xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision cfee94752b8f8f09b984335ae77b29af15b967fd)
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 #include <time.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_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  * @param[out] link
477  *   Storage for current link status.
478  *
479  * @return
480  *   0 on success, a negative errno value otherwise and rte_errno is set.
481  */
482 static int
483 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
484 			       struct rte_eth_link *link)
485 {
486 	struct priv *priv = dev->data->dev_private;
487 	struct ethtool_cmd edata = {
488 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
489 	};
490 	struct ifreq ifr;
491 	struct rte_eth_link dev_link;
492 	int link_speed = 0;
493 	int ret;
494 
495 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
496 	if (ret) {
497 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
498 			dev->data->port_id, strerror(rte_errno));
499 		return ret;
500 	}
501 	memset(&dev_link, 0, sizeof(dev_link));
502 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
503 				(ifr.ifr_flags & IFF_RUNNING));
504 	ifr.ifr_data = (void *)&edata;
505 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
506 	if (ret) {
507 		DRV_LOG(WARNING,
508 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
509 			dev->data->port_id, strerror(rte_errno));
510 		return ret;
511 	}
512 	link_speed = ethtool_cmd_speed(&edata);
513 	if (link_speed == -1)
514 		dev_link.link_speed = 0;
515 	else
516 		dev_link.link_speed = link_speed;
517 	priv->link_speed_capa = 0;
518 	if (edata.supported & SUPPORTED_Autoneg)
519 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
520 	if (edata.supported & (SUPPORTED_1000baseT_Full |
521 			       SUPPORTED_1000baseKX_Full))
522 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
523 	if (edata.supported & SUPPORTED_10000baseKR_Full)
524 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
525 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
526 			       SUPPORTED_40000baseCR4_Full |
527 			       SUPPORTED_40000baseSR4_Full |
528 			       SUPPORTED_40000baseLR4_Full))
529 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
530 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
531 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
532 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
533 			ETH_LINK_SPEED_FIXED);
534 	if ((dev_link.link_speed && !dev_link.link_status) ||
535 	    (!dev_link.link_speed && dev_link.link_status)) {
536 		rte_errno = EAGAIN;
537 		return -rte_errno;
538 	}
539 	*link = dev_link;
540 	return 0;
541 }
542 
543 /**
544  * Retrieve physical link information (unlocked version using new ioctl).
545  *
546  * @param dev
547  *   Pointer to Ethernet device structure.
548  * @param[out] link
549  *   Storage for current link status.
550  *
551  * @return
552  *   0 on success, a negative errno value otherwise and rte_errno is set.
553  */
554 static int
555 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
556 			     struct rte_eth_link *link)
557 
558 {
559 	struct priv *priv = dev->data->dev_private;
560 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
561 	struct ifreq ifr;
562 	struct rte_eth_link dev_link;
563 	uint64_t sc;
564 	int ret;
565 
566 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
567 	if (ret) {
568 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
569 			dev->data->port_id, strerror(rte_errno));
570 		return ret;
571 	}
572 	memset(&dev_link, 0, sizeof(dev_link));
573 	dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
574 				(ifr.ifr_flags & IFF_RUNNING));
575 	ifr.ifr_data = (void *)&gcmd;
576 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
577 	if (ret) {
578 		DRV_LOG(DEBUG,
579 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
580 			" failed: %s",
581 			dev->data->port_id, strerror(rte_errno));
582 		return ret;
583 	}
584 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
585 
586 	alignas(struct ethtool_link_settings)
587 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
588 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
589 	struct ethtool_link_settings *ecmd = (void *)data;
590 
591 	*ecmd = gcmd;
592 	ifr.ifr_data = (void *)ecmd;
593 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
594 	if (ret) {
595 		DRV_LOG(DEBUG,
596 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
597 			" failed: %s",
598 			dev->data->port_id, strerror(rte_errno));
599 		return ret;
600 	}
601 	dev_link.link_speed = ecmd->speed;
602 	sc = ecmd->link_mode_masks[0] |
603 		((uint64_t)ecmd->link_mode_masks[1] << 32);
604 	priv->link_speed_capa = 0;
605 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
606 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
607 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
608 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
609 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
610 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
611 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
612 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
613 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
614 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
615 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
616 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
617 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
618 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
619 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
620 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
621 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
622 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
623 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
624 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
625 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
626 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
627 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
628 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
629 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
630 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
631 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
632 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
633 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
634 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
635 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
636 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
637 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
638 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
639 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
640 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
641 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
642 				  ETH_LINK_SPEED_FIXED);
643 	if ((dev_link.link_speed && !dev_link.link_status) ||
644 	    (!dev_link.link_speed && dev_link.link_status)) {
645 		rte_errno = EAGAIN;
646 		return -rte_errno;
647 	}
648 	*link = dev_link;
649 	return 0;
650 }
651 
652 /**
653  * DPDK callback to retrieve physical link information.
654  *
655  * @param dev
656  *   Pointer to Ethernet device structure.
657  * @param wait_to_complete
658  *   Wait for request completion.
659  *
660  * @return
661  *   0 if link status was not updated, positive if it was, a negative errno
662  *   value otherwise and rte_errno is set.
663  */
664 int
665 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
666 {
667 	int ret;
668 	struct rte_eth_link dev_link;
669 	time_t start_time = time(NULL);
670 
671 	do {
672 		ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
673 		if (ret)
674 			ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
675 		if (ret == 0)
676 			break;
677 		/* Handle wait to complete situation. */
678 		if (wait_to_complete && ret == -EAGAIN) {
679 			if (abs((int)difftime(time(NULL), start_time)) <
680 			    MLX5_LINK_STATUS_TIMEOUT) {
681 				usleep(0);
682 				continue;
683 			} else {
684 				rte_errno = EBUSY;
685 				return -rte_errno;
686 			}
687 		} else if (ret < 0) {
688 			return ret;
689 		}
690 	} while (wait_to_complete);
691 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
692 		       sizeof(struct rte_eth_link));
693 	dev->data->dev_link = dev_link;
694 	return ret;
695 }
696 
697 /**
698  * DPDK callback to change the MTU.
699  *
700  * @param dev
701  *   Pointer to Ethernet device structure.
702  * @param in_mtu
703  *   New MTU.
704  *
705  * @return
706  *   0 on success, a negative errno value otherwise and rte_errno is set.
707  */
708 int
709 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
710 {
711 	struct priv *priv = dev->data->dev_private;
712 	uint16_t kern_mtu = 0;
713 	int ret;
714 
715 	ret = mlx5_get_mtu(dev, &kern_mtu);
716 	if (ret)
717 		return ret;
718 	/* Set kernel interface MTU first. */
719 	ret = mlx5_set_mtu(dev, mtu);
720 	if (ret)
721 		return ret;
722 	ret = mlx5_get_mtu(dev, &kern_mtu);
723 	if (ret)
724 		return ret;
725 	if (kern_mtu == mtu) {
726 		priv->mtu = mtu;
727 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
728 			dev->data->port_id, mtu);
729 		return 0;
730 	}
731 	rte_errno = EAGAIN;
732 	return -rte_errno;
733 }
734 
735 /**
736  * DPDK callback to get flow control status.
737  *
738  * @param dev
739  *   Pointer to Ethernet device structure.
740  * @param[out] fc_conf
741  *   Flow control output buffer.
742  *
743  * @return
744  *   0 on success, a negative errno value otherwise and rte_errno is set.
745  */
746 int
747 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
748 {
749 	struct ifreq ifr;
750 	struct ethtool_pauseparam ethpause = {
751 		.cmd = ETHTOOL_GPAUSEPARAM
752 	};
753 	int ret;
754 
755 	ifr.ifr_data = (void *)&ethpause;
756 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
757 	if (ret) {
758 		DRV_LOG(WARNING,
759 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
760 			" %s",
761 			dev->data->port_id, strerror(rte_errno));
762 		return ret;
763 	}
764 	fc_conf->autoneg = ethpause.autoneg;
765 	if (ethpause.rx_pause && ethpause.tx_pause)
766 		fc_conf->mode = RTE_FC_FULL;
767 	else if (ethpause.rx_pause)
768 		fc_conf->mode = RTE_FC_RX_PAUSE;
769 	else if (ethpause.tx_pause)
770 		fc_conf->mode = RTE_FC_TX_PAUSE;
771 	else
772 		fc_conf->mode = RTE_FC_NONE;
773 	return 0;
774 }
775 
776 /**
777  * DPDK callback to modify flow control parameters.
778  *
779  * @param dev
780  *   Pointer to Ethernet device structure.
781  * @param[in] fc_conf
782  *   Flow control parameters.
783  *
784  * @return
785  *   0 on success, a negative errno value otherwise and rte_errno is set.
786  */
787 int
788 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
789 {
790 	struct ifreq ifr;
791 	struct ethtool_pauseparam ethpause = {
792 		.cmd = ETHTOOL_SPAUSEPARAM
793 	};
794 	int ret;
795 
796 	ifr.ifr_data = (void *)&ethpause;
797 	ethpause.autoneg = fc_conf->autoneg;
798 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
799 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
800 		ethpause.rx_pause = 1;
801 	else
802 		ethpause.rx_pause = 0;
803 
804 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
805 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
806 		ethpause.tx_pause = 1;
807 	else
808 		ethpause.tx_pause = 0;
809 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
810 	if (ret) {
811 		DRV_LOG(WARNING,
812 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
813 			" failed: %s",
814 			dev->data->port_id, strerror(rte_errno));
815 		return ret;
816 	}
817 	return 0;
818 }
819 
820 /**
821  * Get PCI information from struct ibv_device.
822  *
823  * @param device
824  *   Pointer to Ethernet device structure.
825  * @param[out] pci_addr
826  *   PCI bus address output buffer.
827  *
828  * @return
829  *   0 on success, a negative errno value otherwise and rte_errno is set.
830  */
831 int
832 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
833 			    struct rte_pci_addr *pci_addr)
834 {
835 	FILE *file;
836 	char line[32];
837 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
838 
839 	file = fopen(path, "rb");
840 	if (file == NULL) {
841 		rte_errno = errno;
842 		return -rte_errno;
843 	}
844 	while (fgets(line, sizeof(line), file) == line) {
845 		size_t len = strlen(line);
846 		int ret;
847 
848 		/* Truncate long lines. */
849 		if (len == (sizeof(line) - 1))
850 			while (line[(len - 1)] != '\n') {
851 				ret = fgetc(file);
852 				if (ret == EOF)
853 					break;
854 				line[(len - 1)] = ret;
855 			}
856 		/* Extract information. */
857 		if (sscanf(line,
858 			   "PCI_SLOT_NAME="
859 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
860 			   &pci_addr->domain,
861 			   &pci_addr->bus,
862 			   &pci_addr->devid,
863 			   &pci_addr->function) == 4) {
864 			ret = 0;
865 			break;
866 		}
867 	}
868 	fclose(file);
869 	return 0;
870 }
871 
872 /**
873  * Device status handler.
874  *
875  * @param dev
876  *   Pointer to Ethernet device.
877  * @param events
878  *   Pointer to event flags holder.
879  *
880  * @return
881  *   Events bitmap of callback process which can be called immediately.
882  */
883 static uint32_t
884 mlx5_dev_status_handler(struct rte_eth_dev *dev)
885 {
886 	struct priv *priv = dev->data->dev_private;
887 	struct ibv_async_event event;
888 	uint32_t ret = 0;
889 
890 	if (mlx5_link_update(dev, 0) == -EAGAIN) {
891 		usleep(0);
892 		return 0;
893 	}
894 	/* Read all message and acknowledge them. */
895 	for (;;) {
896 		if (mlx5_glue->get_async_event(priv->ctx, &event))
897 			break;
898 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
899 			event.event_type == IBV_EVENT_PORT_ERR) &&
900 			(dev->data->dev_conf.intr_conf.lsc == 1))
901 			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
902 		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
903 			dev->data->dev_conf.intr_conf.rmv == 1)
904 			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
905 		else
906 			DRV_LOG(DEBUG,
907 				"port %u event type %d on not handled",
908 				dev->data->port_id, event.event_type);
909 		mlx5_glue->ack_async_event(&event);
910 	}
911 	return ret;
912 }
913 
914 /**
915  * Handle interrupts from the NIC.
916  *
917  * @param[in] intr_handle
918  *   Interrupt handler.
919  * @param cb_arg
920  *   Callback argument.
921  */
922 void
923 mlx5_dev_interrupt_handler(void *cb_arg)
924 {
925 	struct rte_eth_dev *dev = cb_arg;
926 	uint32_t events;
927 
928 	events = mlx5_dev_status_handler(dev);
929 	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
930 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
931 	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
932 		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
933 }
934 
935 /**
936  * Handle interrupts from the socket.
937  *
938  * @param cb_arg
939  *   Callback argument.
940  */
941 static void
942 mlx5_dev_handler_socket(void *cb_arg)
943 {
944 	struct rte_eth_dev *dev = cb_arg;
945 
946 	mlx5_socket_handle(dev);
947 }
948 
949 /**
950  * Uninstall interrupt handler.
951  *
952  * @param dev
953  *   Pointer to Ethernet device.
954  */
955 void
956 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
957 {
958 	struct priv *priv = dev->data->dev_private;
959 
960 	if (dev->data->dev_conf.intr_conf.lsc ||
961 	    dev->data->dev_conf.intr_conf.rmv)
962 		rte_intr_callback_unregister(&priv->intr_handle,
963 					     mlx5_dev_interrupt_handler, dev);
964 	if (priv->primary_socket)
965 		rte_intr_callback_unregister(&priv->intr_handle_socket,
966 					     mlx5_dev_handler_socket, dev);
967 	priv->intr_handle.fd = 0;
968 	priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
969 	priv->intr_handle_socket.fd = 0;
970 	priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
971 }
972 
973 /**
974  * Install interrupt handler.
975  *
976  * @param dev
977  *   Pointer to Ethernet device.
978  */
979 void
980 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
981 {
982 	struct priv *priv = dev->data->dev_private;
983 	int ret;
984 	int flags;
985 
986 	assert(priv->ctx->async_fd > 0);
987 	flags = fcntl(priv->ctx->async_fd, F_GETFL);
988 	ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
989 	if (ret) {
990 		DRV_LOG(INFO,
991 			"port %u failed to change file descriptor async event"
992 			" queue",
993 			dev->data->port_id);
994 		dev->data->dev_conf.intr_conf.lsc = 0;
995 		dev->data->dev_conf.intr_conf.rmv = 0;
996 	}
997 	if (dev->data->dev_conf.intr_conf.lsc ||
998 	    dev->data->dev_conf.intr_conf.rmv) {
999 		priv->intr_handle.fd = priv->ctx->async_fd;
1000 		priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1001 		rte_intr_callback_register(&priv->intr_handle,
1002 					   mlx5_dev_interrupt_handler, dev);
1003 	}
1004 	ret = mlx5_socket_init(dev);
1005 	if (ret)
1006 		DRV_LOG(ERR, "port %u cannot initialise socket: %s",
1007 			dev->data->port_id, strerror(rte_errno));
1008 	else if (priv->primary_socket) {
1009 		priv->intr_handle_socket.fd = priv->primary_socket;
1010 		priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1011 		rte_intr_callback_register(&priv->intr_handle_socket,
1012 					   mlx5_dev_handler_socket, dev);
1013 	}
1014 }
1015 
1016 /**
1017  * DPDK callback to bring the link DOWN.
1018  *
1019  * @param dev
1020  *   Pointer to Ethernet device structure.
1021  *
1022  * @return
1023  *   0 on success, a negative errno value otherwise and rte_errno is set.
1024  */
1025 int
1026 mlx5_set_link_down(struct rte_eth_dev *dev)
1027 {
1028 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1029 }
1030 
1031 /**
1032  * DPDK callback to bring the link UP.
1033  *
1034  * @param dev
1035  *   Pointer to Ethernet device structure.
1036  *
1037  * @return
1038  *   0 on success, a negative errno value otherwise and rte_errno is set.
1039  */
1040 int
1041 mlx5_set_link_up(struct rte_eth_dev *dev)
1042 {
1043 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1044 }
1045 
1046 /**
1047  * Configure the TX function to use.
1048  *
1049  * @param dev
1050  *   Pointer to private data structure.
1051  *
1052  * @return
1053  *   Pointer to selected Tx burst function.
1054  */
1055 eth_tx_burst_t
1056 mlx5_select_tx_function(struct rte_eth_dev *dev)
1057 {
1058 	struct priv *priv = dev->data->dev_private;
1059 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1060 	struct mlx5_dev_config *config = &priv->config;
1061 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1062 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1063 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1064 				    DEV_TX_OFFLOAD_GRE_TNL_TSO));
1065 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1066 
1067 	assert(priv != NULL);
1068 	/* Select appropriate TX function. */
1069 	if (vlan_insert || tso)
1070 		return tx_pkt_burst;
1071 	if (config->mps == MLX5_MPW_ENHANCED) {
1072 		if (mlx5_check_vec_tx_support(dev) > 0) {
1073 			if (mlx5_check_raw_vec_tx_support(dev) > 0)
1074 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1075 			else
1076 				tx_pkt_burst = mlx5_tx_burst_vec;
1077 			DRV_LOG(DEBUG,
1078 				"port %u selected enhanced MPW Tx vectorized"
1079 				" function",
1080 				dev->data->port_id);
1081 		} else {
1082 			tx_pkt_burst = mlx5_tx_burst_empw;
1083 			DRV_LOG(DEBUG,
1084 				"port %u selected enhanced MPW Tx function",
1085 				dev->data->port_id);
1086 		}
1087 	} else if (config->mps && (config->txq_inline > 0)) {
1088 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1089 		DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1090 			dev->data->port_id);
1091 	} else if (config->mps) {
1092 		tx_pkt_burst = mlx5_tx_burst_mpw;
1093 		DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1094 			dev->data->port_id);
1095 	}
1096 	return tx_pkt_burst;
1097 }
1098 
1099 /**
1100  * Configure the RX function to use.
1101  *
1102  * @param dev
1103  *   Pointer to private data structure.
1104  *
1105  * @return
1106  *   Pointer to selected Rx burst function.
1107  */
1108 eth_rx_burst_t
1109 mlx5_select_rx_function(struct rte_eth_dev *dev)
1110 {
1111 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1112 
1113 	assert(dev != NULL);
1114 	if (mlx5_check_vec_rx_support(dev) > 0) {
1115 		rx_pkt_burst = mlx5_rx_burst_vec;
1116 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1117 			dev->data->port_id);
1118 	}
1119 	return rx_pkt_burst;
1120 }
1121 
1122 /**
1123  * Check if mlx5 device was removed.
1124  *
1125  * @param dev
1126  *   Pointer to Ethernet device structure.
1127  *
1128  * @return
1129  *   1 when device is removed, otherwise 0.
1130  */
1131 int
1132 mlx5_is_removed(struct rte_eth_dev *dev)
1133 {
1134 	struct ibv_device_attr device_attr;
1135 	struct priv *priv = dev->data->dev_private;
1136 
1137 	if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1138 		return 1;
1139 	return 0;
1140 }
1141