)dev, priv->rxqs_n, rxqs_n);
608 	priv->rxqs_n = rxqs_n;
609 	/* If the requested number of RX queues is not a power of two, use the
610 	 * maximum indirection table size for better balancing.
611 	 * The result is always rounded to the next power of two. */
612 	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
613 				     priv->config.ind_table_max_size :
614 				     rxqs_n));
615 	if (priv_rss_reta_index_resize(priv, reta_idx_n))
616 		return ENOMEM;
617 	/* When the number of RX queues is not a power of two, the remaining
618 	 * table entries are padded with reused WQs and hashes are not spread
619 	 * uniformly. */
620 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
621 		(*priv->reta_idx)[i] = j;
622 		if (++j == rxqs_n)
623 			j = 0;
624 	}
625 	return 0;
626 }
627 
628 /**
629  * DPDK callback for Ethernet device configuration.
630  *
631  * @param dev
632  *   Pointer to Ethernet device structure.
633  *
634  * @return
635  *   0 on success, negative errno value on failure.
636  */
637 int
638 mlx5_dev_configure(struct rte_eth_dev *dev)
639 {
640 	struct priv *priv = dev->data->dev_private;
641 	int ret;
642 
643 	priv_lock(priv);
644 	ret = dev_configure(dev);
645 	assert(ret >= 0);
646 	priv_unlock(priv);
647 	return -ret;
648 }
649 
650 /**
651  * DPDK callback to get information about the device.
652  *
653  * @param dev
654  *   Pointer to Ethernet device structure.
655  * @param[out] info
656  *   Info structure output buffer.
657  */
658 void
659 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
660 {
661 	struct priv *priv = dev->data->dev_private;
662 	struct mlx5_dev_config *config = &priv->config;
663 	unsigned int max;
664 	char ifname[IF_NAMESIZE];
665 
666 	info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
667 
668 	priv_lock(priv);
669 	/* FIXME: we should ask the device for these values. */
670 	info->min_rx_bufsize = 32;
671 	info->max_rx_pktlen = 65536;
672 	/*
673 	 * Since we need one CQ per QP, the limit is the minimum number
674 	 * between the two values.
675 	 */
676 	max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
677 		      priv->device_attr.orig_attr.max_qp);
678 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
679 	if (max >= 65535)
680 		max = 65535;
681 	info->max_rx_queues = max;
682 	info->max_tx_queues = max;
683 	info->max_mac_addrs = RTE_DIM(priv->mac);
684 	info->rx_queue_offload_capa =
685 		mlx5_priv_get_rx_queue_offloads(priv);
686 	info->rx_offload_capa = (mlx5_priv_get_rx_port_offloads(priv) |
687 				 info->rx_queue_offload_capa);
688 	info->tx_offload_capa = mlx5_priv_get_tx_port_offloads(priv);
689 	if (priv_get_ifname(priv, &ifname) == 0)
690 		info->if_index = if_nametoindex(ifname);
691 	info->reta_size = priv->reta_idx_n ?
692 		priv->reta_idx_n : config->ind_table_max_size;
693 	info->hash_key_size = priv->rss_conf.rss_key_len;
694 	info->speed_capa = priv->link_speed_capa;
695 	priv_unlock(priv);
696 }
697 
698 const uint32_t *
699 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
700 {
701 	static const uint32_t ptypes[] = {
702 		/* refers to rxq_cq_to_pkt_type() */
703 		RTE_PTYPE_L2_ETHER,
704 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
705 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
706 		RTE_PTYPE_L4_NONFRAG,
707 		RTE_PTYPE_L4_FRAG,
708 		RTE_PTYPE_L4_TCP,
709 		RTE_PTYPE_L4_UDP,
710 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
711 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
712 		RTE_PTYPE_INNER_L4_NONFRAG,
713 		RTE_PTYPE_INNER_L4_FRAG,
714 		RTE_PTYPE_INNER_L4_TCP,
715 		RTE_PTYPE_INNER_L4_UDP,
716 		RTE_PTYPE_UNKNOWN
717 	};
718 
719 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
720 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
721 		return ptypes;
722 	return NULL;
723 }
724 
725 /**
726  * DPDK callback to retrieve physical link information.
727  *
728  * @param dev
729  *   Pointer to Ethernet device structure.
730  * @param wait_to_complete
731  *   Wait for request completion (ignored).
732  */
733 static int
734 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
735 {
736 	struct priv *priv = dev->data->dev_private;
737 	struct ethtool_cmd edata = {
738 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
739 	};
740 	struct ifreq ifr;
741 	struct rte_eth_link dev_link;
742 	int link_speed = 0;
743 
744 	/* priv_lock() is not taken to allow concurrent calls. */
745 
746 	(void)wait_to_complete;
747 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
748 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
749 		return -1;
750 	}
751 	memset(&dev_link, 0, sizeof(dev_link));
752 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
753 				(ifr.ifr_flags & IFF_RUNNING));
754 	ifr.ifr_data = (void *)&edata;
755 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
756 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
757 		     strerror(errno));
758 		return -1;
759 	}
760 	link_speed = ethtool_cmd_speed(&edata);
761 	if (link_speed == -1)
762 		dev_link.link_speed = 0;
763 	else
764 		dev_link.link_speed = link_speed;
765 	priv->link_speed_capa = 0;
766 	if (edata.supported & SUPPORTED_Autoneg)
767 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
768 	if (edata.supported & (SUPPORTED_1000baseT_Full |
769 			       SUPPORTED_1000baseKX_Full))
770 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
771 	if (edata.supported & SUPPORTED_10000baseKR_Full)
772 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
773 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
774 			       SUPPORTED_40000baseCR4_Full |
775 			       SUPPORTED_40000baseSR4_Full |
776 			       SUPPORTED_40000baseLR4_Full))
777 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
778 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
779 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
780 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
781 			ETH_LINK_SPEED_FIXED);
782 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
783 		/* Link status changed. */
784 		dev->data->dev_link = dev_link;
785 		return 0;
786 	}
787 	/* Link status is still the same. */
788 	return -1;
789 }
790 
791 /**
792  * Retrieve physical link information (unlocked version using new ioctl).
793  *
794  * @param dev
795  *   Pointer to Ethernet device structure.
796  * @param wait_to_complete
797  *   Wait for request completion (ignored).
798  */
799 static int
800 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
801 {
802 	struct priv *priv = dev->data->dev_private;
803 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
804 	struct ifreq ifr;
805 	struct rte_eth_link dev_link;
806 	uint64_t sc;
807 
808 	(void)wait_to_complete;
809 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
810 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
811 		return -1;
812 	}
813 	memset(&dev_link, 0, sizeof(dev_link));
814 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
815 				(ifr.ifr_flags & IFF_RUNNING));
816 	ifr.ifr_data = (void *)&gcmd;
817 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
818 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
819 		      strerror(errno));
820 		return -1;
821 	}
822 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
823 
824 	alignas(struct ethtool_link_settings)
825 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
826 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
827 	struct ethtool_link_settings *ecmd = (void *)data;
828 
829 	*ecmd = gcmd;
830 	ifr.ifr_data = (void *)ecmd;
831 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
832 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
833 		      strerror(errno));
834 		return -1;
835 	}
836 	dev_link.link_speed = ecmd->speed;
837 	sc = ecmd->link_mode_masks[0] |
838 		((uint64_t)ecmd->link_mode_masks[1] << 32);
839 	priv->link_speed_capa = 0;
840 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
841 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
842 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
843 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
844 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
845 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
846 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
847 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
848 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
849 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
850 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
851 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
852 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
853 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
854 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
855 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
856 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
857 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
858 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
859 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
860 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
861 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
862 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
863 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
864 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
865 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
866 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
867 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
868 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
869 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
870 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
871 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
872 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
873 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
874 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
875 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
876 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
877 				  ETH_LINK_SPEED_FIXED);
878 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
879 		/* Link status changed. */
880 		dev->data->dev_link = dev_link;
881 		return 0;
882 	}
883 	/* Link status is still the same. */
884 	return -1;
885 }
886 
887 /**
888  * Enable receiving and transmitting traffic.
889  *
890  * @param priv
891  *   Pointer to private structure.
892  */
893 static void
894 priv_link_start(struct priv *priv)
895 {
896 	struct rte_eth_dev *dev = priv->dev;
897 	int err;
898 
899 	dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
900 	dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
901 	err = priv_dev_traffic_enable(priv, dev);
902 	if (err)
903 		ERROR("%p: error occurred while configuring control flows: %s",
904 		      (void *)priv, strerror(err));
905 	err = priv_flow_start(priv, &priv->flows);
906 	if (err)
907 		ERROR("%p: error occurred while configuring flows: %s",
908 		      (void *)priv, strerror(err));
909 }
910 
911 /**
912  * Disable receiving and transmitting traffic.
913  *
914  * @param priv
915  *   Pointer to private structure.
916  */
917 static void
918 priv_link_stop(struct priv *priv)
919 {
920 	struct rte_eth_dev *dev = priv->dev;
921 
922 	priv_flow_stop(priv, &priv->flows);
923 	priv_dev_traffic_disable(priv, dev);
924 	dev->rx_pkt_burst = removed_rx_burst;
925 	dev->tx_pkt_burst = removed_tx_burst;
926 }
927 
928 /**
929  * Retrieve physical link information and update rx/tx_pkt_burst callbacks
930  * accordingly.
931  *
932  * @param priv
933  *   Pointer to private structure.
934  * @param wait_to_complete
935  *   Wait for request completion (ignored).
936  */
937 int
938 priv_link_update(struct priv *priv, int wait_to_complete)
939 {
940 	struct rte_eth_dev *dev = priv->dev;
941 	struct utsname utsname;
942 	int ver[3];
943 	int ret;
944 	struct rte_eth_link dev_link = dev->data->dev_link;
945 
946 	if (uname(&utsname) == -1 ||
947 	    sscanf(utsname.release, "%d.%d.%d",
948 		   &ver[0], &ver[1], &ver[2]) != 3 ||
949 	    KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
950 		ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
951 	else
952 		ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
953 	/* If lsc interrupt is disabled, should always be ready for traffic. */
954 	if (!dev->data->dev_conf.intr_conf.lsc) {
955 		priv_link_start(priv);
956 		return ret;
957 	}
958 	/* Re-select burst callbacks only if link status has been changed. */
959 	if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
960 		if (dev->data->dev_link.link_status == ETH_LINK_UP)
961 			priv_link_start(priv);
962 		else
963 			priv_link_stop(priv);
964 	}
965 	return ret;
966 }
967 
968 /**
969  * DPDK callback to retrieve physical link information.
970  *
971  * @param dev
972  *   Pointer to Ethernet device structure.
973  * @param wait_to_complete
974  *   Wait for request completion (ignored).
975  */
976 int
977 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
978 {
979 	struct priv *priv = dev->data->dev_private;
980 	int ret;
981 
982 	priv_lock(priv);
983 	ret = priv_link_update(priv, wait_to_complete);
984 	priv_unlock(priv);
985 	return ret;
986 }
987 
988 /**
989  * DPDK callback to change the MTU.
990  *
991  * @param dev
992  *   Pointer to Ethernet device structure.
993  * @param in_mtu
994  *   New MTU.
995  *
996  * @return
997  *   0 on success, negative errno value on failure.
998  */
999 int
1000 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1001 {
1002 	struct priv *priv = dev->data->dev_private;
1003 	uint16_t kern_mtu;
1004 	int ret = 0;
1005 
1006 	priv_lock(priv);
1007 	ret = priv_get_mtu(priv, &kern_mtu);
1008 	if (ret)
1009 		goto out;
1010 	/* Set kernel interface MTU first. */
1011 	ret = priv_set_mtu(priv, mtu);
1012 	if (ret)
1013 		goto out;
1014 	ret = priv_get_mtu(priv, &kern_mtu);
1015 	if (ret)
1016 		goto out;
1017 	if (kern_mtu == mtu) {
1018 		priv->mtu = mtu;
1019 		DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
1020 	}
1021 	priv_unlock(priv);
1022 	return 0;
1023 out:
1024 	ret = errno;
1025 	WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
1026 	     strerror(ret));
1027 	priv_unlock(priv);
1028 	assert(ret >= 0);
1029 	return -ret;
1030 }
1031 
1032 /**
1033  * DPDK callback to get flow control status.
1034  *
1035  * @param dev
1036  *   Pointer to Ethernet device structure.
1037  * @param[out] fc_conf
1038  *   Flow control output buffer.
1039  *
1040  * @return
1041  *   0 on success, negative errno value on failure.
1042  */
1043 int
1044 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1045 {
1046 	struct priv *priv = dev->data->dev_private;
1047 	struct ifreq ifr;
1048 	struct ethtool_pauseparam ethpause = {
1049 		.cmd = ETHTOOL_GPAUSEPARAM
1050 	};
1051 	int ret;
1052 
1053 	ifr.ifr_data = (void *)&ethpause;
1054 	priv_lock(priv);
1055 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1056 		ret = errno;
1057 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1058 		     " failed: %s",
1059 		     strerror(ret));
1060 		goto out;
1061 	}
1062 
1063 	fc_conf->autoneg = ethpause.autoneg;
1064 	if (ethpause.rx_pause && ethpause.tx_pause)
1065 		fc_conf->mode = RTE_FC_FULL;
1066 	else if (ethpause.rx_pause)
1067 		fc_conf->mode = RTE_FC_RX_PAUSE;
1068 	else if (ethpause.tx_pause)
1069 		fc_conf->mode = RTE_FC_TX_PAUSE;
1070 	else
1071 		fc_conf->mode = RTE_FC_NONE;
1072 	ret = 0;
1073 
1074 out:
1075 	priv_unlock(priv);
1076 	assert(ret >= 0);
1077 	return -ret;
1078 }
1079 
1080 /**
1081  * DPDK callback to modify flow control parameters.
1082  *
1083  * @param dev
1084  *   Pointer to Ethernet device structure.
1085  * @param[in] fc_conf
1086  *   Flow control parameters.
1087  *
1088  * @return
1089  *   0 on success, negative errno value on failure.
1090  */
1091 int
1092 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1093 {
1094 	struct priv *priv = dev->data->dev_private;
1095 	struct ifreq ifr;
1096 	struct ethtool_pauseparam ethpause = {
1097 		.cmd = ETHTOOL_SPAUSEPARAM
1098 	};
1099 	int ret;
1100 
1101 	ifr.ifr_data = (void *)&ethpause;
1102 	ethpause.autoneg = fc_conf->autoneg;
1103 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1104 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1105 		ethpause.rx_pause = 1;
1106 	else
1107 		ethpause.rx_pause = 0;
1108 
1109 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1110 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1111 		ethpause.tx_pause = 1;
1112 	else
1113 		ethpause.tx_pause = 0;
1114 
1115 	priv_lock(priv);
1116 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1117 		ret = errno;
1118 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1119 		     " failed: %s",
1120 		     strerror(ret));
1121 		goto out;
1122 	}
1123 	ret = 0;
1124 
1125 out:
1126 	priv_unlock(priv);
1127 	assert(ret >= 0);
1128 	return -ret;
1129 }
1130 
1131 /**
1132  * Get PCI information from struct ibv_device.
1133  *
1134  * @param device
1135  *   Pointer to Ethernet device structure.
1136  * @param[out] pci_addr
1137  *   PCI bus address output buffer.
1138  *
1139  * @return
1140  *   0 on success, -1 on failure and errno is set.
1141  */
1142 int
1143 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1144 			    struct rte_pci_addr *pci_addr)
1145 {
1146 	FILE *file;
1147 	char line[32];
1148 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
1149 
1150 	file = fopen(path, "rb");
1151 	if (file == NULL)
1152 		return -1;
1153 	while (fgets(line, sizeof(line), file) == line) {
1154 		size_t len = strlen(line);
1155 		int ret;
1156 
1157 		/* Truncate long lines. */
1158 		if (len == (sizeof(line) - 1))
1159 			while (line[(len - 1)] != '\n') {
1160 				ret = fgetc(file);
1161 				if (ret == EOF)
1162 					break;
1163 				line[(len - 1)] = ret;
1164 			}
1165 		/* Extract information. */
1166 		if (sscanf(line,
1167 			   "PCI_SLOT_NAME="
1168 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1169 			   &pci_addr->domain,
1170 			   &pci_addr->bus,
1171 			   &pci_addr->devid,
1172 			   &pci_addr->function) == 4) {
1173 			ret = 0;
1174 			break;
1175 		}
1176 	}
1177 	fclose(file);
1178 	return 0;
1179 }
1180 
1181 /**
1182  * Update the link status.
1183  *
1184  * @param priv
1185  *   Pointer to private structure.
1186  *
1187  * @return
1188  *   Zero if the callback process can be called immediately.
1189  */
1190 static int
1191 priv_link_status_update(struct priv *priv)
1192 {
1193 	struct rte_eth_link *link = &priv->dev->data->dev_link;
1194 
1195 	priv_link_update(priv, 0);
1196 	if (((link->link_speed == 0) && link->link_status) ||
1197 		((link->link_speed != 0) && !link->link_status)) {
1198 		/*
1199 		 * Inconsistent status. Event likely occurred before the
1200 		 * kernel netdevice exposes the new status.
1201 		 */
1202 		if (!priv->pending_alarm) {
1203 			priv->pending_alarm = 1;
1204 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1205 					  mlx5_dev_link_status_handler,
1206 					  priv->dev);
1207 		}
1208 		return 1;
1209 	} else if (unlikely(priv->pending_alarm)) {
1210 		/* Link interrupt occurred while alarm is already scheduled. */
1211 		priv->pending_alarm = 0;
1212 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
1213 	}
1214 	return 0;
1215 }
1216 
1217 /**
1218  * Device status handler.
1219  *
1220  * @param priv
1221  *   Pointer to private structure.
1222  * @param events
1223  *   Pointer to event flags holder.
1224  *
1225  * @return
1226  *   Events bitmap of callback process which can be called immediately.
1227  */
1228 static uint32_t
1229 priv_dev_status_handler(struct priv *priv)
1230 {
1231 	struct ibv_async_event event;
1232 	uint32_t ret = 0;
1233 
1234 	/* Read all message and acknowledge them. */
1235 	for (;;) {
1236 		if (ibv_get_async_event(priv->ctx, &event))
1237 			break;
1238 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1239 			event.event_type == IBV_EVENT_PORT_ERR) &&
1240 			(priv->dev->data->dev_conf.intr_conf.lsc == 1))
1241 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1242 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1243 			priv->dev->data->dev_conf.intr_conf.rmv == 1)
1244 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1245 		else
1246 			DEBUG("event type %d on port %d not handled",
1247 			      event.event_type, event.element.port_num);
1248 		ibv_ack_async_event(&event);
1249 	}
1250 	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1251 		if (priv_link_status_update(priv))
1252 			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1253 	return ret;
1254 }
1255 
1256 /**
1257  * Handle delayed link status event.
1258  *
1259  * @param arg
1260  *   Registered argument.
1261  */
1262 void
1263 mlx5_dev_link_status_handler(void *arg)
1264 {
1265 	struct rte_eth_dev *dev = arg;
1266 	struct priv *priv = dev->data->dev_private;
1267 	int ret;
1268 
1269 	while (!priv_trylock(priv)) {
1270 		/* Alarm is being canceled. */
1271 		if (priv->pending_alarm == 0)
1272 			return;
1273 		rte_pause();
1274 	}
1275 	priv->pending_alarm = 0;
1276 	ret = priv_link_status_update(priv);
1277 	priv_unlock(priv);
1278 	if (!ret)
1279 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1280 }
1281 
1282 /**
1283  * Handle interrupts from the NIC.
1284  *
1285  * @param[in] intr_handle
1286  *   Interrupt handler.
1287  * @param cb_arg
1288  *   Callback argument.
1289  */
1290 void
1291 mlx5_dev_interrupt_handler(void *cb_arg)
1292 {
1293 	struct rte_eth_dev *dev = cb_arg;
1294 	struct priv *priv = dev->data->dev_private;
1295 	uint32_t events;
1296 
1297 	priv_lock(priv);
1298 	events = priv_dev_status_handler(priv);
1299 	priv_unlock(priv);
1300 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1301 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1302 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1303 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1304 }
1305 
1306 /**
1307  * Handle interrupts from the socket.
1308  *
1309  * @param cb_arg
1310  *   Callback argument.
1311  */
1312 static void
1313 mlx5_dev_handler_socket(void *cb_arg)
1314 {
1315 	struct rte_eth_dev *dev = cb_arg;
1316 	struct priv *priv = dev->data->dev_private;
1317 
1318 	priv_lock(priv);
1319 	priv_socket_handle(priv);
1320 	priv_unlock(priv);
1321 }
1322 
1323 /**
1324  * Uninstall interrupt handler.
1325  *
1326  * @param priv
1327  *   Pointer to private structure.
1328  * @param dev
1329  *   Pointer to the rte_eth_dev structure.
1330  */
1331 void
1332 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1333 {
1334 	if (dev->data->dev_conf.intr_conf.lsc ||
1335 	    dev->data->dev_conf.intr_conf.rmv)
1336 		rte_intr_callback_unregister(&priv->intr_handle,
1337 					     mlx5_dev_interrupt_handler, dev);
1338 	if (priv->primary_socket)
1339 		rte_intr_callback_unregister(&priv->intr_handle_socket,
1340 					     mlx5_dev_handler_socket, dev);
1341 	if (priv->pending_alarm) {
1342 		priv->pending_alarm = 0;
1343 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1344 	}
1345 	priv->intr_handle.fd = 0;
1346 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1347 	priv->intr_handle_socket.fd = 0;
1348 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1349 }
1350 
1351 /**
1352  * Install interrupt handler.
1353  *
1354  * @param priv
1355  *   Pointer to private structure.
1356  * @param dev
1357  *   Pointer to the rte_eth_dev structure.
1358  */
1359 void
1360 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1361 {
1362 	int rc, flags;
1363 
1364 	assert(priv->ctx->async_fd > 0);
1365 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1366 	rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1367 	if (rc < 0) {
1368 		INFO("failed to change file descriptor async event queue");
1369 		dev->data->dev_conf.intr_conf.lsc = 0;
1370 		dev->data->dev_conf.intr_conf.rmv = 0;
1371 	}
1372 	if (dev->data->dev_conf.intr_conf.lsc ||
1373 	    dev->data->dev_conf.intr_conf.rmv) {
1374 		priv->intr_handle.fd = priv->ctx->async_fd;
1375 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1376 		rte_intr_callback_register(&priv->intr_handle,
1377 					   mlx5_dev_interrupt_handler, dev);
1378 	}
1379 
1380 	rc = priv_socket_init(priv);
1381 	if (!rc && priv->primary_socket) {
1382 		priv->intr_handle_socket.fd = priv->primary_socket;
1383 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1384 		rte_intr_callback_register(&priv->intr_handle_socket,
1385 					   mlx5_dev_handler_socket, dev);
1386 	}
1387 }
1388 
1389 /**
1390  * Change the link state (UP / DOWN).
1391  *
1392  * @param priv
1393  *   Pointer to private data structure.
1394  * @param up
1395  *   Nonzero for link up, otherwise link down.
1396  *
1397  * @return
1398  *   0 on success, errno value on failure.
1399  */
1400 static int
1401 priv_dev_set_link(struct priv *priv, int up)
1402 {
1403 	return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1404 }
1405 
1406 /**
1407  * DPDK callback to bring the link DOWN.
1408  *
1409  * @param dev
1410  *   Pointer to Ethernet device structure.
1411  *
1412  * @return
1413  *   0 on success, errno value on failure.
1414  */
1415 int
1416 mlx5_set_link_down(struct rte_eth_dev *dev)
1417 {
1418 	struct priv *priv = dev->data->dev_private;
1419 	int err;
1420 
1421 	priv_lock(priv);
1422 	err = priv_dev_set_link(priv, 0);
1423 	priv_unlock(priv);
1424 	return err;
1425 }
1426 
1427 /**
1428  * DPDK callback to bring the link UP.
1429  *
1430  * @param dev
1431  *   Pointer to Ethernet device structure.
1432  *
1433  * @return
1434  *   0 on success, errno value on failure.
1435  */
1436 int
1437 mlx5_set_link_up(struct rte_eth_dev *dev)
1438 {
1439 	struct priv *priv = dev->data->dev_private;
1440 	int err;
1441 
1442 	priv_lock(priv);
1443 	err = priv_dev_set_link(priv, 1);
1444 	priv_unlock(priv);
1445 	return err;
1446 }
1447 
1448 /**
1449  * Configure the TX function to use.
1450  *
1451  * @param priv
1452  *   Pointer to private data structure.
1453  * @param dev
1454  *   Pointer to rte_eth_dev structure.
1455  *
1456  * @return
1457  *   Pointer to selected Tx burst function.
1458  */
1459 eth_tx_burst_t
1460 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1461 {
1462 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1463 	struct mlx5_dev_config *config = &priv->config;
1464 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1465 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1466 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1467 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1468 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1469 
1470 	assert(priv != NULL);
1471 	/* Select appropriate TX function. */
1472 	if (vlan_insert || tso)
1473 		return tx_pkt_burst;
1474 	if (config->mps == MLX5_MPW_ENHANCED) {
1475 		if (priv_check_vec_tx_support(priv, dev) > 0) {
1476 			if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1477 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1478 			else
1479 				tx_pkt_burst = mlx5_tx_burst_vec;
1480 			DEBUG("selected Enhanced MPW TX vectorized function");
1481 		} else {
1482 			tx_pkt_burst = mlx5_tx_burst_empw;
1483 			DEBUG("selected Enhanced MPW TX function");
1484 		}
1485 	} else if (config->mps && (config->txq_inline > 0)) {
1486 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1487 		DEBUG("selected MPW inline TX function");
1488 	} else if (config->mps) {
1489 		tx_pkt_burst = mlx5_tx_burst_mpw;
1490 		DEBUG("selected MPW TX function");
1491 	}
1492 	return tx_pkt_burst;
1493 }
1494 
1495 /**
1496  * Configure the RX function to use.
1497  *
1498  * @param priv
1499  *   Pointer to private data structure.
1500  * @param dev
1501  *   Pointer to rte_eth_dev structure.
1502  *
1503  * @return
1504  *   Pointer to selected Rx burst function.
1505  */
1506 eth_rx_burst_t
1507 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1508 {
1509 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1510 
1511 	assert(priv != NULL);
1512 	if (priv_check_vec_rx_support(priv) > 0) {
1513 		rx_pkt_burst = mlx5_rx_burst_vec;
1514 		DEBUG("selected RX vectorized function");
1515 	}
1516 	return rx_pkt_burst;
1517 }
1518 
1519 /**
1520  * Check if mlx5 device was removed.
1521  *
1522  * @param dev
1523  *   Pointer to Ethernet device structure.
1524  *
1525  * @return
1526  *   1 when device is removed, otherwise 0.
1527  */
1528 int
1529 mlx5_is_removed(struct rte_eth_dev *dev)
1530 {
1531 	struct ibv_device_attr device_attr;
1532 	struct priv *priv = dev->data->dev_private;
1533 
1534 	if (ibv_query_device(priv->ctx, &device_attr) == EIO)
1535 		return 1;
1536 	return 0;
1537 }
1538