xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 198a3c339a8ff9ecc5916566cfd25e98d98ce484)
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 #include <stddef.h>
35 #include <assert.h>
36 #include <unistd.h>
37 #include <stdint.h>
38 #include <stdio.h>
39 #include <string.h>
40 #include <stdlib.h>
41 #include <errno.h>
42 #include <dirent.h>
43 #include <net/if.h>
44 #include <sys/ioctl.h>
45 #include <sys/socket.h>
46 #include <netinet/in.h>
47 #include <linux/if.h>
48 #include <linux/ethtool.h>
49 #include <linux/sockios.h>
50 #include <fcntl.h>
51 
52 /* DPDK headers don't like -pedantic. */
53 #ifdef PEDANTIC
54 #pragma GCC diagnostic ignored "-pedantic"
55 #endif
56 #include <rte_atomic.h>
57 #include <rte_ethdev.h>
58 #include <rte_mbuf.h>
59 #include <rte_common.h>
60 #include <rte_interrupts.h>
61 #include <rte_alarm.h>
62 #ifdef PEDANTIC
63 #pragma GCC diagnostic error "-pedantic"
64 #endif
65 
66 #include "mlx5.h"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_utils.h"
69 
70 /**
71  * Get interface name from private structure.
72  *
73  * @param[in] priv
74  *   Pointer to private structure.
75  * @param[out] ifname
76  *   Interface name output buffer.
77  *
78  * @return
79  *   0 on success, -1 on failure and errno is set.
80  */
81 int
82 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
83 {
84 	DIR *dir;
85 	struct dirent *dent;
86 	unsigned int dev_type = 0;
87 	unsigned int dev_port_prev = ~0u;
88 	char match[IF_NAMESIZE] = "";
89 
90 	{
91 		MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
92 
93 		dir = opendir(path);
94 		if (dir == NULL)
95 			return -1;
96 	}
97 	while ((dent = readdir(dir)) != NULL) {
98 		char *name = dent->d_name;
99 		FILE *file;
100 		unsigned int dev_port;
101 		int r;
102 
103 		if ((name[0] == '.') &&
104 		    ((name[1] == '\0') ||
105 		     ((name[1] == '.') && (name[2] == '\0'))))
106 			continue;
107 
108 		MKSTR(path, "%s/device/net/%s/%s",
109 		      priv->ctx->device->ibdev_path, name,
110 		      (dev_type ? "dev_id" : "dev_port"));
111 
112 		file = fopen(path, "rb");
113 		if (file == NULL) {
114 			if (errno != ENOENT)
115 				continue;
116 			/*
117 			 * Switch to dev_id when dev_port does not exist as
118 			 * is the case with Linux kernel versions < 3.15.
119 			 */
120 try_dev_id:
121 			match[0] = '\0';
122 			if (dev_type)
123 				break;
124 			dev_type = 1;
125 			dev_port_prev = ~0u;
126 			rewinddir(dir);
127 			continue;
128 		}
129 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
130 		fclose(file);
131 		if (r != 1)
132 			continue;
133 		/*
134 		 * Switch to dev_id when dev_port returns the same value for
135 		 * all ports. May happen when using a MOFED release older than
136 		 * 3.0 with a Linux kernel >= 3.15.
137 		 */
138 		if (dev_port == dev_port_prev)
139 			goto try_dev_id;
140 		dev_port_prev = dev_port;
141 		if (dev_port == (priv->port - 1u))
142 			snprintf(match, sizeof(match), "%s", name);
143 	}
144 	closedir(dir);
145 	if (match[0] == '\0')
146 		return -1;
147 	strncpy(*ifname, match, sizeof(*ifname));
148 	return 0;
149 }
150 
151 /**
152  * Read from sysfs entry.
153  *
154  * @param[in] priv
155  *   Pointer to private structure.
156  * @param[in] entry
157  *   Entry name relative to sysfs path.
158  * @param[out] buf
159  *   Data output buffer.
160  * @param size
161  *   Buffer size.
162  *
163  * @return
164  *   0 on success, -1 on failure and errno is set.
165  */
166 static int
167 priv_sysfs_read(const struct priv *priv, const char *entry,
168 		char *buf, size_t size)
169 {
170 	char ifname[IF_NAMESIZE];
171 	FILE *file;
172 	int ret;
173 	int err;
174 
175 	if (priv_get_ifname(priv, &ifname))
176 		return -1;
177 
178 	MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
179 	      ifname, entry);
180 
181 	file = fopen(path, "rb");
182 	if (file == NULL)
183 		return -1;
184 	ret = fread(buf, 1, size, file);
185 	err = errno;
186 	if (((size_t)ret < size) && (ferror(file)))
187 		ret = -1;
188 	else
189 		ret = size;
190 	fclose(file);
191 	errno = err;
192 	return ret;
193 }
194 
195 /**
196  * Write to sysfs entry.
197  *
198  * @param[in] priv
199  *   Pointer to private structure.
200  * @param[in] entry
201  *   Entry name relative to sysfs path.
202  * @param[in] buf
203  *   Data buffer.
204  * @param size
205  *   Buffer size.
206  *
207  * @return
208  *   0 on success, -1 on failure and errno is set.
209  */
210 static int
211 priv_sysfs_write(const struct priv *priv, const char *entry,
212 		 char *buf, size_t size)
213 {
214 	char ifname[IF_NAMESIZE];
215 	FILE *file;
216 	int ret;
217 	int err;
218 
219 	if (priv_get_ifname(priv, &ifname))
220 		return -1;
221 
222 	MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
223 	      ifname, entry);
224 
225 	file = fopen(path, "wb");
226 	if (file == NULL)
227 		return -1;
228 	ret = fwrite(buf, 1, size, file);
229 	err = errno;
230 	if (((size_t)ret < size) || (ferror(file)))
231 		ret = -1;
232 	else
233 		ret = size;
234 	fclose(file);
235 	errno = err;
236 	return ret;
237 }
238 
239 /**
240  * Get unsigned long sysfs property.
241  *
242  * @param priv
243  *   Pointer to private structure.
244  * @param[in] name
245  *   Entry name relative to sysfs path.
246  * @param[out] value
247  *   Value output buffer.
248  *
249  * @return
250  *   0 on success, -1 on failure and errno is set.
251  */
252 static int
253 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
254 {
255 	int ret;
256 	unsigned long value_ret;
257 	char value_str[32];
258 
259 	ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
260 	if (ret == -1) {
261 		DEBUG("cannot read %s value from sysfs: %s",
262 		      name, strerror(errno));
263 		return -1;
264 	}
265 	value_str[ret] = '\0';
266 	errno = 0;
267 	value_ret = strtoul(value_str, NULL, 0);
268 	if (errno) {
269 		DEBUG("invalid %s value `%s': %s", name, value_str,
270 		      strerror(errno));
271 		return -1;
272 	}
273 	*value = value_ret;
274 	return 0;
275 }
276 
277 /**
278  * Set unsigned long sysfs property.
279  *
280  * @param priv
281  *   Pointer to private structure.
282  * @param[in] name
283  *   Entry name relative to sysfs path.
284  * @param value
285  *   Value to set.
286  *
287  * @return
288  *   0 on success, -1 on failure and errno is set.
289  */
290 static int
291 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
292 {
293 	int ret;
294 	MKSTR(value_str, "%lu", value);
295 
296 	ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
297 	if (ret == -1) {
298 		DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
299 		      name, value_str, value, strerror(errno));
300 		return -1;
301 	}
302 	return 0;
303 }
304 
305 /**
306  * Perform ifreq ioctl() on associated Ethernet device.
307  *
308  * @param[in] priv
309  *   Pointer to private structure.
310  * @param req
311  *   Request number to pass to ioctl().
312  * @param[out] ifr
313  *   Interface request structure output buffer.
314  *
315  * @return
316  *   0 on success, -1 on failure and errno is set.
317  */
318 int
319 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
320 {
321 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
322 	int ret = -1;
323 
324 	if (sock == -1)
325 		return ret;
326 	if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
327 		ret = ioctl(sock, req, ifr);
328 	close(sock);
329 	return ret;
330 }
331 
332 /**
333  * Get device MTU.
334  *
335  * @param priv
336  *   Pointer to private structure.
337  * @param[out] mtu
338  *   MTU value output buffer.
339  *
340  * @return
341  *   0 on success, -1 on failure and errno is set.
342  */
343 int
344 priv_get_mtu(struct priv *priv, uint16_t *mtu)
345 {
346 	unsigned long ulong_mtu;
347 
348 	if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
349 		return -1;
350 	*mtu = ulong_mtu;
351 	return 0;
352 }
353 
354 /**
355  * Set device MTU.
356  *
357  * @param priv
358  *   Pointer to private structure.
359  * @param mtu
360  *   MTU value to set.
361  *
362  * @return
363  *   0 on success, -1 on failure and errno is set.
364  */
365 static int
366 priv_set_mtu(struct priv *priv, uint16_t mtu)
367 {
368 	return priv_set_sysfs_ulong(priv, "mtu", mtu);
369 }
370 
371 /**
372  * Set device flags.
373  *
374  * @param priv
375  *   Pointer to private structure.
376  * @param keep
377  *   Bitmask for flags that must remain untouched.
378  * @param flags
379  *   Bitmask for flags to modify.
380  *
381  * @return
382  *   0 on success, -1 on failure and errno is set.
383  */
384 int
385 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
386 {
387 	unsigned long tmp;
388 
389 	if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
390 		return -1;
391 	tmp &= keep;
392 	tmp |= flags;
393 	return priv_set_sysfs_ulong(priv, "flags", tmp);
394 }
395 
396 /**
397  * Ethernet device configuration.
398  *
399  * Prepare the driver for a given number of TX and RX queues.
400  *
401  * @param dev
402  *   Pointer to Ethernet device structure.
403  *
404  * @return
405  *   0 on success, errno value on failure.
406  */
407 static int
408 dev_configure(struct rte_eth_dev *dev)
409 {
410 	struct priv *priv = dev->data->dev_private;
411 	unsigned int rxqs_n = dev->data->nb_rx_queues;
412 	unsigned int txqs_n = dev->data->nb_tx_queues;
413 
414 	priv->rxqs = (void *)dev->data->rx_queues;
415 	priv->txqs = (void *)dev->data->tx_queues;
416 	if (txqs_n != priv->txqs_n) {
417 		INFO("%p: TX queues number update: %u -> %u",
418 		     (void *)dev, priv->txqs_n, txqs_n);
419 		priv->txqs_n = txqs_n;
420 	}
421 	if (rxqs_n == priv->rxqs_n)
422 		return 0;
423 	INFO("%p: RX queues number update: %u -> %u",
424 	     (void *)dev, priv->rxqs_n, rxqs_n);
425 	priv->rxqs_n = rxqs_n;
426 	return 0;
427 }
428 
429 /**
430  * DPDK callback for Ethernet device configuration.
431  *
432  * @param dev
433  *   Pointer to Ethernet device structure.
434  *
435  * @return
436  *   0 on success, negative errno value on failure.
437  */
438 int
439 mlx5_dev_configure(struct rte_eth_dev *dev)
440 {
441 	struct priv *priv = dev->data->dev_private;
442 	int ret;
443 
444 	priv_lock(priv);
445 	ret = dev_configure(dev);
446 	assert(ret >= 0);
447 	priv_unlock(priv);
448 	return -ret;
449 }
450 
451 /**
452  * DPDK callback to get information about the device.
453  *
454  * @param dev
455  *   Pointer to Ethernet device structure.
456  * @param[out] info
457  *   Info structure output buffer.
458  */
459 void
460 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
461 {
462 	struct priv *priv = dev->data->dev_private;
463 	unsigned int max;
464 	char ifname[IF_NAMESIZE];
465 
466 	priv_lock(priv);
467 	/* FIXME: we should ask the device for these values. */
468 	info->min_rx_bufsize = 32;
469 	info->max_rx_pktlen = 65536;
470 	/*
471 	 * Since we need one CQ per QP, the limit is the minimum number
472 	 * between the two values.
473 	 */
474 	max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
475 	       priv->device_attr.max_qp : priv->device_attr.max_cq);
476 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
477 	if (max >= 65535)
478 		max = 65535;
479 	info->max_rx_queues = max;
480 	info->max_tx_queues = max;
481 	/* Last array entry is reserved for broadcast. */
482 	info->max_mac_addrs = (RTE_DIM(priv->mac) - 1);
483 	info->rx_offload_capa =
484 		(priv->hw_csum ?
485 		 (DEV_RX_OFFLOAD_IPV4_CKSUM |
486 		  DEV_RX_OFFLOAD_UDP_CKSUM |
487 		  DEV_RX_OFFLOAD_TCP_CKSUM) :
488 		 0);
489 	info->tx_offload_capa =
490 		(priv->hw_csum ?
491 		 (DEV_TX_OFFLOAD_IPV4_CKSUM |
492 		  DEV_TX_OFFLOAD_UDP_CKSUM |
493 		  DEV_TX_OFFLOAD_TCP_CKSUM) :
494 		 0);
495 	if (priv_get_ifname(priv, &ifname) == 0)
496 		info->if_index = if_nametoindex(ifname);
497 	priv_unlock(priv);
498 }
499 
500 /**
501  * DPDK callback to retrieve physical link information (unlocked version).
502  *
503  * @param dev
504  *   Pointer to Ethernet device structure.
505  * @param wait_to_complete
506  *   Wait for request completion (ignored).
507  */
508 static int
509 mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
510 {
511 	struct priv *priv = dev->data->dev_private;
512 	struct ethtool_cmd edata = {
513 		.cmd = ETHTOOL_GSET
514 	};
515 	struct ifreq ifr;
516 	struct rte_eth_link dev_link;
517 	int link_speed = 0;
518 
519 	(void)wait_to_complete;
520 	if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
521 		WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
522 		return -1;
523 	}
524 	memset(&dev_link, 0, sizeof(dev_link));
525 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
526 				(ifr.ifr_flags & IFF_RUNNING));
527 	ifr.ifr_data = &edata;
528 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
529 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
530 		     strerror(errno));
531 		return -1;
532 	}
533 	link_speed = ethtool_cmd_speed(&edata);
534 	if (link_speed == -1)
535 		dev_link.link_speed = 0;
536 	else
537 		dev_link.link_speed = link_speed;
538 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
539 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
540 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
541 		/* Link status changed. */
542 		dev->data->dev_link = dev_link;
543 		return 0;
544 	}
545 	/* Link status is still the same. */
546 	return -1;
547 }
548 
549 /**
550  * DPDK callback to retrieve physical link information.
551  *
552  * @param dev
553  *   Pointer to Ethernet device structure.
554  * @param wait_to_complete
555  *   Wait for request completion (ignored).
556  */
557 int
558 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
559 {
560 	struct priv *priv = dev->data->dev_private;
561 	int ret;
562 
563 	priv_lock(priv);
564 	ret = mlx5_link_update_unlocked(dev, wait_to_complete);
565 	priv_unlock(priv);
566 	return ret;
567 }
568 
569 /**
570  * DPDK callback to change the MTU.
571  *
572  * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
573  * received). Use this as a hint to enable/disable scattered packets support
574  * and improve performance when not needed.
575  * Since failure is not an option, reconfiguring queues on the fly is not
576  * recommended.
577  *
578  * @param dev
579  *   Pointer to Ethernet device structure.
580  * @param in_mtu
581  *   New MTU.
582  *
583  * @return
584  *   0 on success, negative errno value on failure.
585  */
586 int
587 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
588 {
589 	struct priv *priv = dev->data->dev_private;
590 	int ret = 0;
591 	unsigned int i;
592 	uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
593 		mlx5_rx_burst;
594 
595 	priv_lock(priv);
596 	/* Set kernel interface MTU first. */
597 	if (priv_set_mtu(priv, mtu)) {
598 		ret = errno;
599 		WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
600 		     strerror(ret));
601 		goto out;
602 	} else
603 		DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
604 	priv->mtu = mtu;
605 	/* Temporarily replace RX handler with a fake one, assuming it has not
606 	 * been copied elsewhere. */
607 	dev->rx_pkt_burst = removed_rx_burst;
608 	/* Make sure everyone has left mlx5_rx_burst() and uses
609 	 * removed_rx_burst() instead. */
610 	rte_wmb();
611 	usleep(1000);
612 	/* Reconfigure each RX queue. */
613 	for (i = 0; (i != priv->rxqs_n); ++i) {
614 		struct rxq *rxq = (*priv->rxqs)[i];
615 		unsigned int max_frame_len;
616 		int sp;
617 
618 		if (rxq == NULL)
619 			continue;
620 		/* Calculate new maximum frame length according to MTU and
621 		 * toggle scattered support (sp) if necessary. */
622 		max_frame_len = (priv->mtu + ETHER_HDR_LEN +
623 				 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
624 		sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
625 		/* Provide new values to rxq_setup(). */
626 		dev->data->dev_conf.rxmode.jumbo_frame = sp;
627 		dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
628 		ret = rxq_rehash(dev, rxq);
629 		if (ret) {
630 			/* Force SP RX if that queue requires it and abort. */
631 			if (rxq->sp)
632 				rx_func = mlx5_rx_burst_sp;
633 			break;
634 		}
635 		/* Scattered burst function takes priority. */
636 		if (rxq->sp)
637 			rx_func = mlx5_rx_burst_sp;
638 	}
639 	/* Burst functions can now be called again. */
640 	rte_wmb();
641 	dev->rx_pkt_burst = rx_func;
642 out:
643 	priv_unlock(priv);
644 	assert(ret >= 0);
645 	return -ret;
646 }
647 
648 /**
649  * DPDK callback to get flow control status.
650  *
651  * @param dev
652  *   Pointer to Ethernet device structure.
653  * @param[out] fc_conf
654  *   Flow control output buffer.
655  *
656  * @return
657  *   0 on success, negative errno value on failure.
658  */
659 int
660 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
661 {
662 	struct priv *priv = dev->data->dev_private;
663 	struct ifreq ifr;
664 	struct ethtool_pauseparam ethpause = {
665 		.cmd = ETHTOOL_GPAUSEPARAM
666 	};
667 	int ret;
668 
669 	ifr.ifr_data = &ethpause;
670 	priv_lock(priv);
671 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
672 		ret = errno;
673 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
674 		     " failed: %s",
675 		     strerror(ret));
676 		goto out;
677 	}
678 
679 	fc_conf->autoneg = ethpause.autoneg;
680 	if (ethpause.rx_pause && ethpause.tx_pause)
681 		fc_conf->mode = RTE_FC_FULL;
682 	else if (ethpause.rx_pause)
683 		fc_conf->mode = RTE_FC_RX_PAUSE;
684 	else if (ethpause.tx_pause)
685 		fc_conf->mode = RTE_FC_TX_PAUSE;
686 	else
687 		fc_conf->mode = RTE_FC_NONE;
688 	ret = 0;
689 
690 out:
691 	priv_unlock(priv);
692 	assert(ret >= 0);
693 	return -ret;
694 }
695 
696 /**
697  * DPDK callback to modify flow control parameters.
698  *
699  * @param dev
700  *   Pointer to Ethernet device structure.
701  * @param[in] fc_conf
702  *   Flow control parameters.
703  *
704  * @return
705  *   0 on success, negative errno value on failure.
706  */
707 int
708 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
709 {
710 	struct priv *priv = dev->data->dev_private;
711 	struct ifreq ifr;
712 	struct ethtool_pauseparam ethpause = {
713 		.cmd = ETHTOOL_SPAUSEPARAM
714 	};
715 	int ret;
716 
717 	ifr.ifr_data = &ethpause;
718 	ethpause.autoneg = fc_conf->autoneg;
719 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
720 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
721 		ethpause.rx_pause = 1;
722 	else
723 		ethpause.rx_pause = 0;
724 
725 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
726 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
727 		ethpause.tx_pause = 1;
728 	else
729 		ethpause.tx_pause = 0;
730 
731 	priv_lock(priv);
732 	if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
733 		ret = errno;
734 		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
735 		     " failed: %s",
736 		     strerror(ret));
737 		goto out;
738 	}
739 	ret = 0;
740 
741 out:
742 	priv_unlock(priv);
743 	assert(ret >= 0);
744 	return -ret;
745 }
746 
747 /**
748  * Get PCI information from struct ibv_device.
749  *
750  * @param device
751  *   Pointer to Ethernet device structure.
752  * @param[out] pci_addr
753  *   PCI bus address output buffer.
754  *
755  * @return
756  *   0 on success, -1 on failure and errno is set.
757  */
758 int
759 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
760 			    struct rte_pci_addr *pci_addr)
761 {
762 	FILE *file;
763 	char line[32];
764 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
765 
766 	file = fopen(path, "rb");
767 	if (file == NULL)
768 		return -1;
769 	while (fgets(line, sizeof(line), file) == line) {
770 		size_t len = strlen(line);
771 		int ret;
772 
773 		/* Truncate long lines. */
774 		if (len == (sizeof(line) - 1))
775 			while (line[(len - 1)] != '\n') {
776 				ret = fgetc(file);
777 				if (ret == EOF)
778 					break;
779 				line[(len - 1)] = ret;
780 			}
781 		/* Extract information. */
782 		if (sscanf(line,
783 			   "PCI_SLOT_NAME="
784 			   "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
785 			   &pci_addr->domain,
786 			   &pci_addr->bus,
787 			   &pci_addr->devid,
788 			   &pci_addr->function) == 4) {
789 			ret = 0;
790 			break;
791 		}
792 	}
793 	fclose(file);
794 	return 0;
795 }
796 
797 /**
798  * Link status handler.
799  *
800  * @param priv
801  *   Pointer to private structure.
802  * @param dev
803  *   Pointer to the rte_eth_dev structure.
804  *
805  * @return
806  *   Nonzero if the callback process can be called immediately.
807  */
808 static int
809 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
810 {
811 	struct ibv_async_event event;
812 	int port_change = 0;
813 	int ret = 0;
814 
815 	/* Read all message and acknowledge them. */
816 	for (;;) {
817 		if (ibv_get_async_event(priv->ctx, &event))
818 			break;
819 
820 		if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
821 		    event.event_type == IBV_EVENT_PORT_ERR)
822 			port_change = 1;
823 		else
824 			DEBUG("event type %d on port %d not handled",
825 			      event.event_type, event.element.port_num);
826 		ibv_ack_async_event(&event);
827 	}
828 
829 	if (port_change ^ priv->pending_alarm) {
830 		struct rte_eth_link *link = &dev->data->dev_link;
831 
832 		priv->pending_alarm = 0;
833 		mlx5_link_update_unlocked(dev, 0);
834 		if (((link->link_speed == 0) && link->link_status) ||
835 		    ((link->link_speed != 0) && !link->link_status)) {
836 			/* Inconsistent status, check again later. */
837 			priv->pending_alarm = 1;
838 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
839 					  mlx5_dev_link_status_handler,
840 					  dev);
841 		} else
842 			ret = 1;
843 	}
844 	return ret;
845 }
846 
847 /**
848  * Handle delayed link status event.
849  *
850  * @param arg
851  *   Registered argument.
852  */
853 void
854 mlx5_dev_link_status_handler(void *arg)
855 {
856 	struct rte_eth_dev *dev = arg;
857 	struct priv *priv = dev->data->dev_private;
858 	int ret;
859 
860 	priv_lock(priv);
861 	assert(priv->pending_alarm == 1);
862 	ret = priv_dev_link_status_handler(priv, dev);
863 	priv_unlock(priv);
864 	if (ret)
865 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
866 }
867 
868 /**
869  * Handle interrupts from the NIC.
870  *
871  * @param[in] intr_handle
872  *   Interrupt handler.
873  * @param cb_arg
874  *   Callback argument.
875  */
876 void
877 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
878 {
879 	struct rte_eth_dev *dev = cb_arg;
880 	struct priv *priv = dev->data->dev_private;
881 	int ret;
882 
883 	(void)intr_handle;
884 	priv_lock(priv);
885 	ret = priv_dev_link_status_handler(priv, dev);
886 	priv_unlock(priv);
887 	if (ret)
888 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
889 }
890 
891 /**
892  * Uninstall interrupt handler.
893  *
894  * @param priv
895  *   Pointer to private structure.
896  * @param dev
897  *   Pointer to the rte_eth_dev structure.
898  */
899 void
900 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
901 {
902 	if (!dev->data->dev_conf.intr_conf.lsc)
903 		return;
904 	rte_intr_callback_unregister(&priv->intr_handle,
905 				     mlx5_dev_interrupt_handler,
906 				     dev);
907 	if (priv->pending_alarm)
908 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
909 	priv->pending_alarm = 0;
910 	priv->intr_handle.fd = 0;
911 	priv->intr_handle.type = 0;
912 }
913 
914 /**
915  * Install interrupt handler.
916  *
917  * @param priv
918  *   Pointer to private structure.
919  * @param dev
920  *   Pointer to the rte_eth_dev structure.
921  */
922 void
923 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
924 {
925 	int rc, flags;
926 
927 	if (!dev->data->dev_conf.intr_conf.lsc)
928 		return;
929 	assert(priv->ctx->async_fd > 0);
930 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
931 	rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
932 	if (rc < 0) {
933 		INFO("failed to change file descriptor async event queue");
934 		dev->data->dev_conf.intr_conf.lsc = 0;
935 	} else {
936 		priv->intr_handle.fd = priv->ctx->async_fd;
937 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
938 		rte_intr_callback_register(&priv->intr_handle,
939 					   mlx5_dev_interrupt_handler,
940 					   dev);
941 	}
942 }
943