xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision ff45f462b87debf8f8034e5df0e4c2b7fe162559)
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 <assert.h>
8 #include <inttypes.h>
9 #include <unistd.h>
10 #include <stdbool.h>
11 #include <stdint.h>
12 #include <stdio.h>
13 #include <string.h>
14 #include <stdlib.h>
15 #include <errno.h>
16 #include <dirent.h>
17 #include <net/if.h>
18 #include <sys/ioctl.h>
19 #include <sys/socket.h>
20 #include <netinet/in.h>
21 #include <linux/ethtool.h>
22 #include <linux/sockios.h>
23 #include <fcntl.h>
24 #include <stdalign.h>
25 #include <sys/un.h>
26 #include <time.h>
27 
28 #include <rte_atomic.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_bus_pci.h>
31 #include <rte_mbuf.h>
32 #include <rte_common.h>
33 #include <rte_interrupts.h>
34 #include <rte_malloc.h>
35 #include <rte_string_fns.h>
36 #include <rte_rwlock.h>
37 #include <rte_cycles.h>
38 
39 #include "mlx5.h"
40 #include "mlx5_glue.h"
41 #include "mlx5_rxtx.h"
42 #include "mlx5_utils.h"
43 
44 /* Supported speed values found in /usr/include/linux/ethtool.h */
45 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
46 #define SUPPORTED_40000baseKR4_Full (1 << 23)
47 #endif
48 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
49 #define SUPPORTED_40000baseCR4_Full (1 << 24)
50 #endif
51 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
52 #define SUPPORTED_40000baseSR4_Full (1 << 25)
53 #endif
54 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
55 #define SUPPORTED_40000baseLR4_Full (1 << 26)
56 #endif
57 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
58 #define SUPPORTED_56000baseKR4_Full (1 << 27)
59 #endif
60 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
61 #define SUPPORTED_56000baseCR4_Full (1 << 28)
62 #endif
63 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
64 #define SUPPORTED_56000baseSR4_Full (1 << 29)
65 #endif
66 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
67 #define SUPPORTED_56000baseLR4_Full (1 << 30)
68 #endif
69 
70 /* Add defines in case the running kernel is not the same as user headers. */
71 #ifndef ETHTOOL_GLINKSETTINGS
72 struct ethtool_link_settings {
73 	uint32_t cmd;
74 	uint32_t speed;
75 	uint8_t duplex;
76 	uint8_t port;
77 	uint8_t phy_address;
78 	uint8_t autoneg;
79 	uint8_t mdio_support;
80 	uint8_t eth_to_mdix;
81 	uint8_t eth_tp_mdix_ctrl;
82 	int8_t link_mode_masks_nwords;
83 	uint32_t reserved[8];
84 	uint32_t link_mode_masks[];
85 };
86 
87 #define ETHTOOL_GLINKSETTINGS 0x0000004c
88 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
89 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
90 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
91 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
92 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
93 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
94 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
95 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
96 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
97 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
98 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
99 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
100 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
101 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
102 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
103 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
104 #endif
105 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
106 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
107 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
108 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
109 #endif
110 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
111 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
112 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
113 #endif
114 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
115 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
116 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
117 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
118 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
119 #endif
120 
121 /**
122  * Get master interface name from private structure.
123  *
124  * @param[in] dev
125  *   Pointer to Ethernet device.
126  * @param[out] ifname
127  *   Interface name output buffer.
128  *
129  * @return
130  *   0 on success, a negative errno value otherwise and rte_errno is set.
131  */
132 int
133 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
134 {
135 	DIR *dir;
136 	struct dirent *dent;
137 	unsigned int dev_type = 0;
138 	unsigned int dev_port_prev = ~0u;
139 	char match[IF_NAMESIZE] = "";
140 
141 	assert(ibdev_path);
142 	{
143 		MKSTR(path, "%s/device/net", ibdev_path);
144 
145 		dir = opendir(path);
146 		if (dir == NULL) {
147 			rte_errno = errno;
148 			return -rte_errno;
149 		}
150 	}
151 	while ((dent = readdir(dir)) != NULL) {
152 		char *name = dent->d_name;
153 		FILE *file;
154 		unsigned int dev_port;
155 		int r;
156 
157 		if ((name[0] == '.') &&
158 		    ((name[1] == '\0') ||
159 		     ((name[1] == '.') && (name[2] == '\0'))))
160 			continue;
161 
162 		MKSTR(path, "%s/device/net/%s/%s",
163 		      ibdev_path, name,
164 		      (dev_type ? "dev_id" : "dev_port"));
165 
166 		file = fopen(path, "rb");
167 		if (file == NULL) {
168 			if (errno != ENOENT)
169 				continue;
170 			/*
171 			 * Switch to dev_id when dev_port does not exist as
172 			 * is the case with Linux kernel versions < 3.15.
173 			 */
174 try_dev_id:
175 			match[0] = '\0';
176 			if (dev_type)
177 				break;
178 			dev_type = 1;
179 			dev_port_prev = ~0u;
180 			rewinddir(dir);
181 			continue;
182 		}
183 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
184 		fclose(file);
185 		if (r != 1)
186 			continue;
187 		/*
188 		 * Switch to dev_id when dev_port returns the same value for
189 		 * all ports. May happen when using a MOFED release older than
190 		 * 3.0 with a Linux kernel >= 3.15.
191 		 */
192 		if (dev_port == dev_port_prev)
193 			goto try_dev_id;
194 		dev_port_prev = dev_port;
195 		if (dev_port == 0)
196 			strlcpy(match, name, sizeof(match));
197 	}
198 	closedir(dir);
199 	if (match[0] == '\0') {
200 		rte_errno = ENOENT;
201 		return -rte_errno;
202 	}
203 	strncpy(*ifname, match, sizeof(*ifname));
204 	return 0;
205 }
206 
207 /**
208  * Get interface name from private structure.
209  *
210  * This is a port representor-aware version of mlx5_get_master_ifname().
211  *
212  * @param[in] dev
213  *   Pointer to Ethernet device.
214  * @param[out] ifname
215  *   Interface name output buffer.
216  *
217  * @return
218  *   0 on success, a negative errno value otherwise and rte_errno is set.
219  */
220 int
221 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
222 {
223 	struct mlx5_priv *priv = dev->data->dev_private;
224 	unsigned int ifindex;
225 
226 	assert(priv);
227 	assert(priv->sh);
228 	ifindex = mlx5_ifindex(dev);
229 	if (!ifindex) {
230 		if (!priv->representor)
231 			return mlx5_get_master_ifname(priv->sh->ibdev_path,
232 						      ifname);
233 		rte_errno = ENXIO;
234 		return -rte_errno;
235 	}
236 	if (if_indextoname(ifindex, &(*ifname)[0]))
237 		return 0;
238 	rte_errno = errno;
239 	return -rte_errno;
240 }
241 
242 /**
243  * Get the interface index from device name.
244  *
245  * @param[in] dev
246  *   Pointer to Ethernet device.
247  *
248  * @return
249  *   Nonzero interface index on success, zero otherwise and rte_errno is set.
250  */
251 unsigned int
252 mlx5_ifindex(const struct rte_eth_dev *dev)
253 {
254 	struct mlx5_priv *priv = dev->data->dev_private;
255 	unsigned int ifindex;
256 
257 	assert(priv);
258 	assert(priv->if_index);
259 	ifindex = priv->if_index;
260 	if (!ifindex)
261 		rte_errno = ENXIO;
262 	return ifindex;
263 }
264 
265 /**
266  * Perform ifreq ioctl() on associated Ethernet device.
267  *
268  * @param[in] dev
269  *   Pointer to Ethernet device.
270  * @param req
271  *   Request number to pass to ioctl().
272  * @param[out] ifr
273  *   Interface request structure output buffer.
274  *
275  * @return
276  *   0 on success, a negative errno value otherwise and rte_errno is set.
277  */
278 int
279 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
280 {
281 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
282 	int ret = 0;
283 
284 	if (sock == -1) {
285 		rte_errno = errno;
286 		return -rte_errno;
287 	}
288 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
289 	if (ret)
290 		goto error;
291 	ret = ioctl(sock, req, ifr);
292 	if (ret == -1) {
293 		rte_errno = errno;
294 		goto error;
295 	}
296 	close(sock);
297 	return 0;
298 error:
299 	close(sock);
300 	return -rte_errno;
301 }
302 
303 /**
304  * Get device MTU.
305  *
306  * @param dev
307  *   Pointer to Ethernet device.
308  * @param[out] mtu
309  *   MTU value output buffer.
310  *
311  * @return
312  *   0 on success, a negative errno value otherwise and rte_errno is set.
313  */
314 int
315 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
316 {
317 	struct ifreq request;
318 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
319 
320 	if (ret)
321 		return ret;
322 	*mtu = request.ifr_mtu;
323 	return 0;
324 }
325 
326 /**
327  * Set device MTU.
328  *
329  * @param dev
330  *   Pointer to Ethernet device.
331  * @param mtu
332  *   MTU value to set.
333  *
334  * @return
335  *   0 on success, a negative errno value otherwise and rte_errno is set.
336  */
337 static int
338 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
339 {
340 	struct ifreq request = { .ifr_mtu = mtu, };
341 
342 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
343 }
344 
345 /**
346  * Set device flags.
347  *
348  * @param dev
349  *   Pointer to Ethernet device.
350  * @param keep
351  *   Bitmask for flags that must remain untouched.
352  * @param flags
353  *   Bitmask for flags to modify.
354  *
355  * @return
356  *   0 on success, a negative errno value otherwise and rte_errno is set.
357  */
358 int
359 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
360 {
361 	struct ifreq request;
362 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
363 
364 	if (ret)
365 		return ret;
366 	request.ifr_flags &= keep;
367 	request.ifr_flags |= flags & ~keep;
368 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
369 }
370 
371 /**
372  * DPDK callback for Ethernet device configuration.
373  *
374  * @param dev
375  *   Pointer to Ethernet device structure.
376  *
377  * @return
378  *   0 on success, a negative errno value otherwise and rte_errno is set.
379  */
380 int
381 mlx5_dev_configure(struct rte_eth_dev *dev)
382 {
383 	struct mlx5_priv *priv = dev->data->dev_private;
384 	unsigned int rxqs_n = dev->data->nb_rx_queues;
385 	unsigned int txqs_n = dev->data->nb_tx_queues;
386 	unsigned int i;
387 	unsigned int j;
388 	unsigned int reta_idx_n;
389 	const uint8_t use_app_rss_key =
390 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
391 	int ret = 0;
392 
393 	if (use_app_rss_key &&
394 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
395 	     MLX5_RSS_HASH_KEY_LEN)) {
396 		DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
397 			dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
398 		rte_errno = EINVAL;
399 		return -rte_errno;
400 	}
401 	priv->rss_conf.rss_key =
402 		rte_realloc(priv->rss_conf.rss_key,
403 			    MLX5_RSS_HASH_KEY_LEN, 0);
404 	if (!priv->rss_conf.rss_key) {
405 		DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
406 			dev->data->port_id, rxqs_n);
407 		rte_errno = ENOMEM;
408 		return -rte_errno;
409 	}
410 	memcpy(priv->rss_conf.rss_key,
411 	       use_app_rss_key ?
412 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
413 	       rss_hash_default_key,
414 	       MLX5_RSS_HASH_KEY_LEN);
415 	priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
416 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
417 	priv->rxqs = (void *)dev->data->rx_queues;
418 	priv->txqs = (void *)dev->data->tx_queues;
419 	if (txqs_n != priv->txqs_n) {
420 		DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
421 			dev->data->port_id, priv->txqs_n, txqs_n);
422 		priv->txqs_n = txqs_n;
423 	}
424 	if (rxqs_n > priv->config.ind_table_max_size) {
425 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
426 			dev->data->port_id, rxqs_n);
427 		rte_errno = EINVAL;
428 		return -rte_errno;
429 	}
430 	if (rxqs_n != priv->rxqs_n) {
431 		DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
432 			dev->data->port_id, priv->rxqs_n, rxqs_n);
433 		priv->rxqs_n = rxqs_n;
434 		/*
435 		 * If the requested number of RX queues is not a power of two,
436 		 * use the maximum indirection table size for better balancing.
437 		 * The result is always rounded to the next power of two.
438 		 */
439 		reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
440 					     priv->config.ind_table_max_size :
441 					     rxqs_n));
442 		ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
443 		if (ret)
444 			return ret;
445 		/*
446 		 * When the number of RX queues is not a power of two,
447 		 * the remaining table entries are padded with reused WQs
448 		 * and hashes are not spread uniformly.
449 		 */
450 		for (i = 0, j = 0; (i != reta_idx_n); ++i) {
451 			(*priv->reta_idx)[i] = j;
452 			if (++j == rxqs_n)
453 				j = 0;
454 		}
455 	}
456 	ret = mlx5_proc_priv_init(dev);
457 	if (ret)
458 		return ret;
459 	return 0;
460 }
461 
462 /**
463  * Sets default tuning parameters.
464  *
465  * @param dev
466  *   Pointer to Ethernet device.
467  * @param[out] info
468  *   Info structure output buffer.
469  */
470 static void
471 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
472 {
473 	struct mlx5_priv *priv = dev->data->dev_private;
474 
475 	/* Minimum CPU utilization. */
476 	info->default_rxportconf.ring_size = 256;
477 	info->default_txportconf.ring_size = 256;
478 	info->default_rxportconf.burst_size = 64;
479 	info->default_txportconf.burst_size = 64;
480 	if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
481 		info->default_rxportconf.nb_queues = 16;
482 		info->default_txportconf.nb_queues = 16;
483 		if (dev->data->nb_rx_queues > 2 ||
484 		    dev->data->nb_tx_queues > 2) {
485 			/* Max Throughput. */
486 			info->default_rxportconf.ring_size = 2048;
487 			info->default_txportconf.ring_size = 2048;
488 		}
489 	} else {
490 		info->default_rxportconf.nb_queues = 8;
491 		info->default_txportconf.nb_queues = 8;
492 		if (dev->data->nb_rx_queues > 2 ||
493 		    dev->data->nb_tx_queues > 2) {
494 			/* Max Throughput. */
495 			info->default_rxportconf.ring_size = 4096;
496 			info->default_txportconf.ring_size = 4096;
497 		}
498 	}
499 }
500 
501 /**
502  * Sets tx mbuf limiting parameters.
503  *
504  * @param dev
505  *   Pointer to Ethernet device.
506  * @param[out] info
507  *   Info structure output buffer.
508  */
509 static void
510 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
511 {
512 	struct mlx5_priv *priv = dev->data->dev_private;
513 	struct mlx5_dev_config *config = &priv->config;
514 	unsigned int inlen;
515 	uint16_t nb_max;
516 
517 	inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
518 		MLX5_SEND_DEF_INLINE_LEN :
519 		(unsigned int)config->txq_inline_max;
520 	assert(config->txq_inline_min >= 0);
521 	inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
522 	inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
523 			       MLX5_ESEG_MIN_INLINE_SIZE -
524 			       MLX5_WQE_CSEG_SIZE -
525 			       MLX5_WQE_ESEG_SIZE -
526 			       MLX5_WQE_DSEG_SIZE * 2);
527 	nb_max = (MLX5_WQE_SIZE_MAX +
528 		  MLX5_ESEG_MIN_INLINE_SIZE -
529 		  MLX5_WQE_CSEG_SIZE -
530 		  MLX5_WQE_ESEG_SIZE -
531 		  MLX5_WQE_DSEG_SIZE -
532 		  inlen) / MLX5_WSEG_SIZE;
533 	info->tx_desc_lim.nb_seg_max = nb_max;
534 	info->tx_desc_lim.nb_mtu_seg_max = nb_max;
535 }
536 
537 /**
538  * DPDK callback to get information about the device.
539  *
540  * @param dev
541  *   Pointer to Ethernet device structure.
542  * @param[out] info
543  *   Info structure output buffer.
544  */
545 void
546 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
547 {
548 	struct mlx5_priv *priv = dev->data->dev_private;
549 	struct mlx5_dev_config *config = &priv->config;
550 	unsigned int max;
551 
552 	/* FIXME: we should ask the device for these values. */
553 	info->min_rx_bufsize = 32;
554 	info->max_rx_pktlen = 65536;
555 	/*
556 	 * Since we need one CQ per QP, the limit is the minimum number
557 	 * between the two values.
558 	 */
559 	max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
560 		      priv->sh->device_attr.orig_attr.max_qp);
561 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
562 	if (max >= 65535)
563 		max = 65535;
564 	info->max_rx_queues = max;
565 	info->max_tx_queues = max;
566 	info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
567 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
568 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
569 				 info->rx_queue_offload_capa);
570 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
571 	info->if_index = mlx5_ifindex(dev);
572 	info->reta_size = priv->reta_idx_n ?
573 		priv->reta_idx_n : config->ind_table_max_size;
574 	info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
575 	info->speed_capa = priv->link_speed_capa;
576 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
577 	mlx5_set_default_params(dev, info);
578 	mlx5_set_txlimit_params(dev, info);
579 	info->switch_info.name = dev->data->name;
580 	info->switch_info.domain_id = priv->domain_id;
581 	info->switch_info.port_id = priv->representor_id;
582 	if (priv->representor) {
583 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
584 		uint16_t port_id[i];
585 
586 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
587 		while (i--) {
588 			struct mlx5_priv *opriv =
589 				rte_eth_devices[port_id[i]].data->dev_private;
590 
591 			if (!opriv ||
592 			    opriv->representor ||
593 			    opriv->domain_id != priv->domain_id)
594 				continue;
595 			/*
596 			 * Override switch name with that of the master
597 			 * device.
598 			 */
599 			info->switch_info.name = opriv->dev_data->name;
600 			break;
601 		}
602 	}
603 }
604 
605 /**
606  * Get device current raw clock counter
607  *
608  * @param dev
609  *   Pointer to Ethernet device structure.
610  * @param[out] time
611  *   Current raw clock counter of the device.
612  *
613  * @return
614  *   0 if the clock has correctly been read
615  *   The value of errno in case of error
616  */
617 int
618 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
619 {
620 	struct mlx5_priv *priv = dev->data->dev_private;
621 	struct ibv_context *ctx = priv->sh->ctx;
622 	struct ibv_values_ex values;
623 	int err = 0;
624 
625 	values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
626 	err = mlx5_glue->query_rt_values_ex(ctx, &values);
627 	if (err != 0) {
628 		DRV_LOG(WARNING, "Could not query the clock !");
629 		return err;
630 	}
631 	*clock = values.raw_clock.tv_nsec;
632 	return 0;
633 }
634 
635 /**
636  * Get firmware version of a device.
637  *
638  * @param dev
639  *   Ethernet device port.
640  * @param fw_ver
641  *   String output allocated by caller.
642  * @param fw_size
643  *   Size of the output string, including terminating null byte.
644  *
645  * @return
646  *   0 on success, or the size of the non truncated string if too big.
647  */
648 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
649 {
650 	struct mlx5_priv *priv = dev->data->dev_private;
651 	struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
652 	size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
653 
654 	if (fw_size < size)
655 		return size;
656 	if (fw_ver != NULL)
657 		strlcpy(fw_ver, attr->fw_ver, fw_size);
658 	return 0;
659 }
660 
661 /**
662  * Get supported packet types.
663  *
664  * @param dev
665  *   Pointer to Ethernet device structure.
666  *
667  * @return
668  *   A pointer to the supported Packet types array.
669  */
670 const uint32_t *
671 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
672 {
673 	static const uint32_t ptypes[] = {
674 		/* refers to rxq_cq_to_pkt_type() */
675 		RTE_PTYPE_L2_ETHER,
676 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
677 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
678 		RTE_PTYPE_L4_NONFRAG,
679 		RTE_PTYPE_L4_FRAG,
680 		RTE_PTYPE_L4_TCP,
681 		RTE_PTYPE_L4_UDP,
682 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
683 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
684 		RTE_PTYPE_INNER_L4_NONFRAG,
685 		RTE_PTYPE_INNER_L4_FRAG,
686 		RTE_PTYPE_INNER_L4_TCP,
687 		RTE_PTYPE_INNER_L4_UDP,
688 		RTE_PTYPE_UNKNOWN
689 	};
690 
691 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
692 	    dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
693 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
694 		return ptypes;
695 	return NULL;
696 }
697 
698 /**
699  * Retrieve the master device for representor in the same switch domain.
700  *
701  * @param dev
702  *   Pointer to representor Ethernet device structure.
703  *
704  * @return
705  *   Master device structure  on success, NULL otherwise.
706  */
707 
708 static struct rte_eth_dev *
709 mlx5_find_master_dev(struct rte_eth_dev *dev)
710 {
711 	struct mlx5_priv *priv;
712 	uint16_t port_id;
713 	uint16_t domain_id;
714 
715 	priv = dev->data->dev_private;
716 	domain_id = priv->domain_id;
717 	assert(priv->representor);
718 	RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
719 		priv = rte_eth_devices[port_id].data->dev_private;
720 		if (priv &&
721 		    priv->master &&
722 		    priv->domain_id == domain_id)
723 			return &rte_eth_devices[port_id];
724 	}
725 	return NULL;
726 }
727 
728 /**
729  * DPDK callback to retrieve physical link information.
730  *
731  * @param dev
732  *   Pointer to Ethernet device structure.
733  * @param[out] link
734  *   Storage for current link status.
735  *
736  * @return
737  *   0 on success, a negative errno value otherwise and rte_errno is set.
738  */
739 static int
740 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
741 			       struct rte_eth_link *link)
742 {
743 	struct mlx5_priv *priv = dev->data->dev_private;
744 	struct ethtool_cmd edata = {
745 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
746 	};
747 	struct ifreq ifr;
748 	struct rte_eth_link dev_link;
749 	int link_speed = 0;
750 	int ret;
751 
752 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
753 	if (ret) {
754 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
755 			dev->data->port_id, strerror(rte_errno));
756 		return ret;
757 	}
758 	dev_link = (struct rte_eth_link) {
759 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
760 				(ifr.ifr_flags & IFF_RUNNING)),
761 	};
762 	ifr = (struct ifreq) {
763 		.ifr_data = (void *)&edata,
764 	};
765 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
766 	if (ret) {
767 		if (ret == -ENOTSUP && priv->representor) {
768 			struct rte_eth_dev *master;
769 
770 			/*
771 			 * For representors we can try to inherit link
772 			 * settings from the master device. Actually
773 			 * link settings do not make a lot of sense
774 			 * for representors due to missing physical
775 			 * link. The old kernel drivers supported
776 			 * emulated settings query for representors,
777 			 * the new ones do not, so we have to add
778 			 * this code for compatibility issues.
779 			 */
780 			master = mlx5_find_master_dev(dev);
781 			if (master) {
782 				ifr = (struct ifreq) {
783 					.ifr_data = (void *)&edata,
784 				};
785 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
786 			}
787 		}
788 		if (ret) {
789 			DRV_LOG(WARNING,
790 				"port %u ioctl(SIOCETHTOOL,"
791 				" ETHTOOL_GSET) failed: %s",
792 				dev->data->port_id, strerror(rte_errno));
793 			return ret;
794 		}
795 	}
796 	link_speed = ethtool_cmd_speed(&edata);
797 	if (link_speed == -1)
798 		dev_link.link_speed = ETH_SPEED_NUM_NONE;
799 	else
800 		dev_link.link_speed = link_speed;
801 	priv->link_speed_capa = 0;
802 	if (edata.supported & SUPPORTED_Autoneg)
803 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
804 	if (edata.supported & (SUPPORTED_1000baseT_Full |
805 			       SUPPORTED_1000baseKX_Full))
806 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
807 	if (edata.supported & SUPPORTED_10000baseKR_Full)
808 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
809 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
810 			       SUPPORTED_40000baseCR4_Full |
811 			       SUPPORTED_40000baseSR4_Full |
812 			       SUPPORTED_40000baseLR4_Full))
813 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
814 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
815 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
816 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
817 			ETH_LINK_SPEED_FIXED);
818 	if (((dev_link.link_speed && !dev_link.link_status) ||
819 	     (!dev_link.link_speed && dev_link.link_status))) {
820 		rte_errno = EAGAIN;
821 		return -rte_errno;
822 	}
823 	*link = dev_link;
824 	return 0;
825 }
826 
827 /**
828  * Retrieve physical link information (unlocked version using new ioctl).
829  *
830  * @param dev
831  *   Pointer to Ethernet device structure.
832  * @param[out] link
833  *   Storage for current link status.
834  *
835  * @return
836  *   0 on success, a negative errno value otherwise and rte_errno is set.
837  */
838 static int
839 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
840 			     struct rte_eth_link *link)
841 
842 {
843 	struct mlx5_priv *priv = dev->data->dev_private;
844 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
845 	struct ifreq ifr;
846 	struct rte_eth_link dev_link;
847 	struct rte_eth_dev *master = NULL;
848 	uint64_t sc;
849 	int ret;
850 
851 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
852 	if (ret) {
853 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
854 			dev->data->port_id, strerror(rte_errno));
855 		return ret;
856 	}
857 	dev_link = (struct rte_eth_link) {
858 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
859 				(ifr.ifr_flags & IFF_RUNNING)),
860 	};
861 	ifr = (struct ifreq) {
862 		.ifr_data = (void *)&gcmd,
863 	};
864 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
865 	if (ret) {
866 		if (ret == -ENOTSUP && priv->representor) {
867 			/*
868 			 * For representors we can try to inherit link
869 			 * settings from the master device. Actually
870 			 * link settings do not make a lot of sense
871 			 * for representors due to missing physical
872 			 * link. The old kernel drivers supported
873 			 * emulated settings query for representors,
874 			 * the new ones do not, so we have to add
875 			 * this code for compatibility issues.
876 			 */
877 			master = mlx5_find_master_dev(dev);
878 			if (master) {
879 				ifr = (struct ifreq) {
880 					.ifr_data = (void *)&gcmd,
881 				};
882 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
883 			}
884 		}
885 		if (ret) {
886 			DRV_LOG(DEBUG,
887 				"port %u ioctl(SIOCETHTOOL,"
888 				" ETHTOOL_GLINKSETTINGS) failed: %s",
889 				dev->data->port_id, strerror(rte_errno));
890 			return ret;
891 		}
892 
893 	}
894 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
895 
896 	alignas(struct ethtool_link_settings)
897 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
898 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
899 	struct ethtool_link_settings *ecmd = (void *)data;
900 
901 	*ecmd = gcmd;
902 	ifr.ifr_data = (void *)ecmd;
903 	ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
904 	if (ret) {
905 		DRV_LOG(DEBUG,
906 			"port %u ioctl(SIOCETHTOOL,"
907 			"ETHTOOL_GLINKSETTINGS) failed: %s",
908 			dev->data->port_id, strerror(rte_errno));
909 		return ret;
910 	}
911 	dev_link.link_speed = ecmd->speed;
912 	sc = ecmd->link_mode_masks[0] |
913 		((uint64_t)ecmd->link_mode_masks[1] << 32);
914 	priv->link_speed_capa = 0;
915 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
916 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
917 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
918 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
919 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
920 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
921 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
922 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
923 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
924 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
925 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
926 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
927 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
928 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
929 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
930 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
931 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
932 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
933 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
934 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
935 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
936 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
937 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
938 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
939 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
940 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
941 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
942 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
943 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
944 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
945 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
946 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
947 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
948 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
949 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
950 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
951 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
952 				  ETH_LINK_SPEED_FIXED);
953 	if (((dev_link.link_speed && !dev_link.link_status) ||
954 	     (!dev_link.link_speed && dev_link.link_status))) {
955 		rte_errno = EAGAIN;
956 		return -rte_errno;
957 	}
958 	*link = dev_link;
959 	return 0;
960 }
961 
962 /**
963  * DPDK callback to retrieve physical link information.
964  *
965  * @param dev
966  *   Pointer to Ethernet device structure.
967  * @param wait_to_complete
968  *   Wait for request completion.
969  *
970  * @return
971  *   0 if link status was not updated, positive if it was, a negative errno
972  *   value otherwise and rte_errno is set.
973  */
974 int
975 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
976 {
977 	int ret;
978 	struct rte_eth_link dev_link;
979 	time_t start_time = time(NULL);
980 
981 	do {
982 		ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
983 		if (ret == -ENOTSUP)
984 			ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
985 		if (ret == 0)
986 			break;
987 		/* Handle wait to complete situation. */
988 		if (wait_to_complete && ret == -EAGAIN) {
989 			if (abs((int)difftime(time(NULL), start_time)) <
990 			    MLX5_LINK_STATUS_TIMEOUT) {
991 				usleep(0);
992 				continue;
993 			} else {
994 				rte_errno = EBUSY;
995 				return -rte_errno;
996 			}
997 		} else if (ret < 0) {
998 			return ret;
999 		}
1000 	} while (wait_to_complete);
1001 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
1002 		       sizeof(struct rte_eth_link));
1003 	dev->data->dev_link = dev_link;
1004 	return ret;
1005 }
1006 
1007 /**
1008  * DPDK callback to change the MTU.
1009  *
1010  * @param dev
1011  *   Pointer to Ethernet device structure.
1012  * @param in_mtu
1013  *   New MTU.
1014  *
1015  * @return
1016  *   0 on success, a negative errno value otherwise and rte_errno is set.
1017  */
1018 int
1019 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1020 {
1021 	struct mlx5_priv *priv = dev->data->dev_private;
1022 	uint16_t kern_mtu = 0;
1023 	int ret;
1024 
1025 	ret = mlx5_get_mtu(dev, &kern_mtu);
1026 	if (ret)
1027 		return ret;
1028 	/* Set kernel interface MTU first. */
1029 	ret = mlx5_set_mtu(dev, mtu);
1030 	if (ret)
1031 		return ret;
1032 	ret = mlx5_get_mtu(dev, &kern_mtu);
1033 	if (ret)
1034 		return ret;
1035 	if (kern_mtu == mtu) {
1036 		priv->mtu = mtu;
1037 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1038 			dev->data->port_id, mtu);
1039 		return 0;
1040 	}
1041 	rte_errno = EAGAIN;
1042 	return -rte_errno;
1043 }
1044 
1045 /**
1046  * DPDK callback to get flow control status.
1047  *
1048  * @param dev
1049  *   Pointer to Ethernet device structure.
1050  * @param[out] fc_conf
1051  *   Flow control output buffer.
1052  *
1053  * @return
1054  *   0 on success, a negative errno value otherwise and rte_errno is set.
1055  */
1056 int
1057 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1058 {
1059 	struct ifreq ifr;
1060 	struct ethtool_pauseparam ethpause = {
1061 		.cmd = ETHTOOL_GPAUSEPARAM
1062 	};
1063 	int ret;
1064 
1065 	ifr.ifr_data = (void *)&ethpause;
1066 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1067 	if (ret) {
1068 		DRV_LOG(WARNING,
1069 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1070 			" %s",
1071 			dev->data->port_id, strerror(rte_errno));
1072 		return ret;
1073 	}
1074 	fc_conf->autoneg = ethpause.autoneg;
1075 	if (ethpause.rx_pause && ethpause.tx_pause)
1076 		fc_conf->mode = RTE_FC_FULL;
1077 	else if (ethpause.rx_pause)
1078 		fc_conf->mode = RTE_FC_RX_PAUSE;
1079 	else if (ethpause.tx_pause)
1080 		fc_conf->mode = RTE_FC_TX_PAUSE;
1081 	else
1082 		fc_conf->mode = RTE_FC_NONE;
1083 	return 0;
1084 }
1085 
1086 /**
1087  * DPDK callback to modify flow control parameters.
1088  *
1089  * @param dev
1090  *   Pointer to Ethernet device structure.
1091  * @param[in] fc_conf
1092  *   Flow control parameters.
1093  *
1094  * @return
1095  *   0 on success, a negative errno value otherwise and rte_errno is set.
1096  */
1097 int
1098 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1099 {
1100 	struct ifreq ifr;
1101 	struct ethtool_pauseparam ethpause = {
1102 		.cmd = ETHTOOL_SPAUSEPARAM
1103 	};
1104 	int ret;
1105 
1106 	ifr.ifr_data = (void *)&ethpause;
1107 	ethpause.autoneg = fc_conf->autoneg;
1108 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1109 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1110 		ethpause.rx_pause = 1;
1111 	else
1112 		ethpause.rx_pause = 0;
1113 
1114 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1115 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1116 		ethpause.tx_pause = 1;
1117 	else
1118 		ethpause.tx_pause = 0;
1119 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1120 	if (ret) {
1121 		DRV_LOG(WARNING,
1122 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1123 			" failed: %s",
1124 			dev->data->port_id, strerror(rte_errno));
1125 		return ret;
1126 	}
1127 	return 0;
1128 }
1129 
1130 /**
1131  * Get PCI information from struct ibv_device.
1132  *
1133  * @param device
1134  *   Pointer to Ethernet device structure.
1135  * @param[out] pci_addr
1136  *   PCI bus address output buffer.
1137  *
1138  * @return
1139  *   0 on success, a negative errno value otherwise and rte_errno is set.
1140  */
1141 int
1142 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1143 			    struct rte_pci_addr *pci_addr)
1144 {
1145 	FILE *file;
1146 	char line[32];
1147 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
1148 
1149 	file = fopen(path, "rb");
1150 	if (file == NULL) {
1151 		rte_errno = errno;
1152 		return -rte_errno;
1153 	}
1154 	while (fgets(line, sizeof(line), file) == line) {
1155 		size_t len = strlen(line);
1156 		int ret;
1157 
1158 		/* Truncate long lines. */
1159 		if (len == (sizeof(line) - 1))
1160 			while (line[(len - 1)] != '\n') {
1161 				ret = fgetc(file);
1162 				if (ret == EOF)
1163 					break;
1164 				line[(len - 1)] = ret;
1165 			}
1166 		/* Extract information. */
1167 		if (sscanf(line,
1168 			   "PCI_SLOT_NAME="
1169 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1170 			   &pci_addr->domain,
1171 			   &pci_addr->bus,
1172 			   &pci_addr->devid,
1173 			   &pci_addr->function) == 4) {
1174 			ret = 0;
1175 			break;
1176 		}
1177 	}
1178 	fclose(file);
1179 	return 0;
1180 }
1181 
1182 /**
1183  * Handle asynchronous removal event for entire multiport device.
1184  *
1185  * @param sh
1186  *   Infiniband device shared context.
1187  */
1188 static void
1189 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1190 {
1191 	uint32_t i;
1192 
1193 	for (i = 0; i < sh->max_port; ++i) {
1194 		struct rte_eth_dev *dev;
1195 
1196 		if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1197 			/*
1198 			 * Or not existing port either no
1199 			 * handler installed for this port.
1200 			 */
1201 			continue;
1202 		}
1203 		dev = &rte_eth_devices[sh->port[i].ih_port_id];
1204 		assert(dev);
1205 		if (dev->data->dev_conf.intr_conf.rmv)
1206 			_rte_eth_dev_callback_process
1207 				(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1208 	}
1209 }
1210 
1211 /**
1212  * Handle shared asynchronous events the NIC (removal event
1213  * and link status change). Supports multiport IB device.
1214  *
1215  * @param cb_arg
1216  *   Callback argument.
1217  */
1218 void
1219 mlx5_dev_interrupt_handler(void *cb_arg)
1220 {
1221 	struct mlx5_ibv_shared *sh = cb_arg;
1222 	struct ibv_async_event event;
1223 
1224 	/* Read all message from the IB device and acknowledge them. */
1225 	for (;;) {
1226 		struct rte_eth_dev *dev;
1227 		uint32_t tmp;
1228 
1229 		if (mlx5_glue->get_async_event(sh->ctx, &event))
1230 			break;
1231 		/* Retrieve and check IB port index. */
1232 		tmp = (uint32_t)event.element.port_num;
1233 		if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1234 			/*
1235 			 * The DEVICE_FATAL event is called once for
1236 			 * entire device without port specifying.
1237 			 * We should notify all existing ports.
1238 			 */
1239 			mlx5_glue->ack_async_event(&event);
1240 			mlx5_dev_interrupt_device_fatal(sh);
1241 			continue;
1242 		}
1243 		assert(tmp && (tmp <= sh->max_port));
1244 		if (!tmp) {
1245 			/* Unsupported devive level event. */
1246 			mlx5_glue->ack_async_event(&event);
1247 			DRV_LOG(DEBUG,
1248 				"unsupported common event (type %d)",
1249 				event.event_type);
1250 			continue;
1251 		}
1252 		if (tmp > sh->max_port) {
1253 			/* Invalid IB port index. */
1254 			mlx5_glue->ack_async_event(&event);
1255 			DRV_LOG(DEBUG,
1256 				"cannot handle an event (type %d)"
1257 				"due to invalid IB port index (%u)",
1258 				event.event_type, tmp);
1259 			continue;
1260 		}
1261 		if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1262 			/* No handler installed. */
1263 			mlx5_glue->ack_async_event(&event);
1264 			DRV_LOG(DEBUG,
1265 				"cannot handle an event (type %d)"
1266 				"due to no handler installed for port %u",
1267 				event.event_type, tmp);
1268 			continue;
1269 		}
1270 		/* Retrieve ethernet device descriptor. */
1271 		tmp = sh->port[tmp - 1].ih_port_id;
1272 		dev = &rte_eth_devices[tmp];
1273 		assert(dev);
1274 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1275 		     event.event_type == IBV_EVENT_PORT_ERR) &&
1276 			dev->data->dev_conf.intr_conf.lsc) {
1277 			mlx5_glue->ack_async_event(&event);
1278 			if (mlx5_link_update(dev, 0) == -EAGAIN) {
1279 				usleep(0);
1280 				continue;
1281 			}
1282 			_rte_eth_dev_callback_process
1283 				(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1284 			continue;
1285 		}
1286 		DRV_LOG(DEBUG,
1287 			"port %u cannot handle an unknown event (type %d)",
1288 			dev->data->port_id, event.event_type);
1289 		mlx5_glue->ack_async_event(&event);
1290 	}
1291 }
1292 
1293 /*
1294  * Unregister callback handler safely. The handler may be active
1295  * while we are trying to unregister it, in this case code -EAGAIN
1296  * is returned by rte_intr_callback_unregister(). This routine checks
1297  * the return code and tries to unregister handler again.
1298  *
1299  * @param handle
1300  *   interrupt handle
1301  * @param cb_fn
1302  *   pointer to callback routine
1303  * @cb_arg
1304  *   opaque callback parameter
1305  */
1306 void
1307 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1308 			      rte_intr_callback_fn cb_fn, void *cb_arg)
1309 {
1310 	/*
1311 	 * Try to reduce timeout management overhead by not calling
1312 	 * the timer related routines on the first iteration. If the
1313 	 * unregistering succeeds on first call there will be no
1314 	 * timer calls at all.
1315 	 */
1316 	uint64_t twait = 0;
1317 	uint64_t start = 0;
1318 
1319 	do {
1320 		int ret;
1321 
1322 		ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1323 		if (ret >= 0)
1324 			return;
1325 		if (ret != -EAGAIN) {
1326 			DRV_LOG(INFO, "failed to unregister interrupt"
1327 				      " handler (error: %d)", ret);
1328 			assert(false);
1329 			return;
1330 		}
1331 		if (twait) {
1332 			struct timespec onems;
1333 
1334 			/* Wait one millisecond and try again. */
1335 			onems.tv_sec = 0;
1336 			onems.tv_nsec = NS_PER_S / MS_PER_S;
1337 			nanosleep(&onems, 0);
1338 			/* Check whether one second elapsed. */
1339 			if ((rte_get_timer_cycles() - start) <= twait)
1340 				continue;
1341 		} else {
1342 			/*
1343 			 * We get the amount of timer ticks for one second.
1344 			 * If this amount elapsed it means we spent one
1345 			 * second in waiting. This branch is executed once
1346 			 * on first iteration.
1347 			 */
1348 			twait = rte_get_timer_hz();
1349 			assert(twait);
1350 		}
1351 		/*
1352 		 * Timeout elapsed, show message (once a second) and retry.
1353 		 * We have no other acceptable option here, if we ignore
1354 		 * the unregistering return code the handler will not
1355 		 * be unregistered, fd will be closed and we may get the
1356 		 * crush. Hanging and messaging in the loop seems not to be
1357 		 * the worst choice.
1358 		 */
1359 		DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1360 		start = rte_get_timer_cycles();
1361 	} while (true);
1362 }
1363 
1364 /**
1365  * Handle DEVX interrupts from the NIC.
1366  * This function is probably called from the DPDK host thread.
1367  *
1368  * @param cb_arg
1369  *   Callback argument.
1370  */
1371 void
1372 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1373 {
1374 #ifndef HAVE_IBV_DEVX_ASYNC
1375 	(void)cb_arg;
1376 	return;
1377 #else
1378 	struct mlx5_ibv_shared *sh = cb_arg;
1379 	union {
1380 		struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1381 		uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1382 			    MLX5_ST_SZ_BYTES(traffic_counter) +
1383 			    sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1384 	} out;
1385 	uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1386 
1387 	while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1388 						   &out.cmd_resp,
1389 						   sizeof(out.buf)))
1390 		mlx5_flow_async_pool_query_handle
1391 			(sh, (uint64_t)out.cmd_resp.wr_id,
1392 			 mlx5_devx_get_out_command_status(buf));
1393 #endif /* HAVE_IBV_DEVX_ASYNC */
1394 }
1395 
1396 /**
1397  * Uninstall shared asynchronous device events handler.
1398  * This function is implemented to support event sharing
1399  * between multiple ports of single IB device.
1400  *
1401  * @param dev
1402  *   Pointer to Ethernet device.
1403  */
1404 static void
1405 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1406 {
1407 	struct mlx5_priv *priv = dev->data->dev_private;
1408 	struct mlx5_ibv_shared *sh = priv->sh;
1409 
1410 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1411 		return;
1412 	pthread_mutex_lock(&sh->intr_mutex);
1413 	assert(priv->ibv_port);
1414 	assert(priv->ibv_port <= sh->max_port);
1415 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1416 	if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1417 		goto exit;
1418 	assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1419 					(uint32_t)dev->data->port_id);
1420 	assert(sh->intr_cnt);
1421 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1422 	if (!sh->intr_cnt || --sh->intr_cnt)
1423 		goto exit;
1424 	mlx5_intr_callback_unregister(&sh->intr_handle,
1425 				     mlx5_dev_interrupt_handler, sh);
1426 	sh->intr_handle.fd = 0;
1427 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1428 	if (sh->intr_handle_devx.fd) {
1429 		rte_intr_callback_unregister(&sh->intr_handle_devx,
1430 					     mlx5_dev_interrupt_handler_devx,
1431 					     sh);
1432 		sh->intr_handle_devx.fd = 0;
1433 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1434 	}
1435 	if (sh->devx_comp) {
1436 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1437 		sh->devx_comp = NULL;
1438 	}
1439 exit:
1440 	pthread_mutex_unlock(&sh->intr_mutex);
1441 }
1442 
1443 /**
1444  * Install shared asynchronous device events handler.
1445  * This function is implemented to support event sharing
1446  * between multiple ports of single IB device.
1447  *
1448  * @param dev
1449  *   Pointer to Ethernet device.
1450  */
1451 static void
1452 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1453 {
1454 	struct mlx5_priv *priv = dev->data->dev_private;
1455 	struct mlx5_ibv_shared *sh = priv->sh;
1456 	int ret;
1457 	int flags;
1458 
1459 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1460 		return;
1461 	pthread_mutex_lock(&sh->intr_mutex);
1462 	assert(priv->ibv_port);
1463 	assert(priv->ibv_port <= sh->max_port);
1464 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1465 	if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1466 		/* The handler is already installed for this port. */
1467 		assert(sh->intr_cnt);
1468 		goto exit;
1469 	}
1470 	sh->port[priv->ibv_port - 1].ih_port_id = (uint32_t)dev->data->port_id;
1471 	if (sh->intr_cnt) {
1472 		sh->intr_cnt++;
1473 		goto exit;
1474 	}
1475 	/* No shared handler installed. */
1476 	assert(sh->ctx->async_fd > 0);
1477 	flags = fcntl(sh->ctx->async_fd, F_GETFL);
1478 	ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1479 	if (ret) {
1480 		DRV_LOG(INFO, "failed to change file descriptor"
1481 			      " async event queue");
1482 		goto error;
1483 	}
1484 	sh->intr_handle.fd = sh->ctx->async_fd;
1485 	sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1486 	rte_intr_callback_register(&sh->intr_handle,
1487 				   mlx5_dev_interrupt_handler, sh);
1488 	if (priv->config.devx) {
1489 #ifndef HAVE_IBV_DEVX_ASYNC
1490 		goto error_unregister;
1491 #else
1492 		sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1493 		if (sh->devx_comp) {
1494 			flags = fcntl(sh->devx_comp->fd, F_GETFL);
1495 			ret = fcntl(sh->devx_comp->fd, F_SETFL,
1496 				    flags | O_NONBLOCK);
1497 			if (ret) {
1498 				DRV_LOG(INFO, "failed to change file descriptor"
1499 					      " devx async event queue");
1500 				goto error_unregister;
1501 			}
1502 			sh->intr_handle_devx.fd = sh->devx_comp->fd;
1503 			sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1504 			rte_intr_callback_register
1505 				(&sh->intr_handle_devx,
1506 				 mlx5_dev_interrupt_handler_devx, sh);
1507 		} else {
1508 			DRV_LOG(INFO, "failed to create devx async command "
1509 				"completion");
1510 			goto error_unregister;
1511 		}
1512 #endif /* HAVE_IBV_DEVX_ASYNC */
1513 	}
1514 	sh->intr_cnt++;
1515 	goto exit;
1516 error_unregister:
1517 	rte_intr_callback_unregister(&sh->intr_handle,
1518 				     mlx5_dev_interrupt_handler, sh);
1519 error:
1520 	/* Indicate there will be no interrupts. */
1521 	dev->data->dev_conf.intr_conf.lsc = 0;
1522 	dev->data->dev_conf.intr_conf.rmv = 0;
1523 	sh->intr_handle.fd = 0;
1524 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1525 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1526 exit:
1527 	pthread_mutex_unlock(&sh->intr_mutex);
1528 }
1529 
1530 /**
1531  * Uninstall interrupt handler.
1532  *
1533  * @param dev
1534  *   Pointer to Ethernet device.
1535  */
1536 void
1537 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1538 {
1539 	mlx5_dev_shared_handler_uninstall(dev);
1540 }
1541 
1542 /**
1543  * Install interrupt handler.
1544  *
1545  * @param dev
1546  *   Pointer to Ethernet device.
1547  */
1548 void
1549 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1550 {
1551 	mlx5_dev_shared_handler_install(dev);
1552 }
1553 
1554 /**
1555  * DPDK callback to bring the link DOWN.
1556  *
1557  * @param dev
1558  *   Pointer to Ethernet device structure.
1559  *
1560  * @return
1561  *   0 on success, a negative errno value otherwise and rte_errno is set.
1562  */
1563 int
1564 mlx5_set_link_down(struct rte_eth_dev *dev)
1565 {
1566 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1567 }
1568 
1569 /**
1570  * DPDK callback to bring the link UP.
1571  *
1572  * @param dev
1573  *   Pointer to Ethernet device structure.
1574  *
1575  * @return
1576  *   0 on success, a negative errno value otherwise and rte_errno is set.
1577  */
1578 int
1579 mlx5_set_link_up(struct rte_eth_dev *dev)
1580 {
1581 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1582 }
1583 
1584 /**
1585  * Configure the RX function to use.
1586  *
1587  * @param dev
1588  *   Pointer to private data structure.
1589  *
1590  * @return
1591  *   Pointer to selected Rx burst function.
1592  */
1593 eth_rx_burst_t
1594 mlx5_select_rx_function(struct rte_eth_dev *dev)
1595 {
1596 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1597 
1598 	assert(dev != NULL);
1599 	if (mlx5_check_vec_rx_support(dev) > 0) {
1600 		rx_pkt_burst = mlx5_rx_burst_vec;
1601 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1602 			dev->data->port_id);
1603 	} else if (mlx5_mprq_enabled(dev)) {
1604 		rx_pkt_burst = mlx5_rx_burst_mprq;
1605 	}
1606 	return rx_pkt_burst;
1607 }
1608 
1609 /**
1610  * Check if mlx5 device was removed.
1611  *
1612  * @param dev
1613  *   Pointer to Ethernet device structure.
1614  *
1615  * @return
1616  *   1 when device is removed, otherwise 0.
1617  */
1618 int
1619 mlx5_is_removed(struct rte_eth_dev *dev)
1620 {
1621 	struct ibv_device_attr device_attr;
1622 	struct mlx5_priv *priv = dev->data->dev_private;
1623 
1624 	if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1625 		return 1;
1626 	return 0;
1627 }
1628 
1629 /**
1630  * Get port ID list of mlx5 instances sharing a common device.
1631  *
1632  * @param[in] dev
1633  *   Device to look for.
1634  * @param[out] port_list
1635  *   Result buffer for collected port IDs.
1636  * @param port_list_n
1637  *   Maximum number of entries in result buffer. If 0, @p port_list can be
1638  *   NULL.
1639  *
1640  * @return
1641  *   Number of matching instances regardless of the @p port_list_n
1642  *   parameter, 0 if none were found.
1643  */
1644 unsigned int
1645 mlx5_dev_to_port_id(const struct rte_device *dev, uint16_t *port_list,
1646 		    unsigned int port_list_n)
1647 {
1648 	uint16_t id;
1649 	unsigned int n = 0;
1650 
1651 	RTE_ETH_FOREACH_DEV_OF(id, dev) {
1652 		if (n < port_list_n)
1653 			port_list[n] = id;
1654 		n++;
1655 	}
1656 	return n;
1657 }
1658 
1659 /**
1660  * Get the E-Switch domain id this port belongs to.
1661  *
1662  * @param[in] port
1663  *   Device port id.
1664  * @param[out] es_domain_id
1665  *   E-Switch domain id.
1666  * @param[out] es_port_id
1667  *   The port id of the port in the E-Switch.
1668  *
1669  * @return
1670  *   0 on success, a negative errno value otherwise and rte_errno is set.
1671  */
1672 int
1673 mlx5_port_to_eswitch_info(uint16_t port,
1674 			  uint16_t *es_domain_id, uint16_t *es_port_id)
1675 {
1676 	struct rte_eth_dev *dev;
1677 	struct mlx5_priv *priv;
1678 
1679 	if (port >= RTE_MAX_ETHPORTS) {
1680 		rte_errno = EINVAL;
1681 		return -rte_errno;
1682 	}
1683 	if (!rte_eth_dev_is_valid_port(port)) {
1684 		rte_errno = ENODEV;
1685 		return -rte_errno;
1686 	}
1687 	dev = &rte_eth_devices[port];
1688 	priv = dev->data->dev_private;
1689 	if (!(priv->representor || priv->master)) {
1690 		rte_errno = EINVAL;
1691 		return -rte_errno;
1692 	}
1693 	if (es_domain_id)
1694 		*es_domain_id = priv->domain_id;
1695 	if (es_port_id)
1696 		*es_port_id = priv->vport_id;
1697 	return 0;
1698 }
1699 
1700 /**
1701  * Get switch information associated with network interface.
1702  *
1703  * @param ifindex
1704  *   Network interface index.
1705  * @param[out] info
1706  *   Switch information object, populated in case of success.
1707  *
1708  * @return
1709  *   0 on success, a negative errno value otherwise and rte_errno is set.
1710  */
1711 int
1712 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1713 {
1714 	char ifname[IF_NAMESIZE];
1715 	char port_name[IF_NAMESIZE];
1716 	FILE *file;
1717 	struct mlx5_switch_info data = {
1718 		.master = 0,
1719 		.representor = 0,
1720 		.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1721 		.port_name = 0,
1722 		.switch_id = 0,
1723 	};
1724 	DIR *dir;
1725 	bool port_switch_id_set = false;
1726 	bool device_dir = false;
1727 	char c;
1728 	int ret;
1729 
1730 	if (!if_indextoname(ifindex, ifname)) {
1731 		rte_errno = errno;
1732 		return -rte_errno;
1733 	}
1734 
1735 	MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1736 	      ifname);
1737 	MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1738 	      ifname);
1739 	MKSTR(pci_device, "/sys/class/net/%s/device",
1740 	      ifname);
1741 
1742 	file = fopen(phys_port_name, "rb");
1743 	if (file != NULL) {
1744 		ret = fscanf(file, "%s", port_name);
1745 		fclose(file);
1746 		if (ret == 1)
1747 			mlx5_translate_port_name(port_name, &data);
1748 	}
1749 	file = fopen(phys_switch_id, "rb");
1750 	if (file == NULL) {
1751 		rte_errno = errno;
1752 		return -rte_errno;
1753 	}
1754 	port_switch_id_set =
1755 		fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1756 		c == '\n';
1757 	fclose(file);
1758 	dir = opendir(pci_device);
1759 	if (dir != NULL) {
1760 		closedir(dir);
1761 		device_dir = true;
1762 	}
1763 	if (port_switch_id_set) {
1764 		/* We have some E-Switch configuration. */
1765 		mlx5_sysfs_check_switch_info(device_dir, &data);
1766 	}
1767 	*info = data;
1768 	assert(!(data.master && data.representor));
1769 	if (data.master && data.representor) {
1770 		DRV_LOG(ERR, "ifindex %u device is recognized as master"
1771 			     " and as representor", ifindex);
1772 		rte_errno = ENODEV;
1773 		return -rte_errno;
1774 	}
1775 	return 0;
1776 }
1777 
1778 /**
1779  * Analyze gathered port parameters via Netlink to recognize master
1780  * and representor devices for E-Switch configuration.
1781  *
1782  * @param[in] num_vf_set
1783  *   flag of presence of number of VFs port attribute.
1784  * @param[inout] switch_info
1785  *   Port information, including port name as a number and port name
1786  *   type if recognized
1787  *
1788  * @return
1789  *   master and representor flags are set in switch_info according to
1790  *   recognized parameters (if any).
1791  */
1792 void
1793 mlx5_nl_check_switch_info(bool num_vf_set,
1794 			  struct mlx5_switch_info *switch_info)
1795 {
1796 	switch (switch_info->name_type) {
1797 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1798 		/*
1799 		 * Name is not recognized, assume the master,
1800 		 * check the number of VFs key presence.
1801 		 */
1802 		switch_info->master = num_vf_set;
1803 		break;
1804 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1805 		/*
1806 		 * Name is not set, this assumes the legacy naming
1807 		 * schema for master, just check if there is a
1808 		 * number of VFs key.
1809 		 */
1810 		switch_info->master = num_vf_set;
1811 		break;
1812 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1813 		/* New uplink naming schema recognized. */
1814 		switch_info->master = 1;
1815 		break;
1816 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1817 		/* Legacy representors naming schema. */
1818 		switch_info->representor = !num_vf_set;
1819 		break;
1820 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1821 		/* New representors naming schema. */
1822 		switch_info->representor = 1;
1823 		break;
1824 	}
1825 }
1826 
1827 /**
1828  * Analyze gathered port parameters via sysfs to recognize master
1829  * and representor devices for E-Switch configuration.
1830  *
1831  * @param[in] device_dir
1832  *   flag of presence of "device" directory under port device key.
1833  * @param[inout] switch_info
1834  *   Port information, including port name as a number and port name
1835  *   type if recognized
1836  *
1837  * @return
1838  *   master and representor flags are set in switch_info according to
1839  *   recognized parameters (if any).
1840  */
1841 void
1842 mlx5_sysfs_check_switch_info(bool device_dir,
1843 			     struct mlx5_switch_info *switch_info)
1844 {
1845 	switch (switch_info->name_type) {
1846 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1847 		/*
1848 		 * Name is not recognized, assume the master,
1849 		 * check the device directory presence.
1850 		 */
1851 		switch_info->master = device_dir;
1852 		break;
1853 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1854 		/*
1855 		 * Name is not set, this assumes the legacy naming
1856 		 * schema for master, just check if there is
1857 		 * a device directory.
1858 		 */
1859 		switch_info->master = device_dir;
1860 		break;
1861 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1862 		/* New uplink naming schema recognized. */
1863 		switch_info->master = 1;
1864 		break;
1865 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1866 		/* Legacy representors naming schema. */
1867 		switch_info->representor = !device_dir;
1868 		break;
1869 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1870 		/* New representors naming schema. */
1871 		switch_info->representor = 1;
1872 		break;
1873 	}
1874 }
1875 
1876 /**
1877  * Extract port name, as a number, from sysfs or netlink information.
1878  *
1879  * @param[in] port_name_in
1880  *   String representing the port name.
1881  * @param[out] port_info_out
1882  *   Port information, including port name as a number and port name
1883  *   type if recognized
1884  *
1885  * @return
1886  *   port_name field set according to recognized name format.
1887  */
1888 void
1889 mlx5_translate_port_name(const char *port_name_in,
1890 			 struct mlx5_switch_info *port_info_out)
1891 {
1892 	char pf_c1, pf_c2, vf_c1, vf_c2;
1893 	char *end;
1894 	int sc_items;
1895 
1896 	/*
1897 	 * Check for port-name as a string of the form pf0vf0
1898 	 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
1899 	 */
1900 	sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
1901 			  &pf_c1, &pf_c2, &port_info_out->pf_num,
1902 			  &vf_c1, &vf_c2, &port_info_out->port_name);
1903 	if (sc_items == 6 &&
1904 	    pf_c1 == 'p' && pf_c2 == 'f' &&
1905 	    vf_c1 == 'v' && vf_c2 == 'f') {
1906 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
1907 		return;
1908 	}
1909 	/*
1910 	 * Check for port-name as a string of the form p0
1911 	 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
1912 	 */
1913 	sc_items = sscanf(port_name_in, "%c%d",
1914 			  &pf_c1, &port_info_out->port_name);
1915 	if (sc_items == 2 && pf_c1 == 'p') {
1916 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1917 		return;
1918 	}
1919 	/* Check for port-name as a number (support kernel ver < 5.0 */
1920 	errno = 0;
1921 	port_info_out->port_name = strtol(port_name_in, &end, 0);
1922 	if (!errno &&
1923 	    (size_t)(end - port_name_in) == strlen(port_name_in)) {
1924 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
1925 		return;
1926 	}
1927 	port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
1928 	return;
1929 }
1930