xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 7ba5320baa3285bf8beca96cc025ff56b494060d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox.
4  */
5 
6 #define _GNU_SOURCE
7 
8 #include <stddef.h>
9 #include <assert.h>
10 #include <inttypes.h>
11 #include <unistd.h>
12 #include <stdint.h>
13 #include <stdio.h>
14 #include <string.h>
15 #include <stdlib.h>
16 #include <errno.h>
17 #include <dirent.h>
18 #include <net/if.h>
19 #include <sys/ioctl.h>
20 #include <sys/socket.h>
21 #include <netinet/in.h>
22 #include <linux/ethtool.h>
23 #include <linux/sockios.h>
24 #include <fcntl.h>
25 #include <stdalign.h>
26 #include <sys/un.h>
27 
28 #include <rte_atomic.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_bus_pci.h>
31 #include <rte_mbuf.h>
32 #include <rte_common.h>
33 #include <rte_interrupts.h>
34 #include <rte_alarm.h>
35 #include <rte_malloc.h>
36 
37 #include "mlx5.h"
38 #include "mlx5_glue.h"
39 #include "mlx5_rxtx.h"
40 #include "mlx5_utils.h"
41 
42 /* Add defines in case the running kernel is not the same as user headers. */
43 #ifndef ETHTOOL_GLINKSETTINGS
44 struct ethtool_link_settings {
45 	uint32_t cmd;
46 	uint32_t speed;
47 	uint8_t duplex;
48 	uint8_t port;
49 	uint8_t phy_address;
50 	uint8_t autoneg;
51 	uint8_t mdio_support;
52 	uint8_t eth_to_mdix;
53 	uint8_t eth_tp_mdix_ctrl;
54 	int8_t link_mode_masks_nwords;
55 	uint32_t reserved[8];
56 	uint32_t link_mode_masks[];
57 };
58 
59 #define ETHTOOL_GLINKSETTINGS 0x0000004c
60 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
61 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
62 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
63 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
64 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
65 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
66 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
67 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
68 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
69 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
70 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
71 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
72 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
73 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
74 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
75 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
76 #endif
77 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
78 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
79 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
80 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
81 #endif
82 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
83 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
84 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
85 #endif
86 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
87 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
88 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
89 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
90 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
91 #endif
92 
93 /**
94  * Get interface name from private structure.
95  *
96  * @param[in] dev
97  *   Pointer to Ethernet device.
98  * @param[out] ifname
99  *   Interface name output buffer.
100  *
101  * @return
102  *   0 on success, a negative errno value otherwise and rte_errno is set.
103  */
104 int
105 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
106 {
107 	struct priv *priv = dev->data->dev_private;
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 			rte_errno = errno;
120 			return -rte_errno;
121 		}
122 	}
123 	while ((dent = readdir(dir)) != NULL) {
124 		char *name = dent->d_name;
125 		FILE *file;
126 		unsigned int dev_port;
127 		int r;
128 
129 		if ((name[0] == '.') &&
130 		    ((name[1] == '\0') ||
131 		     ((name[1] == '.') && (name[2] == '\0'))))
132 			continue;
133 
134 		MKSTR(path, "%s/device/net/%s/%s",
135 		      priv->ibdev_path, name,
136 		      (dev_type ? "dev_id" : "dev_port"));
137 
138 		file = fopen(path, "rb");
139 		if (file == NULL) {
140 			if (errno != ENOENT)
141 				continue;
142 			/*
143 			 * Switch to dev_id when dev_port does not exist as
144 			 * is the case with Linux kernel versions < 3.15.
145 			 */
146 try_dev_id:
147 			match[0] = '\0';
148 			if (dev_type)
149 				break;
150 			dev_type = 1;
151 			dev_port_prev = ~0u;
152 			rewinddir(dir);
153 			continue;
154 		}
155 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
156 		fclose(file);
157 		if (r != 1)
158 			continue;
159 		/*
160 		 * Switch to dev_id when dev_port returns the same value for
161 		 * all ports. May happen when using a MOFED release older than
162 		 * 3.0 with a Linux kernel >= 3.15.
163 		 */
164 		if (dev_port == dev_port_prev)
165 			goto try_dev_id;
166 		dev_port_prev = dev_port;
167 		if (dev_port == (priv->port - 1u))
168 			snprintf(match, sizeof(match), "%s", name);
169 	}
170 	closedir(dir);
171 	if (match[0] == '\0') {
172 		rte_errno = ENOENT;
173 		return -rte_errno;
174 	}
175 	strncpy(*ifname, match, sizeof(*ifname));
176 	return 0;
177 }
178 
179 /**
180  * Perform ifreq ioctl() on associated Ethernet device.
181  *
182  * @param[in] dev
183  *   Pointer to Ethernet device.
184  * @param req
185  *   Request number to pass to ioctl().
186  * @param[out] ifr
187  *   Interface request structure output buffer.
188  *
189  * @return
190  *   0 on success, a negative errno value otherwise and rte_errno is set.
191  */
192 int
193 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
194 {
195 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
196 	int ret = 0;
197 
198 	if (sock == -1) {
199 		rte_errno = errno;
200 		return -rte_errno;
201 	}
202 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
203 	if (ret)
204 		goto error;
205 	ret = ioctl(sock, req, ifr);
206 	if (ret == -1) {
207 		rte_errno = errno;
208 		goto error;
209 	}
210 	close(sock);
211 	return 0;
212 error:
213 	close(sock);
214 	return -rte_errno;
215 }
216 
217 /**
218  * Get device MTU.
219  *
220  * @param dev
221  *   Pointer to Ethernet device.
222  * @param[out] mtu
223  *   MTU value output buffer.
224  *
225  * @return
226  *   0 on success, a negative errno value otherwise and rte_errno is set.
227  */
228 int
229 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
230 {
231 	struct ifreq request;
232 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
233 
234 	if (ret)
235 		return ret;
236 	*mtu = request.ifr_mtu;
237 	return 0;
238 }
239 
240 /**
241  * Set device MTU.
242  *
243  * @param dev
244  *   Pointer to Ethernet device.
245  * @param mtu
246  *   MTU value to set.
247  *
248  * @return
249  *   0 on success, a negative errno value otherwise and rte_errno is set.
250  */
251 static int
252 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
253 {
254 	struct ifreq request = { .ifr_mtu = mtu, };
255 
256 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
257 }
258 
259 /**
260  * Set device flags.
261  *
262  * @param dev
263  *   Pointer to Ethernet device.
264  * @param keep
265  *   Bitmask for flags that must remain untouched.
266  * @param flags
267  *   Bitmask for flags to modify.
268  *
269  * @return
270  *   0 on success, a negative errno value otherwise and rte_errno is set.
271  */
272 int
273 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
274 {
275 	struct ifreq request;
276 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
277 
278 	if (ret)
279 		return ret;
280 	request.ifr_flags &= keep;
281 	request.ifr_flags |= flags & ~keep;
282 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
283 }
284 
285 /**
286  * DPDK callback for Ethernet device configuration.
287  *
288  * @param dev
289  *   Pointer to Ethernet device structure.
290  *
291  * @return
292  *   0 on success, a negative errno value otherwise and rte_errno is set.
293  */
294 int
295 mlx5_dev_configure(struct rte_eth_dev *dev)
296 {
297 	struct priv *priv = dev->data->dev_private;
298 	unsigned int rxqs_n = dev->data->nb_rx_queues;
299 	unsigned int txqs_n = dev->data->nb_tx_queues;
300 	unsigned int i;
301 	unsigned int j;
302 	unsigned int reta_idx_n;
303 	const uint8_t use_app_rss_key =
304 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
305 	uint64_t supp_tx_offloads = mlx5_get_tx_port_offloads(dev);
306 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
307 	uint64_t supp_rx_offloads =
308 		(mlx5_get_rx_port_offloads() |
309 		 mlx5_get_rx_queue_offloads(dev));
310 	uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
311 	int ret = 0;
312 
313 	if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
314 		DRV_LOG(ERR,
315 			"port %u some Tx offloads are not supported requested"
316 			" 0x%" PRIx64 " supported 0x%" PRIx64,
317 			dev->data->port_id, tx_offloads, supp_tx_offloads);
318 		rte_errno = ENOTSUP;
319 		return -rte_errno;
320 	}
321 	if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
322 		DRV_LOG(ERR,
323 			"port %u some Rx offloads are not supported requested"
324 			" 0x%" PRIx64 " supported 0x%" PRIx64,
325 			dev->data->port_id, rx_offloads, supp_rx_offloads);
326 		rte_errno = ENOTSUP;
327 		return -rte_errno;
328 	}
329 	if (use_app_rss_key &&
330 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
331 	     rss_hash_default_key_len)) {
332 		/* MLX5 RSS only support 40bytes key. */
333 		rte_errno = EINVAL;
334 		return -rte_errno;
335 	}
336 	priv->rss_conf.rss_key =
337 		rte_realloc(priv->rss_conf.rss_key,
338 			    rss_hash_default_key_len, 0);
339 	if (!priv->rss_conf.rss_key) {
340 		DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
341 			dev->data->port_id, rxqs_n);
342 		rte_errno = ENOMEM;
343 		return -rte_errno;
344 	}
345 	memcpy(priv->rss_conf.rss_key,
346 	       use_app_rss_key ?
347 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
348 	       rss_hash_default_key,
349 	       rss_hash_default_key_len);
350 	priv->rss_conf.rss_key_len = rss_hash_default_key_len;
351 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
352 	priv->rxqs = (void *)dev->data->rx_queues;
353 	priv->txqs = (void *)dev->data->tx_queues;
354 	if (txqs_n != priv->txqs_n) {
355 		DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
356 			dev->data->port_id, priv->txqs_n, txqs_n);
357 		priv->txqs_n = txqs_n;
358 	}
359 	if (rxqs_n > priv->config.ind_table_max_size) {
360 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
361 			dev->data->port_id, rxqs_n);
362 		rte_errno = EINVAL;
363 		return -rte_errno;
364 	}
365 	if (rxqs_n == priv->rxqs_n)
366 		return 0;
367 	DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
368 		dev->data->port_id, priv->rxqs_n, rxqs_n);
369 	priv->rxqs_n = rxqs_n;
370 	/* If the requested number of RX queues is not a power of two, use the
371 	 * maximum indirection table size for better balancing.
372 	 * The result is always rounded to the next power of two. */
373 	reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
374 				     priv->config.ind_table_max_size :
375 				     rxqs_n));
376 	ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
377 	if (ret)
378 		return ret;
379 	/* When the number of RX queues is not a power of two, the remaining
380 	 * table entries are padded with reused WQs and hashes are not spread
381 	 * uniformly. */
382 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
383 		(*priv->reta_idx)[i] = j;
384 		if (++j == rxqs_n)
385 			j = 0;
386 	}
387 	return 0;
388 }
389 
390 /**
391  * DPDK callback to get information about the device.
392  *
393  * @param dev
394  *   Pointer to Ethernet device structure.
395  * @param[out] info
396  *   Info structure output buffer.
397  */
398 void
399 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
400 {
401 	struct priv *priv = dev->data->dev_private;
402 	struct mlx5_dev_config *config = &priv->config;
403 	unsigned int max;
404 	char ifname[IF_NAMESIZE];
405 
406 	info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
407 	/* FIXME: we should ask the device for these values. */
408 	info->min_rx_bufsize = 32;
409 	info->max_rx_pktlen = 65536;
410 	/*
411 	 * Since we need one CQ per QP, the limit is the minimum number
412 	 * between the two values.
413 	 */
414 	max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
415 		      priv->device_attr.orig_attr.max_qp);
416 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
417 	if (max >= 65535)
418 		max = 65535;
419 	info->max_rx_queues = max;
420 	info->max_tx_queues = max;
421 	info->max_mac_addrs = RTE_DIM(priv->mac);
422 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
423 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
424 				 info->rx_queue_offload_capa);
425 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
426 	if (mlx5_get_ifname(dev, &ifname) == 0)
427 		info->if_index = if_nametoindex(ifname);
428 	info->reta_size = priv->reta_idx_n ?
429 		priv->reta_idx_n : config->ind_table_max_size;
430 	info->hash_key_size = priv->rss_conf.rss_key_len;
431 	info->speed_capa = priv->link_speed_capa;
432 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
433 }
434 
435 /**
436  * Get supported packet types.
437  *
438  * @param dev
439  *   Pointer to Ethernet device structure.
440  *
441  * @return
442  *   A pointer to the supported Packet types array.
443  */
444 const uint32_t *
445 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
446 {
447 	static const uint32_t ptypes[] = {
448 		/* refers to rxq_cq_to_pkt_type() */
449 		RTE_PTYPE_L2_ETHER,
450 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
451 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
452 		RTE_PTYPE_L4_NONFRAG,
453 		RTE_PTYPE_L4_FRAG,
454 		RTE_PTYPE_L4_TCP,
455 		RTE_PTYPE_L4_UDP,
456 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
457 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
458 		RTE_PTYPE_INNER_L4_NONFRAG,
459 		RTE_PTYPE_INNER_L4_FRAG,
460 		RTE_PTYPE_INNER_L4_TCP,
461 		RTE_PTYPE_INNER_L4_UDP,
462 		RTE_PTYPE_UNKNOWN
463 	};
464 
465 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
466 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
467 		return ptypes;
468 	return NULL;
469 }
470 
471 /**
472  * DPDK callback to retrieve physical link information.
473  *
474  * @param dev
475  *   Pointer to Ethernet device structure.
476  *
477  * @return
478  *   0 on success, a negative errno value otherwise and rte_errno is set.
479  */
480 static int
481 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev)
482 {
483 	struct priv *priv = dev->data->dev_private;
484 	struct ethtool_cmd edata = {
485 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
486 	};
487 	struct ifreq ifr;
488 	struct rte_eth_link dev_link;
489 	int link_speed = 0;
490 	int ret;
491 
492 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
493 	if (ret) {
494 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
495 			dev->data->port_id, strerror(rte_errno));
496 		return ret;
497 	}
498 	memset(&dev_link, 0, sizeof(dev_link));
499 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
500 				(ifr.ifr_flags & IFF_RUNNING));
501 	ifr.ifr_data = (void *)&edata;
502 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
503 	if (ret) {
504 		DRV_LOG(WARNING,
505 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
506 			dev->data->port_id, strerror(rte_errno));
507 		return ret;
508 	}
509 	link_speed = ethtool_cmd_speed(&edata);
510 	if (link_speed == -1)
511 		dev_link.link_speed = 0;
512 	else
513 		dev_link.link_speed = link_speed;
514 	priv->link_speed_capa = 0;
515 	if (edata.supported & SUPPORTED_Autoneg)
516 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
517 	if (edata.supported & (SUPPORTED_1000baseT_Full |
518 			       SUPPORTED_1000baseKX_Full))
519 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
520 	if (edata.supported & SUPPORTED_10000baseKR_Full)
521 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
522 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
523 			       SUPPORTED_40000baseCR4_Full |
524 			       SUPPORTED_40000baseSR4_Full |
525 			       SUPPORTED_40000baseLR4_Full))
526 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
527 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
528 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
529 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
530 			ETH_LINK_SPEED_FIXED);
531 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
532 		/* Link status changed. */
533 		dev->data->dev_link = dev_link;
534 		return 0;
535 	}
536 	/* Link status is still the same. */
537 	rte_errno = EAGAIN;
538 	return -rte_errno;
539 }
540 
541 /**
542  * Retrieve physical link information (unlocked version using new ioctl).
543  *
544  * @param dev
545  *   Pointer to Ethernet device structure.
546  *
547  * @return
548  *   0 on success, a negative errno value otherwise and rte_errno is set.
549  */
550 static int
551 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev)
552 {
553 	struct priv *priv = dev->data->dev_private;
554 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
555 	struct ifreq ifr;
556 	struct rte_eth_link dev_link;
557 	uint64_t sc;
558 	int ret;
559 
560 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
561 	if (ret) {
562 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
563 			dev->data->port_id, strerror(rte_errno));
564 		return ret;
565 	}
566 	memset(&dev_link, 0, sizeof(dev_link));
567 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
568 				(ifr.ifr_flags & IFF_RUNNING));
569 	ifr.ifr_data = (void *)&gcmd;
570 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
571 	if (ret) {
572 		DRV_LOG(DEBUG,
573 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
574 			" failed: %s",
575 			dev->data->port_id, strerror(rte_errno));
576 		return ret;
577 	}
578 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
579 
580 	alignas(struct ethtool_link_settings)
581 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
582 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
583 	struct ethtool_link_settings *ecmd = (void *)data;
584 
585 	*ecmd = gcmd;
586 	ifr.ifr_data = (void *)ecmd;
587 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
588 	if (ret) {
589 		DRV_LOG(DEBUG,
590 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
591 			" failed: %s",
592 			dev->data->port_id, strerror(rte_errno));
593 		return ret;
594 	}
595 	dev_link.link_speed = ecmd->speed;
596 	sc = ecmd->link_mode_masks[0] |
597 		((uint64_t)ecmd->link_mode_masks[1] << 32);
598 	priv->link_speed_capa = 0;
599 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
600 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
601 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
602 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
603 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
604 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
605 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
606 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
607 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
608 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
609 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
610 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
611 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
612 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
613 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
614 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
615 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
616 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
617 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
618 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
619 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
620 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
621 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
622 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
623 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
624 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
625 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
626 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
627 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
628 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
629 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
630 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
631 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
632 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
633 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
634 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
635 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
636 				  ETH_LINK_SPEED_FIXED);
637 	if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
638 		/* Link status changed. */
639 		dev->data->dev_link = dev_link;
640 		return 0;
641 	}
642 	/* Link status is still the same. */
643 	rte_errno = EAGAIN;
644 	return -rte_errno;
645 }
646 
647 /**
648  * DPDK callback to retrieve physical link information.
649  *
650  * @param dev
651  *   Pointer to Ethernet device structure.
652  * @param wait_to_complete
653  *   Wait for request completion (ignored).
654  *
655  * @return
656  *   0 on success, a negative errno value otherwise and rte_errno is set.
657  */
658 int
659 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete __rte_unused)
660 {
661 	int ret;
662 
663 	ret = mlx5_link_update_unlocked_gset(dev);
664 	if (ret)
665 		ret = mlx5_link_update_unlocked_gs(dev);
666 	return 0;
667 }
668 
669 /**
670  * DPDK callback to change the MTU.
671  *
672  * @param dev
673  *   Pointer to Ethernet device structure.
674  * @param in_mtu
675  *   New MTU.
676  *
677  * @return
678  *   0 on success, a negative errno value otherwise and rte_errno is set.
679  */
680 int
681 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
682 {
683 	struct priv *priv = dev->data->dev_private;
684 	uint16_t kern_mtu = 0;
685 	int ret;
686 
687 	ret = mlx5_get_mtu(dev, &kern_mtu);
688 	if (ret)
689 		return ret;
690 	/* Set kernel interface MTU first. */
691 	ret = mlx5_set_mtu(dev, mtu);
692 	if (ret)
693 		return ret;
694 	ret = mlx5_get_mtu(dev, &kern_mtu);
695 	if (ret)
696 		return ret;
697 	if (kern_mtu == mtu) {
698 		priv->mtu = mtu;
699 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
700 			dev->data->port_id, mtu);
701 		return 0;
702 	}
703 	rte_errno = EAGAIN;
704 	return -rte_errno;
705 }
706 
707 /**
708  * DPDK callback to get flow control status.
709  *
710  * @param dev
711  *   Pointer to Ethernet device structure.
712  * @param[out] fc_conf
713  *   Flow control output buffer.
714  *
715  * @return
716  *   0 on success, a negative errno value otherwise and rte_errno is set.
717  */
718 int
719 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
720 {
721 	struct ifreq ifr;
722 	struct ethtool_pauseparam ethpause = {
723 		.cmd = ETHTOOL_GPAUSEPARAM
724 	};
725 	int ret;
726 
727 	ifr.ifr_data = (void *)&ethpause;
728 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
729 	if (ret) {
730 		DRV_LOG(WARNING,
731 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
732 			" %s",
733 			dev->data->port_id, strerror(rte_errno));
734 		return ret;
735 	}
736 	fc_conf->autoneg = ethpause.autoneg;
737 	if (ethpause.rx_pause && ethpause.tx_pause)
738 		fc_conf->mode = RTE_FC_FULL;
739 	else if (ethpause.rx_pause)
740 		fc_conf->mode = RTE_FC_RX_PAUSE;
741 	else if (ethpause.tx_pause)
742 		fc_conf->mode = RTE_FC_TX_PAUSE;
743 	else
744 		fc_conf->mode = RTE_FC_NONE;
745 	return 0;
746 }
747 
748 /**
749  * DPDK callback to modify flow control parameters.
750  *
751  * @param dev
752  *   Pointer to Ethernet device structure.
753  * @param[in] fc_conf
754  *   Flow control parameters.
755  *
756  * @return
757  *   0 on success, a negative errno value otherwise and rte_errno is set.
758  */
759 int
760 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
761 {
762 	struct ifreq ifr;
763 	struct ethtool_pauseparam ethpause = {
764 		.cmd = ETHTOOL_SPAUSEPARAM
765 	};
766 	int ret;
767 
768 	ifr.ifr_data = (void *)&ethpause;
769 	ethpause.autoneg = fc_conf->autoneg;
770 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
771 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
772 		ethpause.rx_pause = 1;
773 	else
774 		ethpause.rx_pause = 0;
775 
776 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
777 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
778 		ethpause.tx_pause = 1;
779 	else
780 		ethpause.tx_pause = 0;
781 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
782 	if (ret) {
783 		DRV_LOG(WARNING,
784 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
785 			" failed: %s",
786 			dev->data->port_id, strerror(rte_errno));
787 		return ret;
788 	}
789 	return 0;
790 }
791 
792 /**
793  * Get PCI information from struct ibv_device.
794  *
795  * @param device
796  *   Pointer to Ethernet device structure.
797  * @param[out] pci_addr
798  *   PCI bus address output buffer.
799  *
800  * @return
801  *   0 on success, a negative errno value otherwise and rte_errno is set.
802  */
803 int
804 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
805 			    struct rte_pci_addr *pci_addr)
806 {
807 	FILE *file;
808 	char line[32];
809 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
810 
811 	file = fopen(path, "rb");
812 	if (file == NULL) {
813 		rte_errno = errno;
814 		return -rte_errno;
815 	}
816 	while (fgets(line, sizeof(line), file) == line) {
817 		size_t len = strlen(line);
818 		int ret;
819 
820 		/* Truncate long lines. */
821 		if (len == (sizeof(line) - 1))
822 			while (line[(len - 1)] != '\n') {
823 				ret = fgetc(file);
824 				if (ret == EOF)
825 					break;
826 				line[(len - 1)] = ret;
827 			}
828 		/* Extract information. */
829 		if (sscanf(line,
830 			   "PCI_SLOT_NAME="
831 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
832 			   &pci_addr->domain,
833 			   &pci_addr->bus,
834 			   &pci_addr->devid,
835 			   &pci_addr->function) == 4) {
836 			ret = 0;
837 			break;
838 		}
839 	}
840 	fclose(file);
841 	return 0;
842 }
843 
844 /**
845  * Update the link status.
846  *
847  * @param dev
848  *   Pointer to Ethernet device.
849  *
850  * @return
851  *   Zero if the callback process can be called immediately, negative errno
852  *   value otherwise and rte_errno is set.
853  */
854 static int
855 mlx5_link_status_update(struct rte_eth_dev *dev)
856 {
857 	struct priv *priv = dev->data->dev_private;
858 	struct rte_eth_link *link = &dev->data->dev_link;
859 	int ret;
860 
861 	ret = mlx5_link_update(dev, 0);
862 	if (ret)
863 		return ret;
864 	if (((link->link_speed == 0) && link->link_status) ||
865 		((link->link_speed != 0) && !link->link_status)) {
866 		/*
867 		 * Inconsistent status. Event likely occurred before the
868 		 * kernel netdevice exposes the new status.
869 		 */
870 		if (!priv->pending_alarm) {
871 			priv->pending_alarm = 1;
872 			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
873 					  mlx5_dev_link_status_handler,
874 					  priv->dev);
875 		}
876 		return 1;
877 	} else if (unlikely(priv->pending_alarm)) {
878 		/* Link interrupt occurred while alarm is already scheduled. */
879 		priv->pending_alarm = 0;
880 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
881 	}
882 	return 0;
883 }
884 
885 /**
886  * Device status handler.
887  *
888  * @param dev
889  *   Pointer to Ethernet device.
890  * @param events
891  *   Pointer to event flags holder.
892  *
893  * @return
894  *   Events bitmap of callback process which can be called immediately.
895  */
896 static uint32_t
897 mlx5_dev_status_handler(struct rte_eth_dev *dev)
898 {
899 	struct priv *priv = dev->data->dev_private;
900 	struct ibv_async_event event;
901 	uint32_t ret = 0;
902 
903 	/* Read all message and acknowledge them. */
904 	for (;;) {
905 		if (mlx5_glue->get_async_event(priv->ctx, &event))
906 			break;
907 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
908 			event.event_type == IBV_EVENT_PORT_ERR) &&
909 			(dev->data->dev_conf.intr_conf.lsc == 1))
910 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
911 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
912 			dev->data->dev_conf.intr_conf.rmv == 1)
913 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
914 		else
915 			DRV_LOG(DEBUG,
916 				"port %u event type %d on not handled",
917 				dev->data->port_id, event.event_type);
918 		mlx5_glue->ack_async_event(&event);
919 	}
920 	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
921 		if (mlx5_link_status_update(dev))
922 			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
923 	return ret;
924 }
925 
926 /**
927  * Handle delayed link status event.
928  *
929  * @param arg
930  *   Registered argument.
931  */
932 void
933 mlx5_dev_link_status_handler(void *arg)
934 {
935 	struct rte_eth_dev *dev = arg;
936 	struct priv *priv = dev->data->dev_private;
937 	int ret;
938 
939 	priv->pending_alarm = 0;
940 	ret = mlx5_link_status_update(dev);
941 	if (!ret)
942 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
943 }
944 
945 /**
946  * Handle interrupts from the NIC.
947  *
948  * @param[in] intr_handle
949  *   Interrupt handler.
950  * @param cb_arg
951  *   Callback argument.
952  */
953 void
954 mlx5_dev_interrupt_handler(void *cb_arg)
955 {
956 	struct rte_eth_dev *dev = cb_arg;
957 	uint32_t events;
958 
959 	events = mlx5_dev_status_handler(dev);
960 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
961 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
962 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
963 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
964 }
965 
966 /**
967  * Handle interrupts from the socket.
968  *
969  * @param cb_arg
970  *   Callback argument.
971  */
972 static void
973 mlx5_dev_handler_socket(void *cb_arg)
974 {
975 	struct rte_eth_dev *dev = cb_arg;
976 
977 	mlx5_socket_handle(dev);
978 }
979 
980 /**
981  * Uninstall interrupt handler.
982  *
983  * @param dev
984  *   Pointer to Ethernet device.
985  */
986 void
987 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
988 {
989 	struct priv *priv = dev->data->dev_private;
990 
991 	if (dev->data->dev_conf.intr_conf.lsc ||
992 	    dev->data->dev_conf.intr_conf.rmv)
993 		rte_intr_callback_unregister(&priv->intr_handle,
994 					     mlx5_dev_interrupt_handler, dev);
995 	if (priv->primary_socket)
996 		rte_intr_callback_unregister(&priv->intr_handle_socket,
997 					     mlx5_dev_handler_socket, dev);
998 	if (priv->pending_alarm) {
999 		priv->pending_alarm = 0;
1000 		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1001 	}
1002 	priv->intr_handle.fd = 0;
1003 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1004 	priv->intr_handle_socket.fd = 0;
1005 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1006 }
1007 
1008 /**
1009  * Install interrupt handler.
1010  *
1011  * @param dev
1012  *   Pointer to Ethernet device.
1013  */
1014 void
1015 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1016 {
1017 	struct priv *priv = dev->data->dev_private;
1018 	int ret;
1019 	int flags;
1020 
1021 	assert(priv->ctx->async_fd > 0);
1022 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
1023 	ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1024 	if (ret) {
1025 		DRV_LOG(INFO,
1026 			"port %u failed to change file descriptor async event"
1027 			" queue",
1028 			dev->data->port_id);
1029 		dev->data->dev_conf.intr_conf.lsc = 0;
1030 		dev->data->dev_conf.intr_conf.rmv = 0;
1031 	}
1032 	if (dev->data->dev_conf.intr_conf.lsc ||
1033 	    dev->data->dev_conf.intr_conf.rmv) {
1034 		priv->intr_handle.fd = priv->ctx->async_fd;
1035 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1036 		rte_intr_callback_register(&priv->intr_handle,
1037 					   mlx5_dev_interrupt_handler, dev);
1038 	}
1039 	ret = mlx5_socket_init(dev);
1040 	if (ret)
1041 		DRV_LOG(ERR, "port %u cannot initialise socket: %s",
1042 			dev->data->port_id, strerror(rte_errno));
1043 	else if (priv->primary_socket) {
1044 		priv->intr_handle_socket.fd = priv->primary_socket;
1045 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1046 		rte_intr_callback_register(&priv->intr_handle_socket,
1047 					   mlx5_dev_handler_socket, dev);
1048 	}
1049 }
1050 
1051 /**
1052  * DPDK callback to bring the link DOWN.
1053  *
1054  * @param dev
1055  *   Pointer to Ethernet device structure.
1056  *
1057  * @return
1058  *   0 on success, a negative errno value otherwise and rte_errno is set.
1059  */
1060 int
1061 mlx5_set_link_down(struct rte_eth_dev *dev)
1062 {
1063 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1064 }
1065 
1066 /**
1067  * DPDK callback to bring the link UP.
1068  *
1069  * @param dev
1070  *   Pointer to Ethernet device structure.
1071  *
1072  * @return
1073  *   0 on success, a negative errno value otherwise and rte_errno is set.
1074  */
1075 int
1076 mlx5_set_link_up(struct rte_eth_dev *dev)
1077 {
1078 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1079 }
1080 
1081 /**
1082  * Configure the TX function to use.
1083  *
1084  * @param dev
1085  *   Pointer to private data structure.
1086  *
1087  * @return
1088  *   Pointer to selected Tx burst function.
1089  */
1090 eth_tx_burst_t
1091 mlx5_select_tx_function(struct rte_eth_dev *dev)
1092 {
1093 	struct priv *priv = dev->data->dev_private;
1094 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1095 	struct mlx5_dev_config *config = &priv->config;
1096 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1097 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1098 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1099 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1100 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1101 
1102 	assert(priv != NULL);
1103 	/* Select appropriate TX function. */
1104 	if (vlan_insert || tso)
1105 		return tx_pkt_burst;
1106 	if (config->mps == MLX5_MPW_ENHANCED) {
1107 		if (mlx5_check_vec_tx_support(dev) > 0) {
1108 			if (mlx5_check_raw_vec_tx_support(dev) > 0)
1109 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1110 			else
1111 				tx_pkt_burst = mlx5_tx_burst_vec;
1112 			DRV_LOG(DEBUG,
1113 				"port %u selected enhanced MPW Tx vectorized"
1114 				" function",
1115 				dev->data->port_id);
1116 		} else {
1117 			tx_pkt_burst = mlx5_tx_burst_empw;
1118 			DRV_LOG(DEBUG,
1119 				"port %u selected enhanced MPW Tx function",
1120 				dev->data->port_id);
1121 		}
1122 	} else if (config->mps && (config->txq_inline > 0)) {
1123 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1124 		DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1125 			dev->data->port_id);
1126 	} else if (config->mps) {
1127 		tx_pkt_burst = mlx5_tx_burst_mpw;
1128 		DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1129 			dev->data->port_id);
1130 	}
1131 	return tx_pkt_burst;
1132 }
1133 
1134 /**
1135  * Configure the RX function to use.
1136  *
1137  * @param dev
1138  *   Pointer to private data structure.
1139  *
1140  * @return
1141  *   Pointer to selected Rx burst function.
1142  */
1143 eth_rx_burst_t
1144 mlx5_select_rx_function(struct rte_eth_dev *dev)
1145 {
1146 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1147 
1148 	assert(dev != NULL);
1149 	if (mlx5_check_vec_rx_support(dev) > 0) {
1150 		rx_pkt_burst = mlx5_rx_burst_vec;
1151 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1152 			dev->data->port_id);
1153 	}
1154 	return rx_pkt_burst;
1155 }
1156 
1157 /**
1158  * Check if mlx5 device was removed.
1159  *
1160  * @param dev
1161  *   Pointer to Ethernet device structure.
1162  *
1163  * @return
1164  *   1 when device is removed, otherwise 0.
1165  */
1166 int
1167 mlx5_is_removed(struct rte_eth_dev *dev)
1168 {
1169 	struct ibv_device_attr device_attr;
1170 	struct priv *priv = dev->data->dev_private;
1171 
1172 	if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1173 		return 1;
1174 	return 0;
1175 }
1176