xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision e313ef4c2fe83509785ba195e75389d1ed6b6afb)
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 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
696 	priv_unlock(priv);
697 }
698 
699 const uint32_t *
700 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
701 {
702 	static const uint32_t ptypes[] = {
703 		/* refers to rxq_cq_to_pkt_type() */
704 		RTE_PTYPE_L2_ETHER,
705 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
706 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
707 		RTE_PTYPE_L4_NONFRAG,
708 		RTE_PTYPE_L4_FRAG,
709 		RTE_PTYPE_L4_TCP,
710 		RTE_PTYPE_L4_UDP,
711 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
712 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
713 		RTE_PTYPE_INNER_L4_NONFRAG,
714 		RTE_PTYPE_INNER_L4_FRAG,
715 		RTE_PTYPE_INNER_L4_TCP,
716 		RTE_PTYPE_INNER_L4_UDP,
717 		RTE_PTYPE_UNKNOWN
718 	};
719 
720 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
721 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
722 		return ptypes;
723 	return NULL;
724 }
725 
726 /**
727  * DPDK callback to retrieve physical link information.
728  *
729  * @param dev
730  *   Pointer to Ethernet device structure.
731  * @param wait_to_complete
732  *   Wait for request completion (ignored).
733  */
734 static int
735 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
736 {
737 	struct priv *priv = dev->data->dev_private;
738 	struct ethtool_cmd edata = {
739 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
740 	};
741 	struct ifreq ifr;
742 	struct rte_eth_link dev_link;
743 	int link_speed = 0;
744 
745 	/* priv_lock() is not taken to allow concurrent calls. */
746 
747 	(void)wait_to_complete;
748 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
749 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
750 		return -1;
751 	}
752 	memset(&dev_link, 0, sizeof(dev_link));
753 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
754 				(ifr.ifr_flags & IFF_RUNNING));
755 	ifr.ifr_data = (void *)&edata;
756 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
757 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
758 		     strerror(errno));
759 		return -1;
760 	}
761 	link_speed = ethtool_cmd_speed(&edata);
762 	if (link_speed == -1)
763 		dev_link.link_speed = 0;
764 	else
765 		dev_link.link_speed = link_speed;
766 	priv->link_speed_capa = 0;
767 	if (edata.supported & SUPPORTED_Autoneg)
768 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
769 	if (edata.supported & (SUPPORTED_1000baseT_Full |
770 			       SUPPORTED_1000baseKX_Full))
771 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
772 	if (edata.supported & SUPPORTED_10000baseKR_Full)
773 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
774 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
775 			       SUPPORTED_40000baseCR4_Full |
776 			       SUPPORTED_40000baseSR4_Full |
777 			       SUPPORTED_40000baseLR4_Full))
778 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
779 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
780 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
781 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
782 			ETH_LINK_SPEED_FIXED);
783 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
784 		/* Link status changed. */
785 		dev->data->dev_link = dev_link;
786 		return 0;
787 	}
788 	/* Link status is still the same. */
789 	return -1;
790 }
791 
792 /**
793  * Retrieve physical link information (unlocked version using new ioctl).
794  *
795  * @param dev
796  *   Pointer to Ethernet device structure.
797  * @param wait_to_complete
798  *   Wait for request completion (ignored).
799  */
800 static int
801 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
802 {
803 	struct priv *priv = dev->data->dev_private;
804 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
805 	struct ifreq ifr;
806 	struct rte_eth_link dev_link;
807 	uint64_t sc;
808 
809 	(void)wait_to_complete;
810 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
811 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
812 		return -1;
813 	}
814 	memset(&dev_link, 0, sizeof(dev_link));
815 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
816 				(ifr.ifr_flags & IFF_RUNNING));
817 	ifr.ifr_data = (void *)&gcmd;
818 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
819 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
820 		      strerror(errno));
821 		return -1;
822 	}
823 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
824 
825 	alignas(struct ethtool_link_settings)
826 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
827 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
828 	struct ethtool_link_settings *ecmd = (void *)data;
829 
830 	*ecmd = gcmd;
831 	ifr.ifr_data = (void *)ecmd;
832 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
833 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
834 		      strerror(errno));
835 		return -1;
836 	}
837 	dev_link.link_speed = ecmd->speed;
838 	sc = ecmd->link_mode_masks[0] |
839 		((uint64_t)ecmd->link_mode_masks[1] << 32);
840 	priv->link_speed_capa = 0;
841 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
842 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
843 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
844 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
845 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
846 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
847 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
848 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
849 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
850 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
851 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
852 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
853 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
854 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
855 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
856 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
857 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
858 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
859 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
860 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
861 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
862 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
863 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
864 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
865 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
866 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
867 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
868 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
869 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
870 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
871 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
872 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
873 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
874 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
875 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
876 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
877 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
878 				  ETH_LINK_SPEED_FIXED);
879 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
880 		/* Link status changed. */
881 		dev->data->dev_link = dev_link;
882 		return 0;
883 	}
884 	/* Link status is still the same. */
885 	return -1;
886 }
887 
888 /**
889  * Enable receiving and transmitting traffic.
890  *
891  * @param priv
892  *   Pointer to private structure.
893  */
894 static void
895 priv_link_start(struct priv *priv)
896 {
897 	struct rte_eth_dev *dev = priv->dev;
898 	int err;
899 
900 	dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
901 	dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
902 	err = priv_dev_traffic_enable(priv, dev);
903 	if (err)
904 		ERROR("%p: error occurred while configuring control flows: %s",
905 		      (void *)priv, strerror(err));
906 	err = priv_flow_start(priv, &priv->flows);
907 	if (err)
908 		ERROR("%p: error occurred while configuring flows: %s",
909 		      (void *)priv, strerror(err));
910 }
911 
912 /**
913  * Disable receiving and transmitting traffic.
914  *
915  * @param priv
916  *   Pointer to private structure.
917  */
918 static void
919 priv_link_stop(struct priv *priv)
920 {
921 	struct rte_eth_dev *dev = priv->dev;
922 
923 	priv_flow_stop(priv, &priv->flows);
924 	priv_dev_traffic_disable(priv, dev);
925 	dev->rx_pkt_burst = removed_rx_burst;
926 	dev->tx_pkt_burst = removed_tx_burst;
927 }
928 
929 /**
930  * Retrieve physical link information and update rx/tx_pkt_burst callbacks
931  * accordingly.
932  *
933  * @param priv
934  *   Pointer to private structure.
935  * @param wait_to_complete
936  *   Wait for request completion (ignored).
937  */
938 int
939 priv_link_update(struct priv *priv, int wait_to_complete)
940 {
941 	struct rte_eth_dev *dev = priv->dev;
942 	struct utsname utsname;
943 	int ver[3];
944 	int ret;
945 	struct rte_eth_link dev_link = dev->data->dev_link;
946 
947 	if (uname(&utsname) == -1 ||
948 	    sscanf(utsname.release, "%d.%d.%d",
949 		   &ver[0], &ver[1], &ver[2]) != 3 ||
950 	    KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
951 		ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
952 	else
953 		ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
954 	/* If lsc interrupt is disabled, should always be ready for traffic. */
955 	if (!dev->data->dev_conf.intr_conf.lsc) {
956 		priv_link_start(priv);
957 		return ret;
958 	}
959 	/* Re-select burst callbacks only if link status has been changed. */
960 	if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
961 		if (dev->data->dev_link.link_status == ETH_LINK_UP)
962 			priv_link_start(priv);
963 		else
964 			priv_link_stop(priv);
965 	}
966 	return ret;
967 }
968 
969 /**
970  * Querying the link status till it changes to the desired state.
971  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
972  *
973  * @param priv
974  *   Pointer to private structure.
975  * @param status
976  *   Link desired status.
977  *
978  * @return
979  *   0 on success, negative errno value on failure.
980  */
981 int
982 priv_force_link_status_change(struct priv *priv, int status)
983 {
984 	int try = 0;
985 
986 	while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
987 		priv_link_update(priv, 0);
988 		if (priv->dev->data->dev_link.link_status == status)
989 			return 0;
990 		try++;
991 		sleep(1);
992 	}
993 	return -EAGAIN;
994 }
995 
996 /**
997  * DPDK callback to retrieve physical link information.
998  *
999  * @param dev
1000  *   Pointer to Ethernet device structure.
1001  * @param wait_to_complete
1002  *   Wait for request completion (ignored).
1003  */
1004 int
1005 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1006 {
1007 	struct priv *priv = dev->data->dev_private;
1008 	int ret;
1009 
1010 	priv_lock(priv);
1011 	ret = priv_link_update(priv, wait_to_complete);
1012 	priv_unlock(priv);
1013 	return ret;
1014 }
1015 
1016 /**
1017  * DPDK callback to change the MTU.
1018  *
1019  * @param dev
1020  *   Pointer to Ethernet device structure.
1021  * @param in_mtu
1022  *   New MTU.
1023  *
1024  * @return
1025  *   0 on success, negative errno value on failure.
1026  */
1027 int
1028 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1029 {
1030 	struct priv *priv = dev->data->dev_private;
1031 	uint16_t kern_mtu;
1032 	int ret = 0;
1033 
1034 	priv_lock(priv);
1035 	ret = priv_get_mtu(priv, &kern_mtu);
1036 	if (ret)
1037 		goto out;
1038 	/* Set kernel interface MTU first. */
1039 	ret = priv_set_mtu(priv, mtu);
1040 	if (ret)
1041 		goto out;
1042 	ret = priv_get_mtu(priv, &kern_mtu);
1043 	if (ret)
1044 		goto out;
1045 	if (kern_mtu == mtu) {
1046 		priv->mtu = mtu;
1047 		DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
1048 	}
1049 	priv_unlock(priv);
1050 	return 0;
1051 out:
1052 	ret = errno;
1053 	WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
1054 	     strerror(ret));
1055 	priv_unlock(priv);
1056 	assert(ret >= 0);
1057 	return -ret;
1058 }
1059 
1060 /**
1061  * DPDK callback to get flow control status.
1062  *
1063  * @param dev
1064  *   Pointer to Ethernet device structure.
1065  * @param[out] fc_conf
1066  *   Flow control output buffer.
1067  *
1068  * @return
1069  *   0 on success, negative errno value on failure.
1070  */
1071 int
1072 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1073 {
1074 	struct priv *priv = dev->data->dev_private;
1075 	struct ifreq ifr;
1076 	struct ethtool_pauseparam ethpause = {
1077 		.cmd = ETHTOOL_GPAUSEPARAM
1078 	};
1079 	int ret;
1080 
1081 	ifr.ifr_data = (void *)&ethpause;
1082 	priv_lock(priv);
1083 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1084 		ret = errno;
1085 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1086 		     " failed: %s",
1087 		     strerror(ret));
1088 		goto out;
1089 	}
1090 
1091 	fc_conf->autoneg = ethpause.autoneg;
1092 	if (ethpause.rx_pause && ethpause.tx_pause)
1093 		fc_conf->mode = RTE_FC_FULL;
1094 	else if (ethpause.rx_pause)
1095 		fc_conf->mode = RTE_FC_RX_PAUSE;
1096 	else if (ethpause.tx_pause)
1097 		fc_conf->mode = RTE_FC_TX_PAUSE;
1098 	else
1099 		fc_conf->mode = RTE_FC_NONE;
1100 	ret = 0;
1101 
1102 out:
1103 	priv_unlock(priv);
1104 	assert(ret >= 0);
1105 	return -ret;
1106 }
1107 
1108 /**
1109  * DPDK callback to modify flow control parameters.
1110  *
1111  * @param dev
1112  *   Pointer to Ethernet device structure.
1113  * @param[in] fc_conf
1114  *   Flow control parameters.
1115  *
1116  * @return
1117  *   0 on success, negative errno value on failure.
1118  */
1119 int
1120 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1121 {
1122 	struct priv *priv = dev->data->dev_private;
1123 	struct ifreq ifr;
1124 	struct ethtool_pauseparam ethpause = {
1125 		.cmd = ETHTOOL_SPAUSEPARAM
1126 	};
1127 	int ret;
1128 
1129 	ifr.ifr_data = (void *)&ethpause;
1130 	ethpause.autoneg = fc_conf->autoneg;
1131 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1132 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1133 		ethpause.rx_pause = 1;
1134 	else
1135 		ethpause.rx_pause = 0;
1136 
1137 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1138 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1139 		ethpause.tx_pause = 1;
1140 	else
1141 		ethpause.tx_pause = 0;
1142 
1143 	priv_lock(priv);
1144 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1145 		ret = errno;
1146 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1147 		     " failed: %s",
1148 		     strerror(ret));
1149 		goto out;
1150 	}
1151 	ret = 0;
1152 
1153 out:
1154 	priv_unlock(priv);
1155 	assert(ret >= 0);
1156 	return -ret;
1157 }
1158 
1159 /**
1160  * Get PCI information from struct ibv_device.
1161  *
1162  * @param device
1163  *   Pointer to Ethernet device structure.
1164  * @param[out] pci_addr
1165  *   PCI bus address output buffer.
1166  *
1167  * @return
1168  *   0 on success, -1 on failure and errno is set.
1169  */
1170 int
1171 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1172 			    struct rte_pci_addr *pci_addr)
1173 {
1174 	FILE *file;
1175 	char line[32];
1176 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
1177 
1178 	file = fopen(path, "rb");
1179 	if (file == NULL)
1180 		return -1;
1181 	while (fgets(line, sizeof(line), file) == line) {
1182 		size_t len = strlen(line);
1183 		int ret;
1184 
1185 		/* Truncate long lines. */
1186 		if (len == (sizeof(line) - 1))
1187 			while (line[(len - 1)] != '\n') {
1188 				ret = fgetc(file);
1189 				if (ret == EOF)
1190 					break;
1191 				line[(len - 1)] = ret;
1192 			}
1193 		/* Extract information. */
1194 		if (sscanf(line,
1195 			   "PCI_SLOT_NAME="
1196 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1197 			   &pci_addr->domain,
1198 			   &pci_addr->bus,
1199 			   &pci_addr->devid,
1200 			   &pci_addr->function) == 4) {
1201 			ret = 0;
1202 			break;
1203 		}
1204 	}
1205 	fclose(file);
1206 	return 0;
1207 }
1208 
1209 /**
1210  * Update the link status.
1211  *
1212  * @param priv
1213  *   Pointer to private structure.
1214  *
1215  * @return
1216  *   Zero if the callback process can be called immediately.
1217  */
1218 static int
1219 priv_link_status_update(struct priv *priv)
1220 {
1221 	struct rte_eth_link *link = &priv->dev->data->dev_link;
1222 
1223 	priv_link_update(priv, 0);
1224 	if (((link->link_speed == 0) && link->link_status) ||
1225 		((link->link_speed != 0) && !link->link_status)) {
1226 		/*
1227 		 * Inconsistent status. Event likely occurred before the
1228 		 * kernel netdevice exposes the new status.
1229 		 */
1230 		if (!priv->pending_alarm) {
1231 			priv->pending_alarm = 1;
1232 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1233 					  mlx5_dev_link_status_handler,
1234 					  priv->dev);
1235 		}
1236 		return 1;
1237 	} else if (unlikely(priv->pending_alarm)) {
1238 		/* Link interrupt occurred while alarm is already scheduled. */
1239 		priv->pending_alarm = 0;
1240 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
1241 	}
1242 	return 0;
1243 }
1244 
1245 /**
1246  * Device status handler.
1247  *
1248  * @param priv
1249  *   Pointer to private structure.
1250  * @param events
1251  *   Pointer to event flags holder.
1252  *
1253  * @return
1254  *   Events bitmap of callback process which can be called immediately.
1255  */
1256 static uint32_t
1257 priv_dev_status_handler(struct priv *priv)
1258 {
1259 	struct ibv_async_event event;
1260 	uint32_t ret = 0;
1261 
1262 	/* Read all message and acknowledge them. */
1263 	for (;;) {
1264 		if (ibv_get_async_event(priv->ctx, &event))
1265 			break;
1266 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1267 			event.event_type == IBV_EVENT_PORT_ERR) &&
1268 			(priv->dev->data->dev_conf.intr_conf.lsc == 1))
1269 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1270 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1271 			priv->dev->data->dev_conf.intr_conf.rmv == 1)
1272 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1273 		else
1274 			DEBUG("event type %d on port %d not handled",
1275 			      event.event_type, event.element.port_num);
1276 		ibv_ack_async_event(&event);
1277 	}
1278 	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1279 		if (priv_link_status_update(priv))
1280 			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1281 	return ret;
1282 }
1283 
1284 /**
1285  * Handle delayed link status event.
1286  *
1287  * @param arg
1288  *   Registered argument.
1289  */
1290 void
1291 mlx5_dev_link_status_handler(void *arg)
1292 {
1293 	struct rte_eth_dev *dev = arg;
1294 	struct priv *priv = dev->data->dev_private;
1295 	int ret;
1296 
1297 	while (!priv_trylock(priv)) {
1298 		/* Alarm is being canceled. */
1299 		if (priv->pending_alarm == 0)
1300 			return;
1301 		rte_pause();
1302 	}
1303 	priv->pending_alarm = 0;
1304 	ret = priv_link_status_update(priv);
1305 	priv_unlock(priv);
1306 	if (!ret)
1307 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1308 }
1309 
1310 /**
1311  * Handle interrupts from the NIC.
1312  *
1313  * @param[in] intr_handle
1314  *   Interrupt handler.
1315  * @param cb_arg
1316  *   Callback argument.
1317  */
1318 void
1319 mlx5_dev_interrupt_handler(void *cb_arg)
1320 {
1321 	struct rte_eth_dev *dev = cb_arg;
1322 	struct priv *priv = dev->data->dev_private;
1323 	uint32_t events;
1324 
1325 	priv_lock(priv);
1326 	events = priv_dev_status_handler(priv);
1327 	priv_unlock(priv);
1328 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1329 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1330 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1331 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1332 }
1333 
1334 /**
1335  * Handle interrupts from the socket.
1336  *
1337  * @param cb_arg
1338  *   Callback argument.
1339  */
1340 static void
1341 mlx5_dev_handler_socket(void *cb_arg)
1342 {
1343 	struct rte_eth_dev *dev = cb_arg;
1344 	struct priv *priv = dev->data->dev_private;
1345 
1346 	priv_lock(priv);
1347 	priv_socket_handle(priv);
1348 	priv_unlock(priv);
1349 }
1350 
1351 /**
1352  * Uninstall 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_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1361 {
1362 	if (dev->data->dev_conf.intr_conf.lsc ||
1363 	    dev->data->dev_conf.intr_conf.rmv)
1364 		rte_intr_callback_unregister(&priv->intr_handle,
1365 					     mlx5_dev_interrupt_handler, dev);
1366 	if (priv->primary_socket)
1367 		rte_intr_callback_unregister(&priv->intr_handle_socket,
1368 					     mlx5_dev_handler_socket, dev);
1369 	if (priv->pending_alarm) {
1370 		priv->pending_alarm = 0;
1371 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1372 	}
1373 	priv->intr_handle.fd = 0;
1374 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1375 	priv->intr_handle_socket.fd = 0;
1376 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1377 }
1378 
1379 /**
1380  * Install interrupt handler.
1381  *
1382  * @param priv
1383  *   Pointer to private structure.
1384  * @param dev
1385  *   Pointer to the rte_eth_dev structure.
1386  */
1387 void
1388 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1389 {
1390 	int rc, flags;
1391 
1392 	assert(priv->ctx->async_fd > 0);
1393 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1394 	rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1395 	if (rc < 0) {
1396 		INFO("failed to change file descriptor async event queue");
1397 		dev->data->dev_conf.intr_conf.lsc = 0;
1398 		dev->data->dev_conf.intr_conf.rmv = 0;
1399 	}
1400 	if (dev->data->dev_conf.intr_conf.lsc ||
1401 	    dev->data->dev_conf.intr_conf.rmv) {
1402 		priv->intr_handle.fd = priv->ctx->async_fd;
1403 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1404 		rte_intr_callback_register(&priv->intr_handle,
1405 					   mlx5_dev_interrupt_handler, dev);
1406 	}
1407 
1408 	rc = priv_socket_init(priv);
1409 	if (!rc && priv->primary_socket) {
1410 		priv->intr_handle_socket.fd = priv->primary_socket;
1411 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1412 		rte_intr_callback_register(&priv->intr_handle_socket,
1413 					   mlx5_dev_handler_socket, dev);
1414 	}
1415 }
1416 
1417 /**
1418  * Change the link state (UP / DOWN).
1419  *
1420  * @param priv
1421  *   Pointer to private data structure.
1422  * @param up
1423  *   Nonzero for link up, otherwise link down.
1424  *
1425  * @return
1426  *   0 on success, errno value on failure.
1427  */
1428 static int
1429 priv_dev_set_link(struct priv *priv, int up)
1430 {
1431 	return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1432 }
1433 
1434 /**
1435  * DPDK callback to bring the link DOWN.
1436  *
1437  * @param dev
1438  *   Pointer to Ethernet device structure.
1439  *
1440  * @return
1441  *   0 on success, errno value on failure.
1442  */
1443 int
1444 mlx5_set_link_down(struct rte_eth_dev *dev)
1445 {
1446 	struct priv *priv = dev->data->dev_private;
1447 	int err;
1448 
1449 	priv_lock(priv);
1450 	err = priv_dev_set_link(priv, 0);
1451 	priv_unlock(priv);
1452 	return err;
1453 }
1454 
1455 /**
1456  * DPDK callback to bring the link UP.
1457  *
1458  * @param dev
1459  *   Pointer to Ethernet device structure.
1460  *
1461  * @return
1462  *   0 on success, errno value on failure.
1463  */
1464 int
1465 mlx5_set_link_up(struct rte_eth_dev *dev)
1466 {
1467 	struct priv *priv = dev->data->dev_private;
1468 	int err;
1469 
1470 	priv_lock(priv);
1471 	err = priv_dev_set_link(priv, 1);
1472 	priv_unlock(priv);
1473 	return err;
1474 }
1475 
1476 /**
1477  * Configure the TX function to use.
1478  *
1479  * @param priv
1480  *   Pointer to private data structure.
1481  * @param dev
1482  *   Pointer to rte_eth_dev structure.
1483  *
1484  * @return
1485  *   Pointer to selected Tx burst function.
1486  */
1487 eth_tx_burst_t
1488 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1489 {
1490 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1491 	struct mlx5_dev_config *config = &priv->config;
1492 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1493 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1494 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1495 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1496 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1497 
1498 	assert(priv != NULL);
1499 	/* Select appropriate TX function. */
1500 	if (vlan_insert || tso)
1501 		return tx_pkt_burst;
1502 	if (config->mps == MLX5_MPW_ENHANCED) {
1503 		if (priv_check_vec_tx_support(priv, dev) > 0) {
1504 			if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1505 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1506 			else
1507 				tx_pkt_burst = mlx5_tx_burst_vec;
1508 			DEBUG("selected Enhanced MPW TX vectorized function");
1509 		} else {
1510 			tx_pkt_burst = mlx5_tx_burst_empw;
1511 			DEBUG("selected Enhanced MPW TX function");
1512 		}
1513 	} else if (config->mps && (config->txq_inline > 0)) {
1514 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1515 		DEBUG("selected MPW inline TX function");
1516 	} else if (config->mps) {
1517 		tx_pkt_burst = mlx5_tx_burst_mpw;
1518 		DEBUG("selected MPW TX function");
1519 	}
1520 	return tx_pkt_burst;
1521 }
1522 
1523 /**
1524  * Configure the RX function to use.
1525  *
1526  * @param priv
1527  *   Pointer to private data structure.
1528  * @param dev
1529  *   Pointer to rte_eth_dev structure.
1530  *
1531  * @return
1532  *   Pointer to selected Rx burst function.
1533  */
1534 eth_rx_burst_t
1535 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1536 {
1537 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1538 
1539 	assert(priv != NULL);
1540 	if (priv_check_vec_rx_support(priv) > 0) {
1541 		rx_pkt_burst = mlx5_rx_burst_vec;
1542 		DEBUG("selected RX vectorized function");
1543 	}
1544 	return rx_pkt_burst;
1545 }
1546 
1547 /**
1548  * Check if mlx5 device was removed.
1549  *
1550  * @param dev
1551  *   Pointer to Ethernet device structure.
1552  *
1553  * @return
1554  *   1 when device is removed, otherwise 0.
1555  */
1556 int
1557 mlx5_is_removed(struct rte_eth_dev *dev)
1558 {
1559 	struct ibv_device_attr device_attr;
1560 	struct priv *priv = dev->data->dev_private;
1561 
1562 	if (ibv_query_device(priv->ctx, &device_attr) == EIO)
1563 		return 1;
1564 	return 0;
1565 }
1566