xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 2df00d562d203fb1450261c26d0ffa47b8007089)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <inttypes.h>
8 #include <unistd.h>
9 #include <stdbool.h>
10 #include <stdint.h>
11 #include <stdio.h>
12 #include <string.h>
13 #include <stdlib.h>
14 #include <errno.h>
15 #include <dirent.h>
16 #include <net/if.h>
17 #include <sys/ioctl.h>
18 #include <sys/socket.h>
19 #include <netinet/in.h>
20 #include <linux/ethtool.h>
21 #include <linux/sockios.h>
22 #include <fcntl.h>
23 #include <stdalign.h>
24 #include <sys/un.h>
25 #include <time.h>
26 
27 #include <rte_atomic.h>
28 #include <rte_ethdev_driver.h>
29 #include <rte_bus_pci.h>
30 #include <rte_mbuf.h>
31 #include <rte_common.h>
32 #include <rte_interrupts.h>
33 #include <rte_malloc.h>
34 #include <rte_string_fns.h>
35 #include <rte_rwlock.h>
36 #include <rte_cycles.h>
37 
38 #include <mlx5_glue.h>
39 #include <mlx5_devx_cmds.h>
40 #include <mlx5_common.h>
41 
42 #include "mlx5.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_utils.h"
45 
46 /* Supported speed values found in /usr/include/linux/ethtool.h */
47 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
48 #define SUPPORTED_40000baseKR4_Full (1 << 23)
49 #endif
50 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
51 #define SUPPORTED_40000baseCR4_Full (1 << 24)
52 #endif
53 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
54 #define SUPPORTED_40000baseSR4_Full (1 << 25)
55 #endif
56 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
57 #define SUPPORTED_40000baseLR4_Full (1 << 26)
58 #endif
59 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
60 #define SUPPORTED_56000baseKR4_Full (1 << 27)
61 #endif
62 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
63 #define SUPPORTED_56000baseCR4_Full (1 << 28)
64 #endif
65 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
66 #define SUPPORTED_56000baseSR4_Full (1 << 29)
67 #endif
68 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
69 #define SUPPORTED_56000baseLR4_Full (1 << 30)
70 #endif
71 
72 /* Add defines in case the running kernel is not the same as user headers. */
73 #ifndef ETHTOOL_GLINKSETTINGS
74 struct ethtool_link_settings {
75 	uint32_t cmd;
76 	uint32_t speed;
77 	uint8_t duplex;
78 	uint8_t port;
79 	uint8_t phy_address;
80 	uint8_t autoneg;
81 	uint8_t mdio_support;
82 	uint8_t eth_to_mdix;
83 	uint8_t eth_tp_mdix_ctrl;
84 	int8_t link_mode_masks_nwords;
85 	uint32_t reserved[8];
86 	uint32_t link_mode_masks[];
87 };
88 
89 #define ETHTOOL_GLINKSETTINGS 0x0000004c
90 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
91 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
92 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
93 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
94 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
95 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
96 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
97 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
98 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
99 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
100 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
101 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
102 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
103 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
104 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
105 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
106 #endif
107 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
108 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
109 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
110 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
111 #endif
112 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
113 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
114 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
115 #endif
116 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
117 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
118 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
119 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
120 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
121 #endif
122 
123 /**
124  * Get master interface name from private structure.
125  *
126  * @param[in] dev
127  *   Pointer to Ethernet device.
128  * @param[out] ifname
129  *   Interface name output buffer.
130  *
131  * @return
132  *   0 on success, a negative errno value otherwise and rte_errno is set.
133  */
134 int
135 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
136 {
137 	DIR *dir;
138 	struct dirent *dent;
139 	unsigned int dev_type = 0;
140 	unsigned int dev_port_prev = ~0u;
141 	char match[IF_NAMESIZE] = "";
142 
143 	MLX5_ASSERT(ibdev_path);
144 	{
145 		MKSTR(path, "%s/device/net", ibdev_path);
146 
147 		dir = opendir(path);
148 		if (dir == NULL) {
149 			rte_errno = errno;
150 			return -rte_errno;
151 		}
152 	}
153 	while ((dent = readdir(dir)) != NULL) {
154 		char *name = dent->d_name;
155 		FILE *file;
156 		unsigned int dev_port;
157 		int r;
158 
159 		if ((name[0] == '.') &&
160 		    ((name[1] == '\0') ||
161 		     ((name[1] == '.') && (name[2] == '\0'))))
162 			continue;
163 
164 		MKSTR(path, "%s/device/net/%s/%s",
165 		      ibdev_path, name,
166 		      (dev_type ? "dev_id" : "dev_port"));
167 
168 		file = fopen(path, "rb");
169 		if (file == NULL) {
170 			if (errno != ENOENT)
171 				continue;
172 			/*
173 			 * Switch to dev_id when dev_port does not exist as
174 			 * is the case with Linux kernel versions < 3.15.
175 			 */
176 try_dev_id:
177 			match[0] = '\0';
178 			if (dev_type)
179 				break;
180 			dev_type = 1;
181 			dev_port_prev = ~0u;
182 			rewinddir(dir);
183 			continue;
184 		}
185 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
186 		fclose(file);
187 		if (r != 1)
188 			continue;
189 		/*
190 		 * Switch to dev_id when dev_port returns the same value for
191 		 * all ports. May happen when using a MOFED release older than
192 		 * 3.0 with a Linux kernel >= 3.15.
193 		 */
194 		if (dev_port == dev_port_prev)
195 			goto try_dev_id;
196 		dev_port_prev = dev_port;
197 		if (dev_port == 0)
198 			strlcpy(match, name, sizeof(match));
199 	}
200 	closedir(dir);
201 	if (match[0] == '\0') {
202 		rte_errno = ENOENT;
203 		return -rte_errno;
204 	}
205 	strncpy(*ifname, match, sizeof(*ifname));
206 	return 0;
207 }
208 
209 /**
210  * Get interface name from private structure.
211  *
212  * This is a port representor-aware version of mlx5_get_master_ifname().
213  *
214  * @param[in] dev
215  *   Pointer to Ethernet device.
216  * @param[out] ifname
217  *   Interface name output buffer.
218  *
219  * @return
220  *   0 on success, a negative errno value otherwise and rte_errno is set.
221  */
222 int
223 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
224 {
225 	struct mlx5_priv *priv = dev->data->dev_private;
226 	unsigned int ifindex;
227 
228 	MLX5_ASSERT(priv);
229 	MLX5_ASSERT(priv->sh);
230 	ifindex = mlx5_ifindex(dev);
231 	if (!ifindex) {
232 		if (!priv->representor)
233 			return mlx5_get_master_ifname(priv->sh->ibdev_path,
234 						      ifname);
235 		rte_errno = ENXIO;
236 		return -rte_errno;
237 	}
238 	if (if_indextoname(ifindex, &(*ifname)[0]))
239 		return 0;
240 	rte_errno = errno;
241 	return -rte_errno;
242 }
243 
244 /**
245  * Get the interface index from device name.
246  *
247  * @param[in] dev
248  *   Pointer to Ethernet device.
249  *
250  * @return
251  *   Nonzero interface index on success, zero otherwise and rte_errno is set.
252  */
253 unsigned int
254 mlx5_ifindex(const struct rte_eth_dev *dev)
255 {
256 	struct mlx5_priv *priv = dev->data->dev_private;
257 	unsigned int ifindex;
258 
259 	MLX5_ASSERT(priv);
260 	MLX5_ASSERT(priv->if_index);
261 	ifindex = priv->if_index;
262 	if (!ifindex)
263 		rte_errno = ENXIO;
264 	return ifindex;
265 }
266 
267 /**
268  * Perform ifreq ioctl() on associated Ethernet device.
269  *
270  * @param[in] dev
271  *   Pointer to Ethernet device.
272  * @param req
273  *   Request number to pass to ioctl().
274  * @param[out] ifr
275  *   Interface request structure output buffer.
276  *
277  * @return
278  *   0 on success, a negative errno value otherwise and rte_errno is set.
279  */
280 int
281 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
282 {
283 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
284 	int ret = 0;
285 
286 	if (sock == -1) {
287 		rte_errno = errno;
288 		return -rte_errno;
289 	}
290 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
291 	if (ret)
292 		goto error;
293 	ret = ioctl(sock, req, ifr);
294 	if (ret == -1) {
295 		rte_errno = errno;
296 		goto error;
297 	}
298 	close(sock);
299 	return 0;
300 error:
301 	close(sock);
302 	return -rte_errno;
303 }
304 
305 /**
306  * Get device MTU.
307  *
308  * @param dev
309  *   Pointer to Ethernet device.
310  * @param[out] mtu
311  *   MTU value output buffer.
312  *
313  * @return
314  *   0 on success, a negative errno value otherwise and rte_errno is set.
315  */
316 int
317 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
318 {
319 	struct ifreq request;
320 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
321 
322 	if (ret)
323 		return ret;
324 	*mtu = request.ifr_mtu;
325 	return 0;
326 }
327 
328 /**
329  * Set device MTU.
330  *
331  * @param dev
332  *   Pointer to Ethernet device.
333  * @param mtu
334  *   MTU value to set.
335  *
336  * @return
337  *   0 on success, a negative errno value otherwise and rte_errno is set.
338  */
339 static int
340 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
341 {
342 	struct ifreq request = { .ifr_mtu = mtu, };
343 
344 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
345 }
346 
347 /**
348  * Set device flags.
349  *
350  * @param dev
351  *   Pointer to Ethernet device.
352  * @param keep
353  *   Bitmask for flags that must remain untouched.
354  * @param flags
355  *   Bitmask for flags to modify.
356  *
357  * @return
358  *   0 on success, a negative errno value otherwise and rte_errno is set.
359  */
360 int
361 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
362 {
363 	struct ifreq request;
364 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
365 
366 	if (ret)
367 		return ret;
368 	request.ifr_flags &= keep;
369 	request.ifr_flags |= flags & ~keep;
370 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
371 }
372 
373 /**
374  * DPDK callback for Ethernet device configuration.
375  *
376  * @param dev
377  *   Pointer to Ethernet device structure.
378  *
379  * @return
380  *   0 on success, a negative errno value otherwise and rte_errno is set.
381  */
382 int
383 mlx5_dev_configure(struct rte_eth_dev *dev)
384 {
385 	struct mlx5_priv *priv = dev->data->dev_private;
386 	unsigned int rxqs_n = dev->data->nb_rx_queues;
387 	unsigned int txqs_n = dev->data->nb_tx_queues;
388 	const uint8_t use_app_rss_key =
389 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
390 	int ret = 0;
391 
392 	if (use_app_rss_key &&
393 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
394 	     MLX5_RSS_HASH_KEY_LEN)) {
395 		DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
396 			dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
397 		rte_errno = EINVAL;
398 		return -rte_errno;
399 	}
400 	priv->rss_conf.rss_key =
401 		rte_realloc(priv->rss_conf.rss_key,
402 			    MLX5_RSS_HASH_KEY_LEN, 0);
403 	if (!priv->rss_conf.rss_key) {
404 		DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
405 			dev->data->port_id, rxqs_n);
406 		rte_errno = ENOMEM;
407 		return -rte_errno;
408 	}
409 
410 	if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
411 		dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
412 
413 	memcpy(priv->rss_conf.rss_key,
414 	       use_app_rss_key ?
415 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
416 	       rss_hash_default_key,
417 	       MLX5_RSS_HASH_KEY_LEN);
418 	priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
419 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
420 	priv->rxqs = (void *)dev->data->rx_queues;
421 	priv->txqs = (void *)dev->data->tx_queues;
422 	if (txqs_n != priv->txqs_n) {
423 		DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
424 			dev->data->port_id, priv->txqs_n, txqs_n);
425 		priv->txqs_n = txqs_n;
426 	}
427 	if (rxqs_n > priv->config.ind_table_max_size) {
428 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
429 			dev->data->port_id, rxqs_n);
430 		rte_errno = EINVAL;
431 		return -rte_errno;
432 	}
433 	if (rxqs_n != priv->rxqs_n) {
434 		DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
435 			dev->data->port_id, priv->rxqs_n, rxqs_n);
436 		priv->rxqs_n = rxqs_n;
437 	}
438 	priv->skip_default_rss_reta = 0;
439 	ret = mlx5_proc_priv_init(dev);
440 	if (ret)
441 		return ret;
442 	return 0;
443 }
444 
445 /**
446  * Configure default RSS reta.
447  *
448  * @param dev
449  *   Pointer to Ethernet device structure.
450  *
451  * @return
452  *   0 on success, a negative errno value otherwise and rte_errno is set.
453  */
454 int
455 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
456 {
457 	struct mlx5_priv *priv = dev->data->dev_private;
458 	unsigned int rxqs_n = dev->data->nb_rx_queues;
459 	unsigned int i;
460 	unsigned int j;
461 	unsigned int reta_idx_n;
462 	int ret = 0;
463 	unsigned int *rss_queue_arr = NULL;
464 	unsigned int rss_queue_n = 0;
465 
466 	if (priv->skip_default_rss_reta)
467 		return ret;
468 	rss_queue_arr = rte_malloc("", rxqs_n * sizeof(unsigned int), 0);
469 	if (!rss_queue_arr) {
470 		DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
471 			dev->data->port_id, rxqs_n);
472 		rte_errno = ENOMEM;
473 		return -rte_errno;
474 	}
475 	for (i = 0, j = 0; i < rxqs_n; i++) {
476 		struct mlx5_rxq_data *rxq_data;
477 		struct mlx5_rxq_ctrl *rxq_ctrl;
478 
479 		rxq_data = (*priv->rxqs)[i];
480 		rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
481 		if (rxq_ctrl && rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
482 			rss_queue_arr[j++] = i;
483 	}
484 	rss_queue_n = j;
485 	if (rss_queue_n > priv->config.ind_table_max_size) {
486 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
487 			dev->data->port_id, rss_queue_n);
488 		rte_errno = EINVAL;
489 		rte_free(rss_queue_arr);
490 		return -rte_errno;
491 	}
492 	DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
493 		dev->data->port_id, priv->rxqs_n, rxqs_n);
494 	priv->rxqs_n = rxqs_n;
495 	/*
496 	 * If the requested number of RX queues is not a power of two,
497 	 * use the maximum indirection table size for better balancing.
498 	 * The result is always rounded to the next power of two.
499 	 */
500 	reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
501 				priv->config.ind_table_max_size :
502 				rss_queue_n));
503 	ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
504 	if (ret) {
505 		rte_free(rss_queue_arr);
506 		return ret;
507 	}
508 	/*
509 	 * When the number of RX queues is not a power of two,
510 	 * the remaining table entries are padded with reused WQs
511 	 * and hashes are not spread uniformly.
512 	 */
513 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
514 		(*priv->reta_idx)[i] = rss_queue_arr[j];
515 		if (++j == rss_queue_n)
516 			j = 0;
517 	}
518 	rte_free(rss_queue_arr);
519 	return ret;
520 }
521 
522 /**
523  * Sets default tuning parameters.
524  *
525  * @param dev
526  *   Pointer to Ethernet device.
527  * @param[out] info
528  *   Info structure output buffer.
529  */
530 static void
531 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
532 {
533 	struct mlx5_priv *priv = dev->data->dev_private;
534 
535 	/* Minimum CPU utilization. */
536 	info->default_rxportconf.ring_size = 256;
537 	info->default_txportconf.ring_size = 256;
538 	info->default_rxportconf.burst_size = MLX5_RX_DEFAULT_BURST;
539 	info->default_txportconf.burst_size = MLX5_TX_DEFAULT_BURST;
540 	if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
541 		info->default_rxportconf.nb_queues = 16;
542 		info->default_txportconf.nb_queues = 16;
543 		if (dev->data->nb_rx_queues > 2 ||
544 		    dev->data->nb_tx_queues > 2) {
545 			/* Max Throughput. */
546 			info->default_rxportconf.ring_size = 2048;
547 			info->default_txportconf.ring_size = 2048;
548 		}
549 	} else {
550 		info->default_rxportconf.nb_queues = 8;
551 		info->default_txportconf.nb_queues = 8;
552 		if (dev->data->nb_rx_queues > 2 ||
553 		    dev->data->nb_tx_queues > 2) {
554 			/* Max Throughput. */
555 			info->default_rxportconf.ring_size = 4096;
556 			info->default_txportconf.ring_size = 4096;
557 		}
558 	}
559 }
560 
561 /**
562  * Sets tx mbuf limiting parameters.
563  *
564  * @param dev
565  *   Pointer to Ethernet device.
566  * @param[out] info
567  *   Info structure output buffer.
568  */
569 static void
570 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
571 {
572 	struct mlx5_priv *priv = dev->data->dev_private;
573 	struct mlx5_dev_config *config = &priv->config;
574 	unsigned int inlen;
575 	uint16_t nb_max;
576 
577 	inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
578 		MLX5_SEND_DEF_INLINE_LEN :
579 		(unsigned int)config->txq_inline_max;
580 	MLX5_ASSERT(config->txq_inline_min >= 0);
581 	inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
582 	inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
583 			       MLX5_ESEG_MIN_INLINE_SIZE -
584 			       MLX5_WQE_CSEG_SIZE -
585 			       MLX5_WQE_ESEG_SIZE -
586 			       MLX5_WQE_DSEG_SIZE * 2);
587 	nb_max = (MLX5_WQE_SIZE_MAX +
588 		  MLX5_ESEG_MIN_INLINE_SIZE -
589 		  MLX5_WQE_CSEG_SIZE -
590 		  MLX5_WQE_ESEG_SIZE -
591 		  MLX5_WQE_DSEG_SIZE -
592 		  inlen) / MLX5_WSEG_SIZE;
593 	info->tx_desc_lim.nb_seg_max = nb_max;
594 	info->tx_desc_lim.nb_mtu_seg_max = nb_max;
595 }
596 
597 /**
598  * DPDK callback to get information about the device.
599  *
600  * @param dev
601  *   Pointer to Ethernet device structure.
602  * @param[out] info
603  *   Info structure output buffer.
604  */
605 int
606 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
607 {
608 	struct mlx5_priv *priv = dev->data->dev_private;
609 	struct mlx5_dev_config *config = &priv->config;
610 	unsigned int max;
611 
612 	/* FIXME: we should ask the device for these values. */
613 	info->min_rx_bufsize = 32;
614 	info->max_rx_pktlen = 65536;
615 	info->max_lro_pkt_size = MLX5_MAX_LRO_SIZE;
616 	/*
617 	 * Since we need one CQ per QP, the limit is the minimum number
618 	 * between the two values.
619 	 */
620 	max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
621 		      priv->sh->device_attr.orig_attr.max_qp);
622 	/* max_rx_queues is uint16_t. */
623 	max = RTE_MIN(max, (unsigned int)UINT16_MAX);
624 	info->max_rx_queues = max;
625 	info->max_tx_queues = max;
626 	info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
627 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
628 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
629 				 info->rx_queue_offload_capa);
630 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
631 	info->if_index = mlx5_ifindex(dev);
632 	info->reta_size = priv->reta_idx_n ?
633 		priv->reta_idx_n : config->ind_table_max_size;
634 	info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
635 	info->speed_capa = priv->link_speed_capa;
636 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
637 	mlx5_set_default_params(dev, info);
638 	mlx5_set_txlimit_params(dev, info);
639 	info->switch_info.name = dev->data->name;
640 	info->switch_info.domain_id = priv->domain_id;
641 	info->switch_info.port_id = priv->representor_id;
642 	if (priv->representor) {
643 		uint16_t port_id;
644 
645 		if (priv->pf_bond >= 0) {
646 			/*
647 			 * Switch port ID is opaque value with driver defined
648 			 * format. Push the PF index in bonding configurations
649 			 * in upper four bits of port ID. If we get too many
650 			 * representors (more than 4K) or PFs (more than 15)
651 			 * this approach must be reconsidered.
652 			 */
653 			if ((info->switch_info.port_id >>
654 				MLX5_PORT_ID_BONDING_PF_SHIFT) ||
655 			    priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
656 				DRV_LOG(ERR, "can't update switch port ID"
657 					     " for bonding device");
658 				MLX5_ASSERT(false);
659 				return -ENODEV;
660 			}
661 			info->switch_info.port_id |=
662 				priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
663 		}
664 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
665 			struct mlx5_priv *opriv =
666 				rte_eth_devices[port_id].data->dev_private;
667 
668 			if (!opriv ||
669 			    opriv->representor ||
670 			    opriv->sh != priv->sh ||
671 			    opriv->domain_id != priv->domain_id)
672 				continue;
673 			/*
674 			 * Override switch name with that of the master
675 			 * device.
676 			 */
677 			info->switch_info.name = opriv->dev_data->name;
678 			break;
679 		}
680 	}
681 	return 0;
682 }
683 
684 /**
685  * Get device current raw clock counter
686  *
687  * @param dev
688  *   Pointer to Ethernet device structure.
689  * @param[out] time
690  *   Current raw clock counter of the device.
691  *
692  * @return
693  *   0 if the clock has correctly been read
694  *   The value of errno in case of error
695  */
696 int
697 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
698 {
699 	struct mlx5_priv *priv = dev->data->dev_private;
700 	struct ibv_context *ctx = priv->sh->ctx;
701 	struct ibv_values_ex values;
702 	int err = 0;
703 
704 	values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
705 	err = mlx5_glue->query_rt_values_ex(ctx, &values);
706 	if (err != 0) {
707 		DRV_LOG(WARNING, "Could not query the clock !");
708 		return err;
709 	}
710 	*clock = values.raw_clock.tv_nsec;
711 	return 0;
712 }
713 
714 /**
715  * Get firmware version of a device.
716  *
717  * @param dev
718  *   Ethernet device port.
719  * @param fw_ver
720  *   String output allocated by caller.
721  * @param fw_size
722  *   Size of the output string, including terminating null byte.
723  *
724  * @return
725  *   0 on success, or the size of the non truncated string if too big.
726  */
727 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
728 {
729 	struct mlx5_priv *priv = dev->data->dev_private;
730 	struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
731 	size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
732 
733 	if (fw_size < size)
734 		return size;
735 	if (fw_ver != NULL)
736 		strlcpy(fw_ver, attr->fw_ver, fw_size);
737 	return 0;
738 }
739 
740 /**
741  * Get supported packet types.
742  *
743  * @param dev
744  *   Pointer to Ethernet device structure.
745  *
746  * @return
747  *   A pointer to the supported Packet types array.
748  */
749 const uint32_t *
750 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
751 {
752 	static const uint32_t ptypes[] = {
753 		/* refers to rxq_cq_to_pkt_type() */
754 		RTE_PTYPE_L2_ETHER,
755 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
756 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
757 		RTE_PTYPE_L4_NONFRAG,
758 		RTE_PTYPE_L4_FRAG,
759 		RTE_PTYPE_L4_TCP,
760 		RTE_PTYPE_L4_UDP,
761 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
762 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
763 		RTE_PTYPE_INNER_L4_NONFRAG,
764 		RTE_PTYPE_INNER_L4_FRAG,
765 		RTE_PTYPE_INNER_L4_TCP,
766 		RTE_PTYPE_INNER_L4_UDP,
767 		RTE_PTYPE_UNKNOWN
768 	};
769 
770 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
771 	    dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
772 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
773 		return ptypes;
774 	return NULL;
775 }
776 
777 /**
778  * Retrieve the master device for representor in the same switch domain.
779  *
780  * @param dev
781  *   Pointer to representor Ethernet device structure.
782  *
783  * @return
784  *   Master device structure  on success, NULL otherwise.
785  */
786 
787 static struct rte_eth_dev *
788 mlx5_find_master_dev(struct rte_eth_dev *dev)
789 {
790 	struct mlx5_priv *priv;
791 	uint16_t port_id;
792 	uint16_t domain_id;
793 
794 	priv = dev->data->dev_private;
795 	domain_id = priv->domain_id;
796 	MLX5_ASSERT(priv->representor);
797 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
798 		struct mlx5_priv *opriv =
799 			rte_eth_devices[port_id].data->dev_private;
800 		if (opriv &&
801 		    opriv->master &&
802 		    opriv->domain_id == domain_id &&
803 		    opriv->sh == priv->sh)
804 			return &rte_eth_devices[port_id];
805 	}
806 	return NULL;
807 }
808 
809 /**
810  * DPDK callback to retrieve physical link information.
811  *
812  * @param dev
813  *   Pointer to Ethernet device structure.
814  * @param[out] link
815  *   Storage for current link status.
816  *
817  * @return
818  *   0 on success, a negative errno value otherwise and rte_errno is set.
819  */
820 static int
821 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
822 			       struct rte_eth_link *link)
823 {
824 	struct mlx5_priv *priv = dev->data->dev_private;
825 	struct ethtool_cmd edata = {
826 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
827 	};
828 	struct ifreq ifr;
829 	struct rte_eth_link dev_link;
830 	int link_speed = 0;
831 	int ret;
832 
833 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
834 	if (ret) {
835 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
836 			dev->data->port_id, strerror(rte_errno));
837 		return ret;
838 	}
839 	dev_link = (struct rte_eth_link) {
840 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
841 				(ifr.ifr_flags & IFF_RUNNING)),
842 	};
843 	ifr = (struct ifreq) {
844 		.ifr_data = (void *)&edata,
845 	};
846 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
847 	if (ret) {
848 		if (ret == -ENOTSUP && priv->representor) {
849 			struct rte_eth_dev *master;
850 
851 			/*
852 			 * For representors we can try to inherit link
853 			 * settings from the master device. Actually
854 			 * link settings do not make a lot of sense
855 			 * for representors due to missing physical
856 			 * link. The old kernel drivers supported
857 			 * emulated settings query for representors,
858 			 * the new ones do not, so we have to add
859 			 * this code for compatibility issues.
860 			 */
861 			master = mlx5_find_master_dev(dev);
862 			if (master) {
863 				ifr = (struct ifreq) {
864 					.ifr_data = (void *)&edata,
865 				};
866 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
867 			}
868 		}
869 		if (ret) {
870 			DRV_LOG(WARNING,
871 				"port %u ioctl(SIOCETHTOOL,"
872 				" ETHTOOL_GSET) failed: %s",
873 				dev->data->port_id, strerror(rte_errno));
874 			return ret;
875 		}
876 	}
877 	link_speed = ethtool_cmd_speed(&edata);
878 	if (link_speed == -1)
879 		dev_link.link_speed = ETH_SPEED_NUM_NONE;
880 	else
881 		dev_link.link_speed = link_speed;
882 	priv->link_speed_capa = 0;
883 	if (edata.supported & SUPPORTED_Autoneg)
884 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
885 	if (edata.supported & (SUPPORTED_1000baseT_Full |
886 			       SUPPORTED_1000baseKX_Full))
887 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
888 	if (edata.supported & SUPPORTED_10000baseKR_Full)
889 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
890 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
891 			       SUPPORTED_40000baseCR4_Full |
892 			       SUPPORTED_40000baseSR4_Full |
893 			       SUPPORTED_40000baseLR4_Full))
894 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
895 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
896 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
897 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
898 			ETH_LINK_SPEED_FIXED);
899 	if (((dev_link.link_speed && !dev_link.link_status) ||
900 	     (!dev_link.link_speed && dev_link.link_status))) {
901 		rte_errno = EAGAIN;
902 		return -rte_errno;
903 	}
904 	*link = dev_link;
905 	return 0;
906 }
907 
908 /**
909  * Retrieve physical link information (unlocked version using new ioctl).
910  *
911  * @param dev
912  *   Pointer to Ethernet device structure.
913  * @param[out] link
914  *   Storage for current link status.
915  *
916  * @return
917  *   0 on success, a negative errno value otherwise and rte_errno is set.
918  */
919 static int
920 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
921 			     struct rte_eth_link *link)
922 
923 {
924 	struct mlx5_priv *priv = dev->data->dev_private;
925 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
926 	struct ifreq ifr;
927 	struct rte_eth_link dev_link;
928 	struct rte_eth_dev *master = NULL;
929 	uint64_t sc;
930 	int ret;
931 
932 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
933 	if (ret) {
934 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
935 			dev->data->port_id, strerror(rte_errno));
936 		return ret;
937 	}
938 	dev_link = (struct rte_eth_link) {
939 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
940 				(ifr.ifr_flags & IFF_RUNNING)),
941 	};
942 	ifr = (struct ifreq) {
943 		.ifr_data = (void *)&gcmd,
944 	};
945 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
946 	if (ret) {
947 		if (ret == -ENOTSUP && priv->representor) {
948 			/*
949 			 * For representors we can try to inherit link
950 			 * settings from the master device. Actually
951 			 * link settings do not make a lot of sense
952 			 * for representors due to missing physical
953 			 * link. The old kernel drivers supported
954 			 * emulated settings query for representors,
955 			 * the new ones do not, so we have to add
956 			 * this code for compatibility issues.
957 			 */
958 			master = mlx5_find_master_dev(dev);
959 			if (master) {
960 				ifr = (struct ifreq) {
961 					.ifr_data = (void *)&gcmd,
962 				};
963 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
964 			}
965 		}
966 		if (ret) {
967 			DRV_LOG(DEBUG,
968 				"port %u ioctl(SIOCETHTOOL,"
969 				" ETHTOOL_GLINKSETTINGS) failed: %s",
970 				dev->data->port_id, strerror(rte_errno));
971 			return ret;
972 		}
973 
974 	}
975 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
976 
977 	alignas(struct ethtool_link_settings)
978 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
979 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
980 	struct ethtool_link_settings *ecmd = (void *)data;
981 
982 	*ecmd = gcmd;
983 	ifr.ifr_data = (void *)ecmd;
984 	ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
985 	if (ret) {
986 		DRV_LOG(DEBUG,
987 			"port %u ioctl(SIOCETHTOOL,"
988 			"ETHTOOL_GLINKSETTINGS) failed: %s",
989 			dev->data->port_id, strerror(rte_errno));
990 		return ret;
991 	}
992 	dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
993 							    ecmd->speed;
994 	sc = ecmd->link_mode_masks[0] |
995 		((uint64_t)ecmd->link_mode_masks[1] << 32);
996 	priv->link_speed_capa = 0;
997 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
998 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
999 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
1000 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
1001 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
1002 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
1003 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
1004 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
1005 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
1006 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
1007 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
1008 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
1009 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
1010 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
1011 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
1012 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
1013 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
1014 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
1015 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
1016 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
1017 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
1018 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
1019 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
1020 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
1021 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
1022 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
1023 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
1024 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
1025 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
1026 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
1027 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
1028 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
1029 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
1030 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
1031 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
1032 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
1033 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
1034 				  ETH_LINK_SPEED_FIXED);
1035 	if (((dev_link.link_speed && !dev_link.link_status) ||
1036 	     (!dev_link.link_speed && dev_link.link_status))) {
1037 		rte_errno = EAGAIN;
1038 		return -rte_errno;
1039 	}
1040 	*link = dev_link;
1041 	return 0;
1042 }
1043 
1044 /**
1045  * DPDK callback to retrieve physical link information.
1046  *
1047  * @param dev
1048  *   Pointer to Ethernet device structure.
1049  * @param wait_to_complete
1050  *   Wait for request completion.
1051  *
1052  * @return
1053  *   0 if link status was not updated, positive if it was, a negative errno
1054  *   value otherwise and rte_errno is set.
1055  */
1056 int
1057 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1058 {
1059 	int ret;
1060 	struct rte_eth_link dev_link;
1061 	time_t start_time = time(NULL);
1062 	int retry = MLX5_GET_LINK_STATUS_RETRY_COUNT;
1063 
1064 	do {
1065 		ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1066 		if (ret == -ENOTSUP)
1067 			ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1068 		if (ret == 0)
1069 			break;
1070 		/* Handle wait to complete situation. */
1071 		if ((wait_to_complete || retry) && ret == -EAGAIN) {
1072 			if (abs((int)difftime(time(NULL), start_time)) <
1073 			    MLX5_LINK_STATUS_TIMEOUT) {
1074 				usleep(0);
1075 				continue;
1076 			} else {
1077 				rte_errno = EBUSY;
1078 				return -rte_errno;
1079 			}
1080 		} else if (ret < 0) {
1081 			return ret;
1082 		}
1083 	} while (wait_to_complete || retry-- > 0);
1084 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
1085 		       sizeof(struct rte_eth_link));
1086 	dev->data->dev_link = dev_link;
1087 	return ret;
1088 }
1089 
1090 /**
1091  * DPDK callback to change the MTU.
1092  *
1093  * @param dev
1094  *   Pointer to Ethernet device structure.
1095  * @param in_mtu
1096  *   New MTU.
1097  *
1098  * @return
1099  *   0 on success, a negative errno value otherwise and rte_errno is set.
1100  */
1101 int
1102 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1103 {
1104 	struct mlx5_priv *priv = dev->data->dev_private;
1105 	uint16_t kern_mtu = 0;
1106 	int ret;
1107 
1108 	ret = mlx5_get_mtu(dev, &kern_mtu);
1109 	if (ret)
1110 		return ret;
1111 	/* Set kernel interface MTU first. */
1112 	ret = mlx5_set_mtu(dev, mtu);
1113 	if (ret)
1114 		return ret;
1115 	ret = mlx5_get_mtu(dev, &kern_mtu);
1116 	if (ret)
1117 		return ret;
1118 	if (kern_mtu == mtu) {
1119 		priv->mtu = mtu;
1120 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1121 			dev->data->port_id, mtu);
1122 		return 0;
1123 	}
1124 	rte_errno = EAGAIN;
1125 	return -rte_errno;
1126 }
1127 
1128 /**
1129  * DPDK callback to get flow control status.
1130  *
1131  * @param dev
1132  *   Pointer to Ethernet device structure.
1133  * @param[out] fc_conf
1134  *   Flow control output buffer.
1135  *
1136  * @return
1137  *   0 on success, a negative errno value otherwise and rte_errno is set.
1138  */
1139 int
1140 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1141 {
1142 	struct ifreq ifr;
1143 	struct ethtool_pauseparam ethpause = {
1144 		.cmd = ETHTOOL_GPAUSEPARAM
1145 	};
1146 	int ret;
1147 
1148 	ifr.ifr_data = (void *)&ethpause;
1149 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1150 	if (ret) {
1151 		DRV_LOG(WARNING,
1152 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1153 			" %s",
1154 			dev->data->port_id, strerror(rte_errno));
1155 		return ret;
1156 	}
1157 	fc_conf->autoneg = ethpause.autoneg;
1158 	if (ethpause.rx_pause && ethpause.tx_pause)
1159 		fc_conf->mode = RTE_FC_FULL;
1160 	else if (ethpause.rx_pause)
1161 		fc_conf->mode = RTE_FC_RX_PAUSE;
1162 	else if (ethpause.tx_pause)
1163 		fc_conf->mode = RTE_FC_TX_PAUSE;
1164 	else
1165 		fc_conf->mode = RTE_FC_NONE;
1166 	return 0;
1167 }
1168 
1169 /**
1170  * DPDK callback to modify flow control parameters.
1171  *
1172  * @param dev
1173  *   Pointer to Ethernet device structure.
1174  * @param[in] fc_conf
1175  *   Flow control parameters.
1176  *
1177  * @return
1178  *   0 on success, a negative errno value otherwise and rte_errno is set.
1179  */
1180 int
1181 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1182 {
1183 	struct ifreq ifr;
1184 	struct ethtool_pauseparam ethpause = {
1185 		.cmd = ETHTOOL_SPAUSEPARAM
1186 	};
1187 	int ret;
1188 
1189 	ifr.ifr_data = (void *)&ethpause;
1190 	ethpause.autoneg = fc_conf->autoneg;
1191 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1192 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1193 		ethpause.rx_pause = 1;
1194 	else
1195 		ethpause.rx_pause = 0;
1196 
1197 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1198 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1199 		ethpause.tx_pause = 1;
1200 	else
1201 		ethpause.tx_pause = 0;
1202 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1203 	if (ret) {
1204 		DRV_LOG(WARNING,
1205 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1206 			" failed: %s",
1207 			dev->data->port_id, strerror(rte_errno));
1208 		return ret;
1209 	}
1210 	return 0;
1211 }
1212 
1213 /**
1214  * Handle asynchronous removal event for entire multiport device.
1215  *
1216  * @param sh
1217  *   Infiniband device shared context.
1218  */
1219 static void
1220 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1221 {
1222 	uint32_t i;
1223 
1224 	for (i = 0; i < sh->max_port; ++i) {
1225 		struct rte_eth_dev *dev;
1226 
1227 		if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1228 			/*
1229 			 * Or not existing port either no
1230 			 * handler installed for this port.
1231 			 */
1232 			continue;
1233 		}
1234 		dev = &rte_eth_devices[sh->port[i].ih_port_id];
1235 		MLX5_ASSERT(dev);
1236 		if (dev->data->dev_conf.intr_conf.rmv)
1237 			_rte_eth_dev_callback_process
1238 				(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1239 	}
1240 }
1241 
1242 /**
1243  * Handle shared asynchronous events the NIC (removal event
1244  * and link status change). Supports multiport IB device.
1245  *
1246  * @param cb_arg
1247  *   Callback argument.
1248  */
1249 void
1250 mlx5_dev_interrupt_handler(void *cb_arg)
1251 {
1252 	struct mlx5_ibv_shared *sh = cb_arg;
1253 	struct ibv_async_event event;
1254 
1255 	/* Read all message from the IB device and acknowledge them. */
1256 	for (;;) {
1257 		struct rte_eth_dev *dev;
1258 		uint32_t tmp;
1259 
1260 		if (mlx5_glue->get_async_event(sh->ctx, &event))
1261 			break;
1262 		/* Retrieve and check IB port index. */
1263 		tmp = (uint32_t)event.element.port_num;
1264 		if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1265 			/*
1266 			 * The DEVICE_FATAL event is called once for
1267 			 * entire device without port specifying.
1268 			 * We should notify all existing ports.
1269 			 */
1270 			mlx5_glue->ack_async_event(&event);
1271 			mlx5_dev_interrupt_device_fatal(sh);
1272 			continue;
1273 		}
1274 		MLX5_ASSERT(tmp && (tmp <= sh->max_port));
1275 		if (!tmp) {
1276 			/* Unsupported devive level event. */
1277 			mlx5_glue->ack_async_event(&event);
1278 			DRV_LOG(DEBUG,
1279 				"unsupported common event (type %d)",
1280 				event.event_type);
1281 			continue;
1282 		}
1283 		if (tmp > sh->max_port) {
1284 			/* Invalid IB port index. */
1285 			mlx5_glue->ack_async_event(&event);
1286 			DRV_LOG(DEBUG,
1287 				"cannot handle an event (type %d)"
1288 				"due to invalid IB port index (%u)",
1289 				event.event_type, tmp);
1290 			continue;
1291 		}
1292 		if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1293 			/* No handler installed. */
1294 			mlx5_glue->ack_async_event(&event);
1295 			DRV_LOG(DEBUG,
1296 				"cannot handle an event (type %d)"
1297 				"due to no handler installed for port %u",
1298 				event.event_type, tmp);
1299 			continue;
1300 		}
1301 		/* Retrieve ethernet device descriptor. */
1302 		tmp = sh->port[tmp - 1].ih_port_id;
1303 		dev = &rte_eth_devices[tmp];
1304 		MLX5_ASSERT(dev);
1305 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1306 		     event.event_type == IBV_EVENT_PORT_ERR) &&
1307 			dev->data->dev_conf.intr_conf.lsc) {
1308 			mlx5_glue->ack_async_event(&event);
1309 			if (mlx5_link_update(dev, 0) == -EAGAIN) {
1310 				usleep(0);
1311 				continue;
1312 			}
1313 			_rte_eth_dev_callback_process
1314 				(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1315 			continue;
1316 		}
1317 		DRV_LOG(DEBUG,
1318 			"port %u cannot handle an unknown event (type %d)",
1319 			dev->data->port_id, event.event_type);
1320 		mlx5_glue->ack_async_event(&event);
1321 	}
1322 }
1323 
1324 /*
1325  * Unregister callback handler safely. The handler may be active
1326  * while we are trying to unregister it, in this case code -EAGAIN
1327  * is returned by rte_intr_callback_unregister(). This routine checks
1328  * the return code and tries to unregister handler again.
1329  *
1330  * @param handle
1331  *   interrupt handle
1332  * @param cb_fn
1333  *   pointer to callback routine
1334  * @cb_arg
1335  *   opaque callback parameter
1336  */
1337 void
1338 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1339 			      rte_intr_callback_fn cb_fn, void *cb_arg)
1340 {
1341 	/*
1342 	 * Try to reduce timeout management overhead by not calling
1343 	 * the timer related routines on the first iteration. If the
1344 	 * unregistering succeeds on first call there will be no
1345 	 * timer calls at all.
1346 	 */
1347 	uint64_t twait = 0;
1348 	uint64_t start = 0;
1349 
1350 	do {
1351 		int ret;
1352 
1353 		ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1354 		if (ret >= 0)
1355 			return;
1356 		if (ret != -EAGAIN) {
1357 			DRV_LOG(INFO, "failed to unregister interrupt"
1358 				      " handler (error: %d)", ret);
1359 			MLX5_ASSERT(false);
1360 			return;
1361 		}
1362 		if (twait) {
1363 			struct timespec onems;
1364 
1365 			/* Wait one millisecond and try again. */
1366 			onems.tv_sec = 0;
1367 			onems.tv_nsec = NS_PER_S / MS_PER_S;
1368 			nanosleep(&onems, 0);
1369 			/* Check whether one second elapsed. */
1370 			if ((rte_get_timer_cycles() - start) <= twait)
1371 				continue;
1372 		} else {
1373 			/*
1374 			 * We get the amount of timer ticks for one second.
1375 			 * If this amount elapsed it means we spent one
1376 			 * second in waiting. This branch is executed once
1377 			 * on first iteration.
1378 			 */
1379 			twait = rte_get_timer_hz();
1380 			MLX5_ASSERT(twait);
1381 		}
1382 		/*
1383 		 * Timeout elapsed, show message (once a second) and retry.
1384 		 * We have no other acceptable option here, if we ignore
1385 		 * the unregistering return code the handler will not
1386 		 * be unregistered, fd will be closed and we may get the
1387 		 * crush. Hanging and messaging in the loop seems not to be
1388 		 * the worst choice.
1389 		 */
1390 		DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1391 		start = rte_get_timer_cycles();
1392 	} while (true);
1393 }
1394 
1395 /**
1396  * Handle DEVX interrupts from the NIC.
1397  * This function is probably called from the DPDK host thread.
1398  *
1399  * @param cb_arg
1400  *   Callback argument.
1401  */
1402 void
1403 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1404 {
1405 #ifndef HAVE_IBV_DEVX_ASYNC
1406 	(void)cb_arg;
1407 	return;
1408 #else
1409 	struct mlx5_ibv_shared *sh = cb_arg;
1410 	union {
1411 		struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1412 		uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1413 			    MLX5_ST_SZ_BYTES(traffic_counter) +
1414 			    sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1415 	} out;
1416 	uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1417 
1418 	while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1419 						   &out.cmd_resp,
1420 						   sizeof(out.buf)))
1421 		mlx5_flow_async_pool_query_handle
1422 			(sh, (uint64_t)out.cmd_resp.wr_id,
1423 			 mlx5_devx_get_out_command_status(buf));
1424 #endif /* HAVE_IBV_DEVX_ASYNC */
1425 }
1426 
1427 /**
1428  * Uninstall shared asynchronous device events handler.
1429  * This function is implemented to support event sharing
1430  * between multiple ports of single IB device.
1431  *
1432  * @param dev
1433  *   Pointer to Ethernet device.
1434  */
1435 static void
1436 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1437 {
1438 	struct mlx5_priv *priv = dev->data->dev_private;
1439 	struct mlx5_ibv_shared *sh = priv->sh;
1440 
1441 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1442 		return;
1443 	pthread_mutex_lock(&sh->intr_mutex);
1444 	MLX5_ASSERT(priv->ibv_port);
1445 	MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1446 	MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1447 	if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1448 		goto exit;
1449 	MLX5_ASSERT(sh->port[priv->ibv_port - 1].ih_port_id ==
1450 					(uint32_t)dev->data->port_id);
1451 	MLX5_ASSERT(sh->intr_cnt);
1452 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1453 	if (!sh->intr_cnt || --sh->intr_cnt)
1454 		goto exit;
1455 	mlx5_intr_callback_unregister(&sh->intr_handle,
1456 				     mlx5_dev_interrupt_handler, sh);
1457 	sh->intr_handle.fd = 0;
1458 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1459 exit:
1460 	pthread_mutex_unlock(&sh->intr_mutex);
1461 }
1462 
1463 /**
1464  * Uninstall devx shared asynchronous device events handler.
1465  * This function is implemeted to support event sharing
1466  * between multiple ports of single IB device.
1467  *
1468  * @param dev
1469  *   Pointer to Ethernet device.
1470  */
1471 static void
1472 mlx5_dev_shared_handler_devx_uninstall(struct rte_eth_dev *dev)
1473 {
1474 	struct mlx5_priv *priv = dev->data->dev_private;
1475 	struct mlx5_ibv_shared *sh = priv->sh;
1476 
1477 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1478 		return;
1479 	pthread_mutex_lock(&sh->intr_mutex);
1480 	MLX5_ASSERT(priv->ibv_port);
1481 	MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1482 	MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1483 	if (sh->port[priv->ibv_port - 1].devx_ih_port_id >= RTE_MAX_ETHPORTS)
1484 		goto exit;
1485 	MLX5_ASSERT(sh->port[priv->ibv_port - 1].devx_ih_port_id ==
1486 		    (uint32_t)dev->data->port_id);
1487 	sh->port[priv->ibv_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1488 	if (!sh->devx_intr_cnt || --sh->devx_intr_cnt)
1489 		goto exit;
1490 	if (sh->intr_handle_devx.fd) {
1491 		rte_intr_callback_unregister(&sh->intr_handle_devx,
1492 					     mlx5_dev_interrupt_handler_devx,
1493 					     sh);
1494 		sh->intr_handle_devx.fd = 0;
1495 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1496 	}
1497 	if (sh->devx_comp) {
1498 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1499 		sh->devx_comp = NULL;
1500 	}
1501 exit:
1502 	pthread_mutex_unlock(&sh->intr_mutex);
1503 }
1504 
1505 /**
1506  * Install shared asynchronous device events handler.
1507  * This function is implemented to support event sharing
1508  * between multiple ports of single IB device.
1509  *
1510  * @param dev
1511  *   Pointer to Ethernet device.
1512  */
1513 static void
1514 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1515 {
1516 	struct mlx5_priv *priv = dev->data->dev_private;
1517 	struct mlx5_ibv_shared *sh = priv->sh;
1518 	int ret;
1519 	int flags;
1520 
1521 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1522 		return;
1523 	pthread_mutex_lock(&sh->intr_mutex);
1524 	MLX5_ASSERT(priv->ibv_port);
1525 	MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1526 	MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1527 	if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1528 		/* The handler is already installed for this port. */
1529 		MLX5_ASSERT(sh->intr_cnt);
1530 		goto exit;
1531 	}
1532 	if (sh->intr_cnt) {
1533 		sh->port[priv->ibv_port - 1].ih_port_id =
1534 						(uint32_t)dev->data->port_id;
1535 		sh->intr_cnt++;
1536 		goto exit;
1537 	}
1538 	/* No shared handler installed. */
1539 	MLX5_ASSERT(sh->ctx->async_fd > 0);
1540 	flags = fcntl(sh->ctx->async_fd, F_GETFL);
1541 	ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1542 	if (ret) {
1543 		DRV_LOG(INFO, "failed to change file descriptor async event"
1544 			" queue");
1545 		/* Indicate there will be no interrupts. */
1546 		dev->data->dev_conf.intr_conf.lsc = 0;
1547 		dev->data->dev_conf.intr_conf.rmv = 0;
1548 	} else {
1549 		sh->intr_handle.fd = sh->ctx->async_fd;
1550 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1551 		rte_intr_callback_register(&sh->intr_handle,
1552 					   mlx5_dev_interrupt_handler, sh);
1553 		sh->intr_cnt++;
1554 		sh->port[priv->ibv_port - 1].ih_port_id =
1555 						(uint32_t)dev->data->port_id;
1556 	}
1557 exit:
1558 	pthread_mutex_unlock(&sh->intr_mutex);
1559 }
1560 
1561 /**
1562  * Install devx shared asyncronous device events handler.
1563  * This function is implemeted to support event sharing
1564  * between multiple ports of single IB device.
1565  *
1566  * @param dev
1567  *   Pointer to Ethernet device.
1568  */
1569 static void
1570 mlx5_dev_shared_handler_devx_install(struct rte_eth_dev *dev)
1571 {
1572 	struct mlx5_priv *priv = dev->data->dev_private;
1573 	struct mlx5_ibv_shared *sh = priv->sh;
1574 
1575 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1576 		return;
1577 	pthread_mutex_lock(&sh->intr_mutex);
1578 	MLX5_ASSERT(priv->ibv_port);
1579 	MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1580 	MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1581 	if (sh->port[priv->ibv_port - 1].devx_ih_port_id < RTE_MAX_ETHPORTS) {
1582 		/* The handler is already installed for this port. */
1583 		MLX5_ASSERT(sh->devx_intr_cnt);
1584 		goto exit;
1585 	}
1586 	if (sh->devx_intr_cnt) {
1587 		sh->devx_intr_cnt++;
1588 		sh->port[priv->ibv_port - 1].devx_ih_port_id =
1589 					(uint32_t)dev->data->port_id;
1590 		goto exit;
1591 	}
1592 	if (priv->config.devx) {
1593 #ifndef HAVE_IBV_DEVX_ASYNC
1594 		goto exit;
1595 #else
1596 		sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1597 		if (sh->devx_comp) {
1598 			int flags = fcntl(sh->devx_comp->fd, F_GETFL);
1599 			int ret = fcntl(sh->devx_comp->fd, F_SETFL,
1600 				    flags | O_NONBLOCK);
1601 
1602 			if (ret) {
1603 				DRV_LOG(INFO, "failed to change file descriptor"
1604 					" devx async event queue");
1605 			} else {
1606 				sh->intr_handle_devx.fd = sh->devx_comp->fd;
1607 				sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1608 				rte_intr_callback_register
1609 					(&sh->intr_handle_devx,
1610 					 mlx5_dev_interrupt_handler_devx, sh);
1611 				sh->devx_intr_cnt++;
1612 				sh->port[priv->ibv_port - 1].devx_ih_port_id =
1613 						(uint32_t)dev->data->port_id;
1614 			}
1615 		}
1616 #endif /* HAVE_IBV_DEVX_ASYNC */
1617 	}
1618 exit:
1619 	pthread_mutex_unlock(&sh->intr_mutex);
1620 }
1621 
1622 /**
1623  * Uninstall interrupt handler.
1624  *
1625  * @param dev
1626  *   Pointer to Ethernet device.
1627  */
1628 void
1629 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1630 {
1631 	mlx5_dev_shared_handler_uninstall(dev);
1632 }
1633 
1634 /**
1635  * Install interrupt handler.
1636  *
1637  * @param dev
1638  *   Pointer to Ethernet device.
1639  */
1640 void
1641 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1642 {
1643 	mlx5_dev_shared_handler_install(dev);
1644 }
1645 
1646 /**
1647  * Devx uninstall interrupt handler.
1648  *
1649  * @param dev
1650  *   Pointer to Ethernet device.
1651  */
1652 void
1653 mlx5_dev_interrupt_handler_devx_uninstall(struct rte_eth_dev *dev)
1654 {
1655 	mlx5_dev_shared_handler_devx_uninstall(dev);
1656 }
1657 
1658 /**
1659  * Devx install interrupt handler.
1660  *
1661  * @param dev
1662  *   Pointer to Ethernet device.
1663  */
1664 void
1665 mlx5_dev_interrupt_handler_devx_install(struct rte_eth_dev *dev)
1666 {
1667 	mlx5_dev_shared_handler_devx_install(dev);
1668 }
1669 
1670 /**
1671  * DPDK callback to bring the link DOWN.
1672  *
1673  * @param dev
1674  *   Pointer to Ethernet device structure.
1675  *
1676  * @return
1677  *   0 on success, a negative errno value otherwise and rte_errno is set.
1678  */
1679 int
1680 mlx5_set_link_down(struct rte_eth_dev *dev)
1681 {
1682 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1683 }
1684 
1685 /**
1686  * DPDK callback to bring the link UP.
1687  *
1688  * @param dev
1689  *   Pointer to Ethernet device structure.
1690  *
1691  * @return
1692  *   0 on success, a negative errno value otherwise and rte_errno is set.
1693  */
1694 int
1695 mlx5_set_link_up(struct rte_eth_dev *dev)
1696 {
1697 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1698 }
1699 
1700 /**
1701  * Configure the RX function to use.
1702  *
1703  * @param dev
1704  *   Pointer to private data structure.
1705  *
1706  * @return
1707  *   Pointer to selected Rx burst function.
1708  */
1709 eth_rx_burst_t
1710 mlx5_select_rx_function(struct rte_eth_dev *dev)
1711 {
1712 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1713 
1714 	MLX5_ASSERT(dev != NULL);
1715 	if (mlx5_check_vec_rx_support(dev) > 0) {
1716 		rx_pkt_burst = mlx5_rx_burst_vec;
1717 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1718 			dev->data->port_id);
1719 	} else if (mlx5_mprq_enabled(dev)) {
1720 		rx_pkt_burst = mlx5_rx_burst_mprq;
1721 	}
1722 	return rx_pkt_burst;
1723 }
1724 
1725 /**
1726  * Check if mlx5 device was removed.
1727  *
1728  * @param dev
1729  *   Pointer to Ethernet device structure.
1730  *
1731  * @return
1732  *   1 when device is removed, otherwise 0.
1733  */
1734 int
1735 mlx5_is_removed(struct rte_eth_dev *dev)
1736 {
1737 	struct ibv_device_attr device_attr;
1738 	struct mlx5_priv *priv = dev->data->dev_private;
1739 
1740 	if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1741 		return 1;
1742 	return 0;
1743 }
1744 
1745 /**
1746  * Get the E-Switch parameters by port id.
1747  *
1748  * @param[in] port
1749  *   Device port id.
1750  * @param[in] valid
1751  *   Device port id is valid, skip check. This flag is useful
1752  *   when trials are performed from probing and device is not
1753  *   flagged as valid yet (in attaching process).
1754  * @param[out] es_domain_id
1755  *   E-Switch domain id.
1756  * @param[out] es_port_id
1757  *   The port id of the port in the E-Switch.
1758  *
1759  * @return
1760  *   pointer to device private data structure containing data needed
1761  *   on success, NULL otherwise and rte_errno is set.
1762  */
1763 struct mlx5_priv *
1764 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
1765 {
1766 	struct rte_eth_dev *dev;
1767 	struct mlx5_priv *priv;
1768 
1769 	if (port >= RTE_MAX_ETHPORTS) {
1770 		rte_errno = EINVAL;
1771 		return NULL;
1772 	}
1773 	if (!valid && !rte_eth_dev_is_valid_port(port)) {
1774 		rte_errno = ENODEV;
1775 		return NULL;
1776 	}
1777 	dev = &rte_eth_devices[port];
1778 	priv = dev->data->dev_private;
1779 	if (!(priv->representor || priv->master)) {
1780 		rte_errno = EINVAL;
1781 		return NULL;
1782 	}
1783 	return priv;
1784 }
1785 
1786 /**
1787  * Get the E-Switch parameters by device instance.
1788  *
1789  * @param[in] port
1790  *   Device port id.
1791  * @param[out] es_domain_id
1792  *   E-Switch domain id.
1793  * @param[out] es_port_id
1794  *   The port id of the port in the E-Switch.
1795  *
1796  * @return
1797  *   pointer to device private data structure containing data needed
1798  *   on success, NULL otherwise and rte_errno is set.
1799  */
1800 struct mlx5_priv *
1801 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1802 {
1803 	struct mlx5_priv *priv;
1804 
1805 	priv = dev->data->dev_private;
1806 	if (!(priv->representor || priv->master)) {
1807 		rte_errno = EINVAL;
1808 		return NULL;
1809 	}
1810 	return priv;
1811 }
1812 
1813 /**
1814  * Get switch information associated with network interface.
1815  *
1816  * @param ifindex
1817  *   Network interface index.
1818  * @param[out] info
1819  *   Switch information object, populated in case of success.
1820  *
1821  * @return
1822  *   0 on success, a negative errno value otherwise and rte_errno is set.
1823  */
1824 int
1825 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1826 {
1827 	char ifname[IF_NAMESIZE];
1828 	char port_name[IF_NAMESIZE];
1829 	FILE *file;
1830 	struct mlx5_switch_info data = {
1831 		.master = 0,
1832 		.representor = 0,
1833 		.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1834 		.port_name = 0,
1835 		.switch_id = 0,
1836 	};
1837 	DIR *dir;
1838 	bool port_switch_id_set = false;
1839 	bool device_dir = false;
1840 	char c;
1841 	int ret;
1842 
1843 	if (!if_indextoname(ifindex, ifname)) {
1844 		rte_errno = errno;
1845 		return -rte_errno;
1846 	}
1847 
1848 	MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1849 	      ifname);
1850 	MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1851 	      ifname);
1852 	MKSTR(pci_device, "/sys/class/net/%s/device",
1853 	      ifname);
1854 
1855 	file = fopen(phys_port_name, "rb");
1856 	if (file != NULL) {
1857 		ret = fscanf(file, "%s", port_name);
1858 		fclose(file);
1859 		if (ret == 1)
1860 			mlx5_translate_port_name(port_name, &data);
1861 	}
1862 	file = fopen(phys_switch_id, "rb");
1863 	if (file == NULL) {
1864 		rte_errno = errno;
1865 		return -rte_errno;
1866 	}
1867 	port_switch_id_set =
1868 		fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1869 		c == '\n';
1870 	fclose(file);
1871 	dir = opendir(pci_device);
1872 	if (dir != NULL) {
1873 		closedir(dir);
1874 		device_dir = true;
1875 	}
1876 	if (port_switch_id_set) {
1877 		/* We have some E-Switch configuration. */
1878 		mlx5_sysfs_check_switch_info(device_dir, &data);
1879 	}
1880 	*info = data;
1881 	MLX5_ASSERT(!(data.master && data.representor));
1882 	if (data.master && data.representor) {
1883 		DRV_LOG(ERR, "ifindex %u device is recognized as master"
1884 			     " and as representor", ifindex);
1885 		rte_errno = ENODEV;
1886 		return -rte_errno;
1887 	}
1888 	return 0;
1889 }
1890 
1891 /**
1892  * Analyze gathered port parameters via sysfs to recognize master
1893  * and representor devices for E-Switch configuration.
1894  *
1895  * @param[in] device_dir
1896  *   flag of presence of "device" directory under port device key.
1897  * @param[inout] switch_info
1898  *   Port information, including port name as a number and port name
1899  *   type if recognized
1900  *
1901  * @return
1902  *   master and representor flags are set in switch_info according to
1903  *   recognized parameters (if any).
1904  */
1905 void
1906 mlx5_sysfs_check_switch_info(bool device_dir,
1907 			     struct mlx5_switch_info *switch_info)
1908 {
1909 	switch (switch_info->name_type) {
1910 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1911 		/*
1912 		 * Name is not recognized, assume the master,
1913 		 * check the device directory presence.
1914 		 */
1915 		switch_info->master = device_dir;
1916 		break;
1917 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1918 		/*
1919 		 * Name is not set, this assumes the legacy naming
1920 		 * schema for master, just check if there is
1921 		 * a device directory.
1922 		 */
1923 		switch_info->master = device_dir;
1924 		break;
1925 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1926 		/* New uplink naming schema recognized. */
1927 		switch_info->master = 1;
1928 		break;
1929 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1930 		/* Legacy representors naming schema. */
1931 		switch_info->representor = !device_dir;
1932 		break;
1933 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1934 		/* New representors naming schema. */
1935 		switch_info->representor = 1;
1936 		break;
1937 	}
1938 }
1939 
1940 /**
1941  * DPDK callback to retrieve plug-in module EEPROM information (type and size).
1942  *
1943  * @param dev
1944  *   Pointer to Ethernet device structure.
1945  * @param[out] modinfo
1946  *   Storage for plug-in module EEPROM information.
1947  *
1948  * @return
1949  *   0 on success, a negative errno value otherwise and rte_errno is set.
1950  */
1951 int
1952 mlx5_get_module_info(struct rte_eth_dev *dev,
1953 		     struct rte_eth_dev_module_info *modinfo)
1954 {
1955 	struct ethtool_modinfo info = {
1956 		.cmd = ETHTOOL_GMODULEINFO,
1957 	};
1958 	struct ifreq ifr = (struct ifreq) {
1959 		.ifr_data = (void *)&info,
1960 	};
1961 	int ret = 0;
1962 
1963 	if (!dev || !modinfo) {
1964 		DRV_LOG(WARNING, "missing argument, cannot get module info");
1965 		rte_errno = EINVAL;
1966 		return -rte_errno;
1967 	}
1968 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1969 	if (ret) {
1970 		DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
1971 			dev->data->port_id, strerror(rte_errno));
1972 		return ret;
1973 	}
1974 	modinfo->type = info.type;
1975 	modinfo->eeprom_len = info.eeprom_len;
1976 	return ret;
1977 }
1978 
1979 /**
1980  * DPDK callback to retrieve plug-in module EEPROM data.
1981  *
1982  * @param dev
1983  *   Pointer to Ethernet device structure.
1984  * @param[out] info
1985  *   Storage for plug-in module EEPROM data.
1986  *
1987  * @return
1988  *   0 on success, a negative errno value otherwise and rte_errno is set.
1989  */
1990 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
1991 			   struct rte_dev_eeprom_info *info)
1992 {
1993 	struct ethtool_eeprom *eeprom;
1994 	struct ifreq ifr;
1995 	int ret = 0;
1996 
1997 	if (!dev || !info) {
1998 		DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
1999 		rte_errno = EINVAL;
2000 		return -rte_errno;
2001 	}
2002 	eeprom = rte_calloc(__func__, 1,
2003 			    (sizeof(struct ethtool_eeprom) + info->length), 0);
2004 	if (!eeprom) {
2005 		DRV_LOG(WARNING, "port %u cannot allocate memory for "
2006 			"eeprom data", dev->data->port_id);
2007 		rte_errno = ENOMEM;
2008 		return -rte_errno;
2009 	}
2010 	eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2011 	eeprom->offset = info->offset;
2012 	eeprom->len = info->length;
2013 	ifr = (struct ifreq) {
2014 		.ifr_data = (void *)eeprom,
2015 	};
2016 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2017 	if (ret)
2018 		DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2019 			dev->data->port_id, strerror(rte_errno));
2020 	else
2021 		rte_memcpy(info->data, eeprom->data, info->length);
2022 	rte_free(eeprom);
2023 	return ret;
2024 }
2025 
2026 /**
2027  * DPDK callback to retrieve hairpin capabilities.
2028  *
2029  * @param dev
2030  *   Pointer to Ethernet device structure.
2031  * @param[out] cap
2032  *   Storage for hairpin capability data.
2033  *
2034  * @return
2035  *   0 on success, a negative errno value otherwise and rte_errno is set.
2036  */
2037 int mlx5_hairpin_cap_get(struct rte_eth_dev *dev,
2038 			 struct rte_eth_hairpin_cap *cap)
2039 {
2040 	struct mlx5_priv *priv = dev->data->dev_private;
2041 
2042 	if (priv->sh->devx == 0) {
2043 		rte_errno = ENOTSUP;
2044 		return -rte_errno;
2045 	}
2046 	cap->max_nb_queues = UINT16_MAX;
2047 	cap->max_rx_2_tx = 1;
2048 	cap->max_tx_2_rx = 1;
2049 	cap->max_nb_desc = 8192;
2050 	return 0;
2051 }
2052