xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 117eaa70584b73eebf6f648cf3ee6f2ab03264a0)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox.
4  */
5 
6 #define _GNU_SOURCE
7 
8 #include <stddef.h>
9 #include <assert.h>
10 #include <unistd.h>
11 #include <stdint.h>
12 #include <stdio.h>
13 #include <string.h>
14 #include <stdlib.h>
15 #include <errno.h>
16 #include <dirent.h>
17 #include <net/if.h>
18 #include <sys/ioctl.h>
19 #include <sys/socket.h>
20 #include <sys/utsname.h>
21 #include <netinet/in.h>
22 #include <linux/ethtool.h>
23 #include <linux/sockios.h>
24 #include <linux/version.h>
25 #include <fcntl.h>
26 #include <stdalign.h>
27 #include <sys/un.h>
28 
29 #include <rte_atomic.h>
30 #include <rte_ethdev_driver.h>
31 #include <rte_bus_pci.h>
32 #include <rte_mbuf.h>
33 #include <rte_common.h>
34 #include <rte_interrupts.h>
35 #include <rte_alarm.h>
36 #include <rte_malloc.h>
37 
38 #include "mlx5.h"
39 #include "mlx5_glue.h"
40 #include "mlx5_rxtx.h"
41 #include "mlx5_utils.h"
42 
43 /* Add defines in case the running kernel is not the same as user headers. */
44 #ifndef ETHTOOL_GLINKSETTINGS
45 struct ethtool_link_settings {
46 	uint32_t cmd;
47 	uint32_t speed;
48 	uint8_t duplex;
49 	uint8_t port;
50 	uint8_t phy_address;
51 	uint8_t autoneg;
52 	uint8_t mdio_support;
53 	uint8_t eth_to_mdix;
54 	uint8_t eth_tp_mdix_ctrl;
55 	int8_t link_mode_masks_nwords;
56 	uint32_t reserved[8];
57 	uint32_t link_mode_masks[];
58 };
59 
60 #define ETHTOOL_GLINKSETTINGS 0x0000004c
61 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
62 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
63 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
64 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
65 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
66 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
67 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
68 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
69 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
70 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
71 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
72 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
73 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
74 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
75 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
76 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
77 #endif
78 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
79 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
80 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
81 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
82 #endif
83 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
84 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
85 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
86 #endif
87 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
88 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
89 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
90 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
91 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
92 #endif
93 
94 /**
95  * Get interface name from private structure.
96  *
97  * @param[in] priv
98  *   Pointer to private structure.
99  * @param[out] ifname
100  *   Interface name output buffer.
101  *
102  * @return
103  *   0 on success, -1 on failure and errno is set.
104  */
105 int
106 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
107 {
108 	DIR *dir;
109 	struct dirent *dent;
110 	unsigned int dev_type = 0;
111 	unsigned int dev_port_prev = ~0u;
112 	char match[IF_NAMESIZE] = "";
113 
114 	{
115 		MKSTR(path, "%s/device/net", priv->ibdev_path);
116 
117 		dir = opendir(path);
118 		if (dir == NULL)
119 			return -1;
120 	}
121 	while ((dent = readdir(dir)) != NULL) {
122 		char *name = dent->d_name;
123 		FILE *file;
124 		unsigned int dev_port;
125 		int r;
126 
127 		if ((name[0] == '.') &&
128 		    ((name[1] == '\0') ||
129 		     ((name[1] == '.') && (name[2] == '\0'))))
130 			continue;
131 
132 		MKSTR(path, "%s/device/net/%s/%s",
133 		      priv->ibdev_path, name,
134 		      (dev_type ? "dev_id" : "dev_port"));
135 
136 		file = fopen(path, "rb");
137 		if (file == NULL) {
138 			if (errno != ENOENT)
139 				continue;
140 			/*
141 			 * Switch to dev_id when dev_port does not exist as
142 			 * is the case with Linux kernel versions < 3.15.
143 			 */
144 try_dev_id:
145 			match[0] = '\0';
146 			if (dev_type)
147 				break;
148 			dev_type = 1;
149 			dev_port_prev = ~0u;
150 			rewinddir(dir);
151 			continue;
152 		}
153 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
154 		fclose(file);
155 		if (r != 1)
156 			continue;
157 		/*
158 		 * Switch to dev_id when dev_port returns the same value for
159 		 * all ports. May happen when using a MOFED release older than
160 		 * 3.0 with a Linux kernel >= 3.15.
161 		 */
162 		if (dev_port == dev_port_prev)
163 			goto try_dev_id;
164 		dev_port_prev = dev_port;
165 		if (dev_port == (priv->port - 1u))
166 			snprintf(match, sizeof(match), "%s", name);
167 	}
168 	closedir(dir);
169 	if (match[0] == '\0')
170 		return -1;
171 	strncpy(*ifname, match, sizeof(*ifname));
172 	return 0;
173 }
174 
175 /**
176  * Check if the counter is located on ib counters file.
177  *
178  * @param[in] cntr
179  *   Counter name.
180  *
181  * @return
182  *   1 if counter is located on ib counters file , 0 otherwise.
183  */
184 int
185 priv_is_ib_cntr(const char *cntr)
186 {
187 	if (!strcmp(cntr, "out_of_buffer"))
188 		return 1;
189 	return 0;
190 }
191 
192 /**
193  * Read from sysfs entry.
194  *
195  * @param[in] priv
196  *   Pointer to private structure.
197  * @param[in] entry
198  *   Entry name relative to sysfs path.
199  * @param[out] buf
200  *   Data output buffer.
201  * @param size
202  *   Buffer size.
203  *
204  * @return
205  *   0 on success, -1 on failure and errno is set.
206  */
207 static int
208 priv_sysfs_read(const struct priv *priv, const char *entry,
209 		char *buf, size_t size)
210 {
211 	char ifname[IF_NAMESIZE];
212 	FILE *file;
213 	int ret;
214 	int err;
215 
216 	if (priv_get_ifname(priv, &ifname))
217 		return -1;
218 
219 	if (priv_is_ib_cntr(entry)) {
220 		MKSTR(path, "%s/ports/1/hw_counters/%s",
221 		      priv->ibdev_path, entry);
222 		file = fopen(path, "rb");
223 	} else {
224 		MKSTR(path, "%s/device/net/%s/%s",
225 		      priv->ibdev_path, ifname, entry);
226 		file = fopen(path, "rb");
227 	}
228 	if (file == NULL)
229 		return -1;
230 	ret = fread(buf, 1, size, file);
231 	err = errno;
232 	if (((size_t)ret < size) && (ferror(file)))
233 		ret = -1;
234 	else
235 		ret = size;
236 	fclose(file);
237 	errno = err;
238 	return ret;
239 }
240 
241 /**
242  * Write to sysfs entry.
243  *
244  * @param[in] priv
245  *   Pointer to private structure.
246  * @param[in] entry
247  *   Entry name relative to sysfs path.
248  * @param[in] buf
249  *   Data buffer.
250  * @param size
251  *   Buffer size.
252  *
253  * @return
254  *   0 on success, -1 on failure and errno is set.
255  */
256 static int
257 priv_sysfs_write(const struct priv *priv, const char *entry,
258 		 char *buf, size_t size)
259 {
260 	char ifname[IF_NAMESIZE];
261 	FILE *file;
262 	int ret;
263 	int err;
264 
265 	if (priv_get_ifname(priv, &ifname))
266 		return -1;
267 
268 	MKSTR(path, "%s/device/net/%s/%s", priv->ibdev_path, ifname, entry);
269 
270 	file = fopen(path, "wb");
271 	if (file == NULL)
272 		return -1;
273 	ret = fwrite(buf, 1, size, file);
274 	err = errno;
275 	if (((size_t)ret < size) || (ferror(file)))
276 		ret = -1;
277 	else
278 		ret = size;
279 	fclose(file);
280 	errno = err;
281 	return ret;
282 }
283 
284 /**
285  * Get unsigned long sysfs property.
286  *
287  * @param priv
288  *   Pointer to private structure.
289  * @param[in] name
290  *   Entry name relative to sysfs path.
291  * @param[out] value
292  *   Value output buffer.
293  *
294  * @return
295  *   0 on success, -1 on failure and errno is set.
296  */
297 static int
298 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
299 {
300 	int ret;
301 	unsigned long value_ret;
302 	char value_str[32];
303 
304 	ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
305 	if (ret == -1) {
306 		DEBUG("cannot read %s value from sysfs: %s",
307 		      name, strerror(errno));
308 		return -1;
309 	}
310 	value_str[ret] = '\0';
311 	errno = 0;
312 	value_ret = strtoul(value_str, NULL, 0);
313 	if (errno) {
314 		DEBUG("invalid %s value `%s': %s", name, value_str,
315 		      strerror(errno));
316 		return -1;
317 	}
318 	*value = value_ret;
319 	return 0;
320 }
321 
322 /**
323  * Set unsigned long sysfs property.
324  *
325  * @param priv
326  *   Pointer to private structure.
327  * @param[in] name
328  *   Entry name relative to sysfs path.
329  * @param value
330  *   Value to set.
331  *
332  * @return
333  *   0 on success, -1 on failure and errno is set.
334  */
335 static int
336 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
337 {
338 	int ret;
339 	MKSTR(value_str, "%lu", value);
340 
341 	ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
342 	if (ret == -1) {
343 		DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
344 		      name, value_str, value, strerror(errno));
345 		return -1;
346 	}
347 	return 0;
348 }
349 
350 /**
351  * Perform ifreq ioctl() on associated Ethernet device.
352  *
353  * @param[in] priv
354  *   Pointer to private structure.
355  * @param req
356  *   Request number to pass to ioctl().
357  * @param[out] ifr
358  *   Interface request structure output buffer.
359  *
360  * @return
361  *   0 on success, -1 on failure and errno is set.
362  */
363 int
364 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
365 {
366 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
367 	int ret = -1;
368 
369 	if (sock == -1)
370 		return ret;
371 	if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
372 		ret = ioctl(sock, req, ifr);
373 	close(sock);
374 	return ret;
375 }
376 
377 /**
378  * Return the number of active VFs for the current device.
379  *
380  * @param[in] priv
381  *   Pointer to private structure.
382  * @param[out] num_vfs
383  *   Number of active VFs.
384  *
385  * @return
386  *   0 on success, -1 on failure and errno is set.
387  */
388 int
389 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
390 {
391 	/* The sysfs entry name depends on the operating system. */
392 	const char **name = (const char *[]){
393 		"device/sriov_numvfs",
394 		"device/mlx5_num_vfs",
395 		NULL,
396 	};
397 	int ret;
398 
399 	do {
400 		unsigned long ulong_num_vfs;
401 
402 		ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
403 		if (!ret)
404 			*num_vfs = ulong_num_vfs;
405 	} while (*(++name) && ret);
406 	return ret;
407 }
408 
409 /**
410  * Get device MTU.
411  *
412  * @param priv
413  *   Pointer to private structure.
414  * @param[out] mtu
415  *   MTU value output buffer.
416  *
417  * @return
418  *   0 on success, -1 on failure and errno is set.
419  */
420 int
421 priv_get_mtu(struct priv *priv, uint16_t *mtu)
422 {
423 	unsigned long ulong_mtu;
424 
425 	if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
426 		return -1;
427 	*mtu = ulong_mtu;
428 	return 0;
429 }
430 
431 /**
432  * Read device counter from sysfs.
433  *
434  * @param priv
435  *   Pointer to private structure.
436  * @param name
437  *   Counter name.
438  * @param[out] cntr
439  *   Counter output buffer.
440  *
441  * @return
442  *   0 on success, -1 on failure and errno is set.
443  */
444 int
445 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr)
446 {
447 	unsigned long ulong_ctr;
448 
449 	if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1)
450 		return -1;
451 	*cntr = ulong_ctr;
452 	return 0;
453 }
454 
455 /**
456  * Set device MTU.
457  *
458  * @param priv
459  *   Pointer to private structure.
460  * @param mtu
461  *   MTU value to set.
462  *
463  * @return
464  *   0 on success, -1 on failure and errno is set.
465  */
466 static int
467 priv_set_mtu(struct priv *priv, uint16_t mtu)
468 {
469 	uint16_t new_mtu;
470 
471 	if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
472 	    priv_get_mtu(priv, &new_mtu))
473 		return -1;
474 	if (new_mtu == mtu)
475 		return 0;
476 	errno = EINVAL;
477 	return -1;
478 }
479 
480 /**
481  * Set device flags.
482  *
483  * @param priv
484  *   Pointer to private structure.
485  * @param keep
486  *   Bitmask for flags that must remain untouched.
487  * @param flags
488  *   Bitmask for flags to modify.
489  *
490  * @return
491  *   0 on success, -1 on failure and errno is set.
492  */
493 int
494 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
495 {
496 	unsigned long tmp;
497 
498 	if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
499 		return -1;
500 	tmp &= keep;
501 	tmp |= (flags & (~keep));
502 	return priv_set_sysfs_ulong(priv, "flags", tmp);
503 }
504 
505 /**
506  * Ethernet device configuration.
507  *
508  * Prepare the driver for a given number of TX and RX queues.
509  *
510  * @param dev
511  *   Pointer to Ethernet device structure.
512  *
513  * @return
514  *   0 on success, errno value on failure.
515  */
516 static int
517 dev_configure(struct rte_eth_dev *dev)
518 {
519 	struct priv *priv = dev->data->dev_private;
520 	unsigned int rxqs_n = dev->data->nb_rx_queues;
521 	unsigned int txqs_n = dev->data->nb_tx_queues;
522 	unsigned int i;
523 	unsigned int j;
524 	unsigned int reta_idx_n;
525 	const uint8_t use_app_rss_key =
526 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
527 	uint64_t supp_tx_offloads = mlx5_priv_get_tx_port_offloads(priv);
528 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
529 	uint64_t supp_rx_offloads =
530 		(mlx5_priv_get_rx_port_offloads(priv) |
531 		 mlx5_priv_get_rx_queue_offloads(priv));
532 	uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
533 
534 	if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
535 		ERROR("Some Tx offloads are not supported "
536 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
537 		      tx_offloads, supp_tx_offloads);
538 		return ENOTSUP;
539 	}
540 	if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
541 		ERROR("Some Rx offloads are not supported "
542 		      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
543 		      rx_offloads, supp_rx_offloads);
544 		return ENOTSUP;
545 	}
546 	if (use_app_rss_key &&
547 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
548 	     rss_hash_default_key_len)) {
549 		/* MLX5 RSS only support 40bytes key. */
550 		return EINVAL;
551 	}
552 	priv->rss_conf.rss_key =
553 		rte_realloc(priv->rss_conf.rss_key,
554 			    rss_hash_default_key_len, 0);
555 	if (!priv->rss_conf.rss_key) {
556 		ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
557 		return ENOMEM;
558 	}
559 	memcpy(priv->rss_conf.rss_key,
560 	       use_app_rss_key ?
561 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
562 	       rss_hash_default_key,
563 	       rss_hash_default_key_len);
564 	priv->rss_conf.rss_key_len = rss_hash_default_key_len;
565 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
566 	priv->rxqs = (void *)dev->data->rx_queues;
567 	priv->txqs = (void *)dev->data->tx_queues;
568 	if (txqs_n != priv->txqs_n) {
569 		INFO("%p: TX queues number update: %u -> %u",
570 		     (void *)dev, priv->txqs_n, txqs_n);
571 		priv->txqs_n = txqs_n;
572 	}
573 	if (rxqs_n > priv->config.ind_table_max_size) {
574 		ERROR("cannot handle this many RX queues (%u)", rxqs_n);
575 		return EINVAL;
576 	}
577 	if (rxqs_n == priv->rxqs_n)
578 		return 0;
579 	INFO("%p: RX queues number update: %u -> %u",
580 	     (void *)dev, priv->rxqs_n, rxqs_n);
581 	priv->rxqs_n = rxqs_n;
582 	/* If the requested number of RX queues is not a power of two, use the
583 	 * maximum indirection table size for better balancing.
584 	 * The result is always rounded to the next power of two. */
585 	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
586 				     priv->config.ind_table_max_size :
587 				     rxqs_n));
588 	if (priv_rss_reta_index_resize(priv, reta_idx_n))
589 		return ENOMEM;
590 	/* When the number of RX queues is not a power of two, the remaining
591 	 * table entries are padded with reused WQs and hashes are not spread
592 	 * uniformly. */
593 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
594 		(*priv->reta_idx)[i] = j;
595 		if (++j == rxqs_n)
596 			j = 0;
597 	}
598 	return 0;
599 }
600 
601 /**
602  * DPDK callback for Ethernet device configuration.
603  *
604  * @param dev
605  *   Pointer to Ethernet device structure.
606  *
607  * @return
608  *   0 on success, negative errno value on failure.
609  */
610 int
611 mlx5_dev_configure(struct rte_eth_dev *dev)
612 {
613 	struct priv *priv = dev->data->dev_private;
614 	int ret;
615 
616 	priv_lock(priv);
617 	ret = dev_configure(dev);
618 	assert(ret >= 0);
619 	priv_unlock(priv);
620 	return -ret;
621 }
622 
623 /**
624  * DPDK callback to get information about the device.
625  *
626  * @param dev
627  *   Pointer to Ethernet device structure.
628  * @param[out] info
629  *   Info structure output buffer.
630  */
631 void
632 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
633 {
634 	struct priv *priv = dev->data->dev_private;
635 	struct mlx5_dev_config *config = &priv->config;
636 	unsigned int max;
637 	char ifname[IF_NAMESIZE];
638 
639 	info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
640 
641 	priv_lock(priv);
642 	/* FIXME: we should ask the device for these values. */
643 	info->min_rx_bufsize = 32;
644 	info->max_rx_pktlen = 65536;
645 	/*
646 	 * Since we need one CQ per QP, the limit is the minimum number
647 	 * between the two values.
648 	 */
649 	max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
650 		      priv->device_attr.orig_attr.max_qp);
651 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
652 	if (max >= 65535)
653 		max = 65535;
654 	info->max_rx_queues = max;
655 	info->max_tx_queues = max;
656 	info->max_mac_addrs = RTE_DIM(priv->mac);
657 	info->rx_queue_offload_capa =
658 		mlx5_priv_get_rx_queue_offloads(priv);
659 	info->rx_offload_capa = (mlx5_priv_get_rx_port_offloads(priv) |
660 				 info->rx_queue_offload_capa);
661 	info->tx_offload_capa = mlx5_priv_get_tx_port_offloads(priv);
662 	if (priv_get_ifname(priv, &ifname) == 0)
663 		info->if_index = if_nametoindex(ifname);
664 	info->reta_size = priv->reta_idx_n ?
665 		priv->reta_idx_n : config->ind_table_max_size;
666 	info->hash_key_size = priv->rss_conf.rss_key_len;
667 	info->speed_capa = priv->link_speed_capa;
668 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
669 	priv_unlock(priv);
670 }
671 
672 const uint32_t *
673 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
674 {
675 	static const uint32_t ptypes[] = {
676 		/* refers to rxq_cq_to_pkt_type() */
677 		RTE_PTYPE_L2_ETHER,
678 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
679 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
680 		RTE_PTYPE_L4_NONFRAG,
681 		RTE_PTYPE_L4_FRAG,
682 		RTE_PTYPE_L4_TCP,
683 		RTE_PTYPE_L4_UDP,
684 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
685 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
686 		RTE_PTYPE_INNER_L4_NONFRAG,
687 		RTE_PTYPE_INNER_L4_FRAG,
688 		RTE_PTYPE_INNER_L4_TCP,
689 		RTE_PTYPE_INNER_L4_UDP,
690 		RTE_PTYPE_UNKNOWN
691 	};
692 
693 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
694 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
695 		return ptypes;
696 	return NULL;
697 }
698 
699 /**
700  * DPDK callback to retrieve physical link information.
701  *
702  * @param dev
703  *   Pointer to Ethernet device structure.
704  * @param wait_to_complete
705  *   Wait for request completion (ignored).
706  */
707 static int
708 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
709 {
710 	struct priv *priv = dev->data->dev_private;
711 	struct ethtool_cmd edata = {
712 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
713 	};
714 	struct ifreq ifr;
715 	struct rte_eth_link dev_link;
716 	int link_speed = 0;
717 
718 	/* priv_lock() is not taken to allow concurrent calls. */
719 
720 	(void)wait_to_complete;
721 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
722 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
723 		return -1;
724 	}
725 	memset(&dev_link, 0, sizeof(dev_link));
726 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
727 				(ifr.ifr_flags & IFF_RUNNING));
728 	ifr.ifr_data = (void *)&edata;
729 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
730 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
731 		     strerror(errno));
732 		return -1;
733 	}
734 	link_speed = ethtool_cmd_speed(&edata);
735 	if (link_speed == -1)
736 		dev_link.link_speed = 0;
737 	else
738 		dev_link.link_speed = link_speed;
739 	priv->link_speed_capa = 0;
740 	if (edata.supported & SUPPORTED_Autoneg)
741 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
742 	if (edata.supported & (SUPPORTED_1000baseT_Full |
743 			       SUPPORTED_1000baseKX_Full))
744 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
745 	if (edata.supported & SUPPORTED_10000baseKR_Full)
746 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
747 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
748 			       SUPPORTED_40000baseCR4_Full |
749 			       SUPPORTED_40000baseSR4_Full |
750 			       SUPPORTED_40000baseLR4_Full))
751 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
752 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
753 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
754 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
755 			ETH_LINK_SPEED_FIXED);
756 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
757 		/* Link status changed. */
758 		dev->data->dev_link = dev_link;
759 		return 0;
760 	}
761 	/* Link status is still the same. */
762 	return -1;
763 }
764 
765 /**
766  * Retrieve physical link information (unlocked version using new ioctl).
767  *
768  * @param dev
769  *   Pointer to Ethernet device structure.
770  * @param wait_to_complete
771  *   Wait for request completion (ignored).
772  */
773 static int
774 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
775 {
776 	struct priv *priv = dev->data->dev_private;
777 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
778 	struct ifreq ifr;
779 	struct rte_eth_link dev_link;
780 	uint64_t sc;
781 
782 	(void)wait_to_complete;
783 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
784 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
785 		return -1;
786 	}
787 	memset(&dev_link, 0, sizeof(dev_link));
788 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
789 				(ifr.ifr_flags & IFF_RUNNING));
790 	ifr.ifr_data = (void *)&gcmd;
791 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
792 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
793 		      strerror(errno));
794 		return -1;
795 	}
796 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
797 
798 	alignas(struct ethtool_link_settings)
799 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
800 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
801 	struct ethtool_link_settings *ecmd = (void *)data;
802 
803 	*ecmd = gcmd;
804 	ifr.ifr_data = (void *)ecmd;
805 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
806 		DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
807 		      strerror(errno));
808 		return -1;
809 	}
810 	dev_link.link_speed = ecmd->speed;
811 	sc = ecmd->link_mode_masks[0] |
812 		((uint64_t)ecmd->link_mode_masks[1] << 32);
813 	priv->link_speed_capa = 0;
814 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
815 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
816 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
817 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
818 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
819 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
820 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
821 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
822 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
823 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
824 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
825 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
826 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
827 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
828 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
829 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
830 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
831 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
832 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
833 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
834 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
835 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
836 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
837 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
838 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
839 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
840 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
841 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
842 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
843 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
844 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
845 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
846 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
847 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
848 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
849 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
850 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
851 				  ETH_LINK_SPEED_FIXED);
852 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
853 		/* Link status changed. */
854 		dev->data->dev_link = dev_link;
855 		return 0;
856 	}
857 	/* Link status is still the same. */
858 	return -1;
859 }
860 
861 /**
862  * Enable receiving and transmitting traffic.
863  *
864  * @param priv
865  *   Pointer to private structure.
866  */
867 static void
868 priv_link_start(struct priv *priv)
869 {
870 	struct rte_eth_dev *dev = priv->dev;
871 	int err;
872 
873 	dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
874 	dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
875 	err = priv_dev_traffic_enable(priv, dev);
876 	if (err)
877 		ERROR("%p: error occurred while configuring control flows: %s",
878 		      (void *)priv, strerror(err));
879 	err = priv_flow_start(priv, &priv->flows);
880 	if (err)
881 		ERROR("%p: error occurred while configuring flows: %s",
882 		      (void *)priv, strerror(err));
883 }
884 
885 /**
886  * Disable receiving and transmitting traffic.
887  *
888  * @param priv
889  *   Pointer to private structure.
890  */
891 static void
892 priv_link_stop(struct priv *priv)
893 {
894 	struct rte_eth_dev *dev = priv->dev;
895 
896 	priv_flow_stop(priv, &priv->flows);
897 	priv_dev_traffic_disable(priv, dev);
898 	dev->rx_pkt_burst = removed_rx_burst;
899 	dev->tx_pkt_burst = removed_tx_burst;
900 }
901 
902 /**
903  * Retrieve physical link information and update rx/tx_pkt_burst callbacks
904  * accordingly.
905  *
906  * @param priv
907  *   Pointer to private structure.
908  * @param wait_to_complete
909  *   Wait for request completion (ignored).
910  */
911 int
912 priv_link_update(struct priv *priv, int wait_to_complete)
913 {
914 	struct rte_eth_dev *dev = priv->dev;
915 	struct utsname utsname;
916 	int ver[3];
917 	int ret;
918 	struct rte_eth_link dev_link = dev->data->dev_link;
919 
920 	if (uname(&utsname) == -1 ||
921 	    sscanf(utsname.release, "%d.%d.%d",
922 		   &ver[0], &ver[1], &ver[2]) != 3 ||
923 	    KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
924 		ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
925 	else
926 		ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
927 	/* If lsc interrupt is disabled, should always be ready for traffic. */
928 	if (!dev->data->dev_conf.intr_conf.lsc) {
929 		priv_link_start(priv);
930 		return ret;
931 	}
932 	/* Re-select burst callbacks only if link status has been changed. */
933 	if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
934 		if (dev->data->dev_link.link_status == ETH_LINK_UP)
935 			priv_link_start(priv);
936 		else
937 			priv_link_stop(priv);
938 	}
939 	return ret;
940 }
941 
942 /**
943  * Querying the link status till it changes to the desired state.
944  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
945  *
946  * @param priv
947  *   Pointer to private structure.
948  * @param status
949  *   Link desired status.
950  *
951  * @return
952  *   0 on success, negative errno value on failure.
953  */
954 int
955 priv_force_link_status_change(struct priv *priv, int status)
956 {
957 	int try = 0;
958 
959 	while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
960 		priv_link_update(priv, 0);
961 		if (priv->dev->data->dev_link.link_status == status)
962 			return 0;
963 		try++;
964 		sleep(1);
965 	}
966 	return -EAGAIN;
967 }
968 
969 /**
970  * DPDK callback to retrieve physical link information.
971  *
972  * @param dev
973  *   Pointer to Ethernet device structure.
974  * @param wait_to_complete
975  *   Wait for request completion (ignored).
976  */
977 int
978 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
979 {
980 	struct priv *priv = dev->data->dev_private;
981 	int ret;
982 
983 	priv_lock(priv);
984 	ret = priv_link_update(priv, wait_to_complete);
985 	priv_unlock(priv);
986 	return ret;
987 }
988 
989 /**
990  * DPDK callback to change the MTU.
991  *
992  * @param dev
993  *   Pointer to Ethernet device structure.
994  * @param in_mtu
995  *   New MTU.
996  *
997  * @return
998  *   0 on success, negative errno value on failure.
999  */
1000 int
1001 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1002 {
1003 	struct priv *priv = dev->data->dev_private;
1004 	uint16_t kern_mtu;
1005 	int ret = 0;
1006 
1007 	priv_lock(priv);
1008 	ret = priv_get_mtu(priv, &kern_mtu);
1009 	if (ret)
1010 		goto out;
1011 	/* Set kernel interface MTU first. */
1012 	ret = priv_set_mtu(priv, mtu);
1013 	if (ret)
1014 		goto out;
1015 	ret = priv_get_mtu(priv, &kern_mtu);
1016 	if (ret)
1017 		goto out;
1018 	if (kern_mtu == mtu) {
1019 		priv->mtu = mtu;
1020 		DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
1021 	}
1022 	priv_unlock(priv);
1023 	return 0;
1024 out:
1025 	ret = errno;
1026 	WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
1027 	     strerror(ret));
1028 	priv_unlock(priv);
1029 	assert(ret >= 0);
1030 	return -ret;
1031 }
1032 
1033 /**
1034  * DPDK callback to get flow control status.
1035  *
1036  * @param dev
1037  *   Pointer to Ethernet device structure.
1038  * @param[out] fc_conf
1039  *   Flow control output buffer.
1040  *
1041  * @return
1042  *   0 on success, negative errno value on failure.
1043  */
1044 int
1045 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1046 {
1047 	struct priv *priv = dev->data->dev_private;
1048 	struct ifreq ifr;
1049 	struct ethtool_pauseparam ethpause = {
1050 		.cmd = ETHTOOL_GPAUSEPARAM
1051 	};
1052 	int ret;
1053 
1054 	ifr.ifr_data = (void *)&ethpause;
1055 	priv_lock(priv);
1056 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1057 		ret = errno;
1058 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1059 		     " failed: %s",
1060 		     strerror(ret));
1061 		goto out;
1062 	}
1063 
1064 	fc_conf->autoneg = ethpause.autoneg;
1065 	if (ethpause.rx_pause && ethpause.tx_pause)
1066 		fc_conf->mode = RTE_FC_FULL;
1067 	else if (ethpause.rx_pause)
1068 		fc_conf->mode = RTE_FC_RX_PAUSE;
1069 	else if (ethpause.tx_pause)
1070 		fc_conf->mode = RTE_FC_TX_PAUSE;
1071 	else
1072 		fc_conf->mode = RTE_FC_NONE;
1073 	ret = 0;
1074 
1075 out:
1076 	priv_unlock(priv);
1077 	assert(ret >= 0);
1078 	return -ret;
1079 }
1080 
1081 /**
1082  * DPDK callback to modify flow control parameters.
1083  *
1084  * @param dev
1085  *   Pointer to Ethernet device structure.
1086  * @param[in] fc_conf
1087  *   Flow control parameters.
1088  *
1089  * @return
1090  *   0 on success, negative errno value on failure.
1091  */
1092 int
1093 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1094 {
1095 	struct priv *priv = dev->data->dev_private;
1096 	struct ifreq ifr;
1097 	struct ethtool_pauseparam ethpause = {
1098 		.cmd = ETHTOOL_SPAUSEPARAM
1099 	};
1100 	int ret;
1101 
1102 	ifr.ifr_data = (void *)&ethpause;
1103 	ethpause.autoneg = fc_conf->autoneg;
1104 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1105 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1106 		ethpause.rx_pause = 1;
1107 	else
1108 		ethpause.rx_pause = 0;
1109 
1110 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1111 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1112 		ethpause.tx_pause = 1;
1113 	else
1114 		ethpause.tx_pause = 0;
1115 
1116 	priv_lock(priv);
1117 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1118 		ret = errno;
1119 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1120 		     " failed: %s",
1121 		     strerror(ret));
1122 		goto out;
1123 	}
1124 	ret = 0;
1125 
1126 out:
1127 	priv_unlock(priv);
1128 	assert(ret >= 0);
1129 	return -ret;
1130 }
1131 
1132 /**
1133  * Get PCI information from struct ibv_device.
1134  *
1135  * @param device
1136  *   Pointer to Ethernet device structure.
1137  * @param[out] pci_addr
1138  *   PCI bus address output buffer.
1139  *
1140  * @return
1141  *   0 on success, -1 on failure and errno is set.
1142  */
1143 int
1144 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1145 			    struct rte_pci_addr *pci_addr)
1146 {
1147 	FILE *file;
1148 	char line[32];
1149 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
1150 
1151 	file = fopen(path, "rb");
1152 	if (file == NULL)
1153 		return -1;
1154 	while (fgets(line, sizeof(line), file) == line) {
1155 		size_t len = strlen(line);
1156 		int ret;
1157 
1158 		/* Truncate long lines. */
1159 		if (len == (sizeof(line) - 1))
1160 			while (line[(len - 1)] != '\n') {
1161 				ret = fgetc(file);
1162 				if (ret == EOF)
1163 					break;
1164 				line[(len - 1)] = ret;
1165 			}
1166 		/* Extract information. */
1167 		if (sscanf(line,
1168 			   "PCI_SLOT_NAME="
1169 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1170 			   &pci_addr->domain,
1171 			   &pci_addr->bus,
1172 			   &pci_addr->devid,
1173 			   &pci_addr->function) == 4) {
1174 			ret = 0;
1175 			break;
1176 		}
1177 	}
1178 	fclose(file);
1179 	return 0;
1180 }
1181 
1182 /**
1183  * Update the link status.
1184  *
1185  * @param priv
1186  *   Pointer to private structure.
1187  *
1188  * @return
1189  *   Zero if the callback process can be called immediately.
1190  */
1191 static int
1192 priv_link_status_update(struct priv *priv)
1193 {
1194 	struct rte_eth_link *link = &priv->dev->data->dev_link;
1195 
1196 	priv_link_update(priv, 0);
1197 	if (((link->link_speed == 0) && link->link_status) ||
1198 		((link->link_speed != 0) && !link->link_status)) {
1199 		/*
1200 		 * Inconsistent status. Event likely occurred before the
1201 		 * kernel netdevice exposes the new status.
1202 		 */
1203 		if (!priv->pending_alarm) {
1204 			priv->pending_alarm = 1;
1205 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1206 					  mlx5_dev_link_status_handler,
1207 					  priv->dev);
1208 		}
1209 		return 1;
1210 	} else if (unlikely(priv->pending_alarm)) {
1211 		/* Link interrupt occurred while alarm is already scheduled. */
1212 		priv->pending_alarm = 0;
1213 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
1214 	}
1215 	return 0;
1216 }
1217 
1218 /**
1219  * Device status handler.
1220  *
1221  * @param priv
1222  *   Pointer to private structure.
1223  * @param events
1224  *   Pointer to event flags holder.
1225  *
1226  * @return
1227  *   Events bitmap of callback process which can be called immediately.
1228  */
1229 static uint32_t
1230 priv_dev_status_handler(struct priv *priv)
1231 {
1232 	struct ibv_async_event event;
1233 	uint32_t ret = 0;
1234 
1235 	/* Read all message and acknowledge them. */
1236 	for (;;) {
1237 		if (mlx5_glue->get_async_event(priv->ctx, &event))
1238 			break;
1239 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1240 			event.event_type == IBV_EVENT_PORT_ERR) &&
1241 			(priv->dev->data->dev_conf.intr_conf.lsc == 1))
1242 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1243 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1244 			priv->dev->data->dev_conf.intr_conf.rmv == 1)
1245 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1246 		else
1247 			DEBUG("event type %d on port %d not handled",
1248 			      event.event_type, event.element.port_num);
1249 		mlx5_glue->ack_async_event(&event);
1250 	}
1251 	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1252 		if (priv_link_status_update(priv))
1253 			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1254 	return ret;
1255 }
1256 
1257 /**
1258  * Handle delayed link status event.
1259  *
1260  * @param arg
1261  *   Registered argument.
1262  */
1263 void
1264 mlx5_dev_link_status_handler(void *arg)
1265 {
1266 	struct rte_eth_dev *dev = arg;
1267 	struct priv *priv = dev->data->dev_private;
1268 	int ret;
1269 
1270 	while (!priv_trylock(priv)) {
1271 		/* Alarm is being canceled. */
1272 		if (priv->pending_alarm == 0)
1273 			return;
1274 		rte_pause();
1275 	}
1276 	priv->pending_alarm = 0;
1277 	ret = priv_link_status_update(priv);
1278 	priv_unlock(priv);
1279 	if (!ret)
1280 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1281 }
1282 
1283 /**
1284  * Handle interrupts from the NIC.
1285  *
1286  * @param[in] intr_handle
1287  *   Interrupt handler.
1288  * @param cb_arg
1289  *   Callback argument.
1290  */
1291 void
1292 mlx5_dev_interrupt_handler(void *cb_arg)
1293 {
1294 	struct rte_eth_dev *dev = cb_arg;
1295 	struct priv *priv = dev->data->dev_private;
1296 	uint32_t events;
1297 
1298 	priv_lock(priv);
1299 	events = priv_dev_status_handler(priv);
1300 	priv_unlock(priv);
1301 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1302 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1303 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1304 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1305 }
1306 
1307 /**
1308  * Handle interrupts from the socket.
1309  *
1310  * @param cb_arg
1311  *   Callback argument.
1312  */
1313 static void
1314 mlx5_dev_handler_socket(void *cb_arg)
1315 {
1316 	struct rte_eth_dev *dev = cb_arg;
1317 	struct priv *priv = dev->data->dev_private;
1318 
1319 	priv_lock(priv);
1320 	priv_socket_handle(priv);
1321 	priv_unlock(priv);
1322 }
1323 
1324 /**
1325  * Uninstall interrupt handler.
1326  *
1327  * @param priv
1328  *   Pointer to private structure.
1329  * @param dev
1330  *   Pointer to the rte_eth_dev structure.
1331  */
1332 void
1333 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1334 {
1335 	if (dev->data->dev_conf.intr_conf.lsc ||
1336 	    dev->data->dev_conf.intr_conf.rmv)
1337 		rte_intr_callback_unregister(&priv->intr_handle,
1338 					     mlx5_dev_interrupt_handler, dev);
1339 	if (priv->primary_socket)
1340 		rte_intr_callback_unregister(&priv->intr_handle_socket,
1341 					     mlx5_dev_handler_socket, dev);
1342 	if (priv->pending_alarm) {
1343 		priv->pending_alarm = 0;
1344 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1345 	}
1346 	priv->intr_handle.fd = 0;
1347 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1348 	priv->intr_handle_socket.fd = 0;
1349 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1350 }
1351 
1352 /**
1353  * Install interrupt handler.
1354  *
1355  * @param priv
1356  *   Pointer to private structure.
1357  * @param dev
1358  *   Pointer to the rte_eth_dev structure.
1359  */
1360 void
1361 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1362 {
1363 	int rc, flags;
1364 
1365 	assert(priv->ctx->async_fd > 0);
1366 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1367 	rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1368 	if (rc < 0) {
1369 		INFO("failed to change file descriptor async event queue");
1370 		dev->data->dev_conf.intr_conf.lsc = 0;
1371 		dev->data->dev_conf.intr_conf.rmv = 0;
1372 	}
1373 	if (dev->data->dev_conf.intr_conf.lsc ||
1374 	    dev->data->dev_conf.intr_conf.rmv) {
1375 		priv->intr_handle.fd = priv->ctx->async_fd;
1376 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1377 		rte_intr_callback_register(&priv->intr_handle,
1378 					   mlx5_dev_interrupt_handler, dev);
1379 	}
1380 
1381 	rc = priv_socket_init(priv);
1382 	if (!rc && priv->primary_socket) {
1383 		priv->intr_handle_socket.fd = priv->primary_socket;
1384 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1385 		rte_intr_callback_register(&priv->intr_handle_socket,
1386 					   mlx5_dev_handler_socket, dev);
1387 	}
1388 }
1389 
1390 /**
1391  * Change the link state (UP / DOWN).
1392  *
1393  * @param priv
1394  *   Pointer to private data structure.
1395  * @param up
1396  *   Nonzero for link up, otherwise link down.
1397  *
1398  * @return
1399  *   0 on success, errno value on failure.
1400  */
1401 static int
1402 priv_dev_set_link(struct priv *priv, int up)
1403 {
1404 	return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1405 }
1406 
1407 /**
1408  * DPDK callback to bring the link DOWN.
1409  *
1410  * @param dev
1411  *   Pointer to Ethernet device structure.
1412  *
1413  * @return
1414  *   0 on success, errno value on failure.
1415  */
1416 int
1417 mlx5_set_link_down(struct rte_eth_dev *dev)
1418 {
1419 	struct priv *priv = dev->data->dev_private;
1420 	int err;
1421 
1422 	priv_lock(priv);
1423 	err = priv_dev_set_link(priv, 0);
1424 	priv_unlock(priv);
1425 	return err;
1426 }
1427 
1428 /**
1429  * DPDK callback to bring the link UP.
1430  *
1431  * @param dev
1432  *   Pointer to Ethernet device structure.
1433  *
1434  * @return
1435  *   0 on success, errno value on failure.
1436  */
1437 int
1438 mlx5_set_link_up(struct rte_eth_dev *dev)
1439 {
1440 	struct priv *priv = dev->data->dev_private;
1441 	int err;
1442 
1443 	priv_lock(priv);
1444 	err = priv_dev_set_link(priv, 1);
1445 	priv_unlock(priv);
1446 	return err;
1447 }
1448 
1449 /**
1450  * Configure the TX function to use.
1451  *
1452  * @param priv
1453  *   Pointer to private data structure.
1454  * @param dev
1455  *   Pointer to rte_eth_dev structure.
1456  *
1457  * @return
1458  *   Pointer to selected Tx burst function.
1459  */
1460 eth_tx_burst_t
1461 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1462 {
1463 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1464 	struct mlx5_dev_config *config = &priv->config;
1465 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1466 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1467 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1468 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1469 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1470 
1471 	assert(priv != NULL);
1472 	/* Select appropriate TX function. */
1473 	if (vlan_insert || tso)
1474 		return tx_pkt_burst;
1475 	if (config->mps == MLX5_MPW_ENHANCED) {
1476 		if (priv_check_vec_tx_support(priv, dev) > 0) {
1477 			if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1478 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1479 			else
1480 				tx_pkt_burst = mlx5_tx_burst_vec;
1481 			DEBUG("selected Enhanced MPW TX vectorized function");
1482 		} else {
1483 			tx_pkt_burst = mlx5_tx_burst_empw;
1484 			DEBUG("selected Enhanced MPW TX function");
1485 		}
1486 	} else if (config->mps && (config->txq_inline > 0)) {
1487 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1488 		DEBUG("selected MPW inline TX function");
1489 	} else if (config->mps) {
1490 		tx_pkt_burst = mlx5_tx_burst_mpw;
1491 		DEBUG("selected MPW TX function");
1492 	}
1493 	return tx_pkt_burst;
1494 }
1495 
1496 /**
1497  * Configure the RX function to use.
1498  *
1499  * @param priv
1500  *   Pointer to private data structure.
1501  * @param dev
1502  *   Pointer to rte_eth_dev structure.
1503  *
1504  * @return
1505  *   Pointer to selected Rx burst function.
1506  */
1507 eth_rx_burst_t
1508 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1509 {
1510 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1511 
1512 	assert(priv != NULL);
1513 	if (priv_check_vec_rx_support(priv) > 0) {
1514 		rx_pkt_burst = mlx5_rx_burst_vec;
1515 		DEBUG("selected RX vectorized function");
1516 	}
1517 	return rx_pkt_burst;
1518 }
1519 
1520 /**
1521  * Check if mlx5 device was removed.
1522  *
1523  * @param dev
1524  *   Pointer to Ethernet device structure.
1525  *
1526  * @return
1527  *   1 when device is removed, otherwise 0.
1528  */
1529 int
1530 mlx5_is_removed(struct rte_eth_dev *dev)
1531 {
1532 	struct ibv_device_attr device_attr;
1533 	struct priv *priv = dev->data->dev_private;
1534 
1535 	if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1536 		return 1;
1537 	return 0;
1538 }
1539