xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 93e30982962b4d05c944555f7c2801e6512db0af)
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_glue.h>
40 #include <mlx5_devx_cmds.h>
41 #include <mlx5_common.h>
42 
43 #include "mlx5.h"
44 #include "mlx5_rxtx.h"
45 #include "mlx5_utils.h"
46 
47 /* Supported speed values found in /usr/include/linux/ethtool.h */
48 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
49 #define SUPPORTED_40000baseKR4_Full (1 << 23)
50 #endif
51 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
52 #define SUPPORTED_40000baseCR4_Full (1 << 24)
53 #endif
54 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
55 #define SUPPORTED_40000baseSR4_Full (1 << 25)
56 #endif
57 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
58 #define SUPPORTED_40000baseLR4_Full (1 << 26)
59 #endif
60 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
61 #define SUPPORTED_56000baseKR4_Full (1 << 27)
62 #endif
63 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
64 #define SUPPORTED_56000baseCR4_Full (1 << 28)
65 #endif
66 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
67 #define SUPPORTED_56000baseSR4_Full (1 << 29)
68 #endif
69 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
70 #define SUPPORTED_56000baseLR4_Full (1 << 30)
71 #endif
72 
73 /* Add defines in case the running kernel is not the same as user headers. */
74 #ifndef ETHTOOL_GLINKSETTINGS
75 struct ethtool_link_settings {
76 	uint32_t cmd;
77 	uint32_t speed;
78 	uint8_t duplex;
79 	uint8_t port;
80 	uint8_t phy_address;
81 	uint8_t autoneg;
82 	uint8_t mdio_support;
83 	uint8_t eth_to_mdix;
84 	uint8_t eth_tp_mdix_ctrl;
85 	int8_t link_mode_masks_nwords;
86 	uint32_t reserved[8];
87 	uint32_t link_mode_masks[];
88 };
89 
90 #define ETHTOOL_GLINKSETTINGS 0x0000004c
91 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
92 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
93 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
94 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
95 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
96 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
97 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
98 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
99 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
100 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
101 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
102 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
103 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
104 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
105 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
106 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
107 #endif
108 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
109 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
110 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
111 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
112 #endif
113 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
114 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
115 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
116 #endif
117 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
118 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
119 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
120 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
121 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
122 #endif
123 
124 /**
125  * Get master interface name from private structure.
126  *
127  * @param[in] dev
128  *   Pointer to Ethernet device.
129  * @param[out] ifname
130  *   Interface name output buffer.
131  *
132  * @return
133  *   0 on success, a negative errno value otherwise and rte_errno is set.
134  */
135 int
136 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
137 {
138 	DIR *dir;
139 	struct dirent *dent;
140 	unsigned int dev_type = 0;
141 	unsigned int dev_port_prev = ~0u;
142 	char match[IF_NAMESIZE] = "";
143 
144 	assert(ibdev_path);
145 	{
146 		MKSTR(path, "%s/device/net", ibdev_path);
147 
148 		dir = opendir(path);
149 		if (dir == NULL) {
150 			rte_errno = errno;
151 			return -rte_errno;
152 		}
153 	}
154 	while ((dent = readdir(dir)) != NULL) {
155 		char *name = dent->d_name;
156 		FILE *file;
157 		unsigned int dev_port;
158 		int r;
159 
160 		if ((name[0] == '.') &&
161 		    ((name[1] == '\0') ||
162 		     ((name[1] == '.') && (name[2] == '\0'))))
163 			continue;
164 
165 		MKSTR(path, "%s/device/net/%s/%s",
166 		      ibdev_path, name,
167 		      (dev_type ? "dev_id" : "dev_port"));
168 
169 		file = fopen(path, "rb");
170 		if (file == NULL) {
171 			if (errno != ENOENT)
172 				continue;
173 			/*
174 			 * Switch to dev_id when dev_port does not exist as
175 			 * is the case with Linux kernel versions < 3.15.
176 			 */
177 try_dev_id:
178 			match[0] = '\0';
179 			if (dev_type)
180 				break;
181 			dev_type = 1;
182 			dev_port_prev = ~0u;
183 			rewinddir(dir);
184 			continue;
185 		}
186 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
187 		fclose(file);
188 		if (r != 1)
189 			continue;
190 		/*
191 		 * Switch to dev_id when dev_port returns the same value for
192 		 * all ports. May happen when using a MOFED release older than
193 		 * 3.0 with a Linux kernel >= 3.15.
194 		 */
195 		if (dev_port == dev_port_prev)
196 			goto try_dev_id;
197 		dev_port_prev = dev_port;
198 		if (dev_port == 0)
199 			strlcpy(match, name, sizeof(match));
200 	}
201 	closedir(dir);
202 	if (match[0] == '\0') {
203 		rte_errno = ENOENT;
204 		return -rte_errno;
205 	}
206 	strncpy(*ifname, match, sizeof(*ifname));
207 	return 0;
208 }
209 
210 /**
211  * Get interface name from private structure.
212  *
213  * This is a port representor-aware version of mlx5_get_master_ifname().
214  *
215  * @param[in] dev
216  *   Pointer to Ethernet device.
217  * @param[out] ifname
218  *   Interface name output buffer.
219  *
220  * @return
221  *   0 on success, a negative errno value otherwise and rte_errno is set.
222  */
223 int
224 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
225 {
226 	struct mlx5_priv *priv = dev->data->dev_private;
227 	unsigned int ifindex;
228 
229 	assert(priv);
230 	assert(priv->sh);
231 	ifindex = mlx5_ifindex(dev);
232 	if (!ifindex) {
233 		if (!priv->representor)
234 			return mlx5_get_master_ifname(priv->sh->ibdev_path,
235 						      ifname);
236 		rte_errno = ENXIO;
237 		return -rte_errno;
238 	}
239 	if (if_indextoname(ifindex, &(*ifname)[0]))
240 		return 0;
241 	rte_errno = errno;
242 	return -rte_errno;
243 }
244 
245 /**
246  * Get the interface index from device name.
247  *
248  * @param[in] dev
249  *   Pointer to Ethernet device.
250  *
251  * @return
252  *   Nonzero interface index on success, zero otherwise and rte_errno is set.
253  */
254 unsigned int
255 mlx5_ifindex(const struct rte_eth_dev *dev)
256 {
257 	struct mlx5_priv *priv = dev->data->dev_private;
258 	unsigned int ifindex;
259 
260 	assert(priv);
261 	assert(priv->if_index);
262 	ifindex = priv->if_index;
263 	if (!ifindex)
264 		rte_errno = ENXIO;
265 	return ifindex;
266 }
267 
268 /**
269  * Perform ifreq ioctl() on associated Ethernet device.
270  *
271  * @param[in] dev
272  *   Pointer to Ethernet device.
273  * @param req
274  *   Request number to pass to ioctl().
275  * @param[out] ifr
276  *   Interface request structure output buffer.
277  *
278  * @return
279  *   0 on success, a negative errno value otherwise and rte_errno is set.
280  */
281 int
282 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
283 {
284 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
285 	int ret = 0;
286 
287 	if (sock == -1) {
288 		rte_errno = errno;
289 		return -rte_errno;
290 	}
291 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
292 	if (ret)
293 		goto error;
294 	ret = ioctl(sock, req, ifr);
295 	if (ret == -1) {
296 		rte_errno = errno;
297 		goto error;
298 	}
299 	close(sock);
300 	return 0;
301 error:
302 	close(sock);
303 	return -rte_errno;
304 }
305 
306 /**
307  * Get device MTU.
308  *
309  * @param dev
310  *   Pointer to Ethernet device.
311  * @param[out] mtu
312  *   MTU value output buffer.
313  *
314  * @return
315  *   0 on success, a negative errno value otherwise and rte_errno is set.
316  */
317 int
318 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
319 {
320 	struct ifreq request;
321 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
322 
323 	if (ret)
324 		return ret;
325 	*mtu = request.ifr_mtu;
326 	return 0;
327 }
328 
329 /**
330  * Set device MTU.
331  *
332  * @param dev
333  *   Pointer to Ethernet device.
334  * @param mtu
335  *   MTU value to set.
336  *
337  * @return
338  *   0 on success, a negative errno value otherwise and rte_errno is set.
339  */
340 static int
341 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
342 {
343 	struct ifreq request = { .ifr_mtu = mtu, };
344 
345 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
346 }
347 
348 /**
349  * Set device flags.
350  *
351  * @param dev
352  *   Pointer to Ethernet device.
353  * @param keep
354  *   Bitmask for flags that must remain untouched.
355  * @param flags
356  *   Bitmask for flags to modify.
357  *
358  * @return
359  *   0 on success, a negative errno value otherwise and rte_errno is set.
360  */
361 int
362 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
363 {
364 	struct ifreq request;
365 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
366 
367 	if (ret)
368 		return ret;
369 	request.ifr_flags &= keep;
370 	request.ifr_flags |= flags & ~keep;
371 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
372 }
373 
374 /**
375  * DPDK callback for Ethernet device configuration.
376  *
377  * @param dev
378  *   Pointer to Ethernet device structure.
379  *
380  * @return
381  *   0 on success, a negative errno value otherwise and rte_errno is set.
382  */
383 int
384 mlx5_dev_configure(struct rte_eth_dev *dev)
385 {
386 	struct mlx5_priv *priv = dev->data->dev_private;
387 	unsigned int rxqs_n = dev->data->nb_rx_queues;
388 	unsigned int txqs_n = dev->data->nb_tx_queues;
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 
411 	if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
412 		dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
413 
414 	memcpy(priv->rss_conf.rss_key,
415 	       use_app_rss_key ?
416 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
417 	       rss_hash_default_key,
418 	       MLX5_RSS_HASH_KEY_LEN);
419 	priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
420 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
421 	priv->rxqs = (void *)dev->data->rx_queues;
422 	priv->txqs = (void *)dev->data->tx_queues;
423 	if (txqs_n != priv->txqs_n) {
424 		DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
425 			dev->data->port_id, priv->txqs_n, txqs_n);
426 		priv->txqs_n = txqs_n;
427 	}
428 	if (rxqs_n > priv->config.ind_table_max_size) {
429 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
430 			dev->data->port_id, rxqs_n);
431 		rte_errno = EINVAL;
432 		return -rte_errno;
433 	}
434 	if (rxqs_n != priv->rxqs_n) {
435 		DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
436 			dev->data->port_id, priv->rxqs_n, rxqs_n);
437 		priv->rxqs_n = rxqs_n;
438 	}
439 	priv->skip_default_rss_reta = 0;
440 	ret = mlx5_proc_priv_init(dev);
441 	if (ret)
442 		return ret;
443 	return 0;
444 }
445 
446 /**
447  * Configure default RSS reta.
448  *
449  * @param dev
450  *   Pointer to Ethernet device structure.
451  *
452  * @return
453  *   0 on success, a negative errno value otherwise and rte_errno is set.
454  */
455 int
456 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
457 {
458 	struct mlx5_priv *priv = dev->data->dev_private;
459 	unsigned int rxqs_n = dev->data->nb_rx_queues;
460 	unsigned int i;
461 	unsigned int j;
462 	unsigned int reta_idx_n;
463 	int ret = 0;
464 	unsigned int *rss_queue_arr = NULL;
465 	unsigned int rss_queue_n = 0;
466 
467 	if (priv->skip_default_rss_reta)
468 		return ret;
469 	rss_queue_arr = rte_malloc("", rxqs_n * sizeof(unsigned int), 0);
470 	if (!rss_queue_arr) {
471 		DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
472 			dev->data->port_id, rxqs_n);
473 		rte_errno = ENOMEM;
474 		return -rte_errno;
475 	}
476 	for (i = 0, j = 0; i < rxqs_n; i++) {
477 		struct mlx5_rxq_data *rxq_data;
478 		struct mlx5_rxq_ctrl *rxq_ctrl;
479 
480 		rxq_data = (*priv->rxqs)[i];
481 		rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
482 		if (rxq_ctrl && rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
483 			rss_queue_arr[j++] = i;
484 	}
485 	rss_queue_n = j;
486 	if (rss_queue_n > priv->config.ind_table_max_size) {
487 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
488 			dev->data->port_id, rss_queue_n);
489 		rte_errno = EINVAL;
490 		rte_free(rss_queue_arr);
491 		return -rte_errno;
492 	}
493 	DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
494 		dev->data->port_id, priv->rxqs_n, rxqs_n);
495 	priv->rxqs_n = rxqs_n;
496 	/*
497 	 * If the requested number of RX queues is not a power of two,
498 	 * use the maximum indirection table size for better balancing.
499 	 * The result is always rounded to the next power of two.
500 	 */
501 	reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
502 				priv->config.ind_table_max_size :
503 				rss_queue_n));
504 	ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
505 	if (ret) {
506 		rte_free(rss_queue_arr);
507 		return ret;
508 	}
509 	/*
510 	 * When the number of RX queues is not a power of two,
511 	 * the remaining table entries are padded with reused WQs
512 	 * and hashes are not spread uniformly.
513 	 */
514 	for (i = 0, j = 0; (i != reta_idx_n); ++i) {
515 		(*priv->reta_idx)[i] = rss_queue_arr[j];
516 		if (++j == rss_queue_n)
517 			j = 0;
518 	}
519 	rte_free(rss_queue_arr);
520 	return ret;
521 }
522 
523 /**
524  * Sets default tuning parameters.
525  *
526  * @param dev
527  *   Pointer to Ethernet device.
528  * @param[out] info
529  *   Info structure output buffer.
530  */
531 static void
532 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
533 {
534 	struct mlx5_priv *priv = dev->data->dev_private;
535 
536 	/* Minimum CPU utilization. */
537 	info->default_rxportconf.ring_size = 256;
538 	info->default_txportconf.ring_size = 256;
539 	info->default_rxportconf.burst_size = MLX5_RX_DEFAULT_BURST;
540 	info->default_txportconf.burst_size = MLX5_TX_DEFAULT_BURST;
541 	if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
542 		info->default_rxportconf.nb_queues = 16;
543 		info->default_txportconf.nb_queues = 16;
544 		if (dev->data->nb_rx_queues > 2 ||
545 		    dev->data->nb_tx_queues > 2) {
546 			/* Max Throughput. */
547 			info->default_rxportconf.ring_size = 2048;
548 			info->default_txportconf.ring_size = 2048;
549 		}
550 	} else {
551 		info->default_rxportconf.nb_queues = 8;
552 		info->default_txportconf.nb_queues = 8;
553 		if (dev->data->nb_rx_queues > 2 ||
554 		    dev->data->nb_tx_queues > 2) {
555 			/* Max Throughput. */
556 			info->default_rxportconf.ring_size = 4096;
557 			info->default_txportconf.ring_size = 4096;
558 		}
559 	}
560 }
561 
562 /**
563  * Sets tx mbuf limiting parameters.
564  *
565  * @param dev
566  *   Pointer to Ethernet device.
567  * @param[out] info
568  *   Info structure output buffer.
569  */
570 static void
571 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
572 {
573 	struct mlx5_priv *priv = dev->data->dev_private;
574 	struct mlx5_dev_config *config = &priv->config;
575 	unsigned int inlen;
576 	uint16_t nb_max;
577 
578 	inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
579 		MLX5_SEND_DEF_INLINE_LEN :
580 		(unsigned int)config->txq_inline_max;
581 	assert(config->txq_inline_min >= 0);
582 	inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
583 	inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
584 			       MLX5_ESEG_MIN_INLINE_SIZE -
585 			       MLX5_WQE_CSEG_SIZE -
586 			       MLX5_WQE_ESEG_SIZE -
587 			       MLX5_WQE_DSEG_SIZE * 2);
588 	nb_max = (MLX5_WQE_SIZE_MAX +
589 		  MLX5_ESEG_MIN_INLINE_SIZE -
590 		  MLX5_WQE_CSEG_SIZE -
591 		  MLX5_WQE_ESEG_SIZE -
592 		  MLX5_WQE_DSEG_SIZE -
593 		  inlen) / MLX5_WSEG_SIZE;
594 	info->tx_desc_lim.nb_seg_max = nb_max;
595 	info->tx_desc_lim.nb_mtu_seg_max = nb_max;
596 }
597 
598 /**
599  * DPDK callback to get information about the device.
600  *
601  * @param dev
602  *   Pointer to Ethernet device structure.
603  * @param[out] info
604  *   Info structure output buffer.
605  */
606 int
607 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
608 {
609 	struct mlx5_priv *priv = dev->data->dev_private;
610 	struct mlx5_dev_config *config = &priv->config;
611 	unsigned int max;
612 
613 	/* FIXME: we should ask the device for these values. */
614 	info->min_rx_bufsize = 32;
615 	info->max_rx_pktlen = 65536;
616 	info->max_lro_pkt_size = MLX5_MAX_LRO_SIZE;
617 	/*
618 	 * Since we need one CQ per QP, the limit is the minimum number
619 	 * between the two values.
620 	 */
621 	max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
622 		      priv->sh->device_attr.orig_attr.max_qp);
623 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
624 	if (max >= 65535)
625 		max = 65535;
626 	info->max_rx_queues = max;
627 	info->max_tx_queues = max;
628 	info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
629 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
630 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
631 				 info->rx_queue_offload_capa);
632 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
633 	info->if_index = mlx5_ifindex(dev);
634 	info->reta_size = priv->reta_idx_n ?
635 		priv->reta_idx_n : config->ind_table_max_size;
636 	info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
637 	info->speed_capa = priv->link_speed_capa;
638 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
639 	mlx5_set_default_params(dev, info);
640 	mlx5_set_txlimit_params(dev, info);
641 	info->switch_info.name = dev->data->name;
642 	info->switch_info.domain_id = priv->domain_id;
643 	info->switch_info.port_id = priv->representor_id;
644 	if (priv->representor) {
645 		uint16_t port_id;
646 
647 		if (priv->pf_bond >= 0) {
648 			/*
649 			 * Switch port ID is opaque value with driver defined
650 			 * format. Push the PF index in bonding configurations
651 			 * in upper four bits of port ID. If we get too many
652 			 * representors (more than 4K) or PFs (more than 15)
653 			 * this approach must be reconsidered.
654 			 */
655 			if ((info->switch_info.port_id >>
656 				MLX5_PORT_ID_BONDING_PF_SHIFT) ||
657 			    priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
658 				DRV_LOG(ERR, "can't update switch port ID"
659 					     " for bonding device");
660 				assert(false);
661 				return -ENODEV;
662 			}
663 			info->switch_info.port_id |=
664 				priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
665 		}
666 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
667 			struct mlx5_priv *opriv =
668 				rte_eth_devices[port_id].data->dev_private;
669 
670 			if (!opriv ||
671 			    opriv->representor ||
672 			    opriv->sh != priv->sh ||
673 			    opriv->domain_id != priv->domain_id)
674 				continue;
675 			/*
676 			 * Override switch name with that of the master
677 			 * device.
678 			 */
679 			info->switch_info.name = opriv->dev_data->name;
680 			break;
681 		}
682 	}
683 	return 0;
684 }
685 
686 /**
687  * Get device current raw clock counter
688  *
689  * @param dev
690  *   Pointer to Ethernet device structure.
691  * @param[out] time
692  *   Current raw clock counter of the device.
693  *
694  * @return
695  *   0 if the clock has correctly been read
696  *   The value of errno in case of error
697  */
698 int
699 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
700 {
701 	struct mlx5_priv *priv = dev->data->dev_private;
702 	struct ibv_context *ctx = priv->sh->ctx;
703 	struct ibv_values_ex values;
704 	int err = 0;
705 
706 	values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
707 	err = mlx5_glue->query_rt_values_ex(ctx, &values);
708 	if (err != 0) {
709 		DRV_LOG(WARNING, "Could not query the clock !");
710 		return err;
711 	}
712 	*clock = values.raw_clock.tv_nsec;
713 	return 0;
714 }
715 
716 /**
717  * Get firmware version of a device.
718  *
719  * @param dev
720  *   Ethernet device port.
721  * @param fw_ver
722  *   String output allocated by caller.
723  * @param fw_size
724  *   Size of the output string, including terminating null byte.
725  *
726  * @return
727  *   0 on success, or the size of the non truncated string if too big.
728  */
729 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
730 {
731 	struct mlx5_priv *priv = dev->data->dev_private;
732 	struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
733 	size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
734 
735 	if (fw_size < size)
736 		return size;
737 	if (fw_ver != NULL)
738 		strlcpy(fw_ver, attr->fw_ver, fw_size);
739 	return 0;
740 }
741 
742 /**
743  * Get supported packet types.
744  *
745  * @param dev
746  *   Pointer to Ethernet device structure.
747  *
748  * @return
749  *   A pointer to the supported Packet types array.
750  */
751 const uint32_t *
752 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
753 {
754 	static const uint32_t ptypes[] = {
755 		/* refers to rxq_cq_to_pkt_type() */
756 		RTE_PTYPE_L2_ETHER,
757 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
758 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
759 		RTE_PTYPE_L4_NONFRAG,
760 		RTE_PTYPE_L4_FRAG,
761 		RTE_PTYPE_L4_TCP,
762 		RTE_PTYPE_L4_UDP,
763 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
764 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
765 		RTE_PTYPE_INNER_L4_NONFRAG,
766 		RTE_PTYPE_INNER_L4_FRAG,
767 		RTE_PTYPE_INNER_L4_TCP,
768 		RTE_PTYPE_INNER_L4_UDP,
769 		RTE_PTYPE_UNKNOWN
770 	};
771 
772 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
773 	    dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
774 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
775 		return ptypes;
776 	return NULL;
777 }
778 
779 /**
780  * Retrieve the master device for representor in the same switch domain.
781  *
782  * @param dev
783  *   Pointer to representor Ethernet device structure.
784  *
785  * @return
786  *   Master device structure  on success, NULL otherwise.
787  */
788 
789 static struct rte_eth_dev *
790 mlx5_find_master_dev(struct rte_eth_dev *dev)
791 {
792 	struct mlx5_priv *priv;
793 	uint16_t port_id;
794 	uint16_t domain_id;
795 
796 	priv = dev->data->dev_private;
797 	domain_id = priv->domain_id;
798 	assert(priv->representor);
799 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
800 		struct mlx5_priv *opriv =
801 			rte_eth_devices[port_id].data->dev_private;
802 		if (opriv &&
803 		    opriv->master &&
804 		    opriv->domain_id == domain_id &&
805 		    opriv->sh == priv->sh)
806 			return &rte_eth_devices[port_id];
807 	}
808 	return NULL;
809 }
810 
811 /**
812  * DPDK callback to retrieve physical link information.
813  *
814  * @param dev
815  *   Pointer to Ethernet device structure.
816  * @param[out] link
817  *   Storage for current link status.
818  *
819  * @return
820  *   0 on success, a negative errno value otherwise and rte_errno is set.
821  */
822 static int
823 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
824 			       struct rte_eth_link *link)
825 {
826 	struct mlx5_priv *priv = dev->data->dev_private;
827 	struct ethtool_cmd edata = {
828 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
829 	};
830 	struct ifreq ifr;
831 	struct rte_eth_link dev_link;
832 	int link_speed = 0;
833 	int ret;
834 
835 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
836 	if (ret) {
837 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
838 			dev->data->port_id, strerror(rte_errno));
839 		return ret;
840 	}
841 	dev_link = (struct rte_eth_link) {
842 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
843 				(ifr.ifr_flags & IFF_RUNNING)),
844 	};
845 	ifr = (struct ifreq) {
846 		.ifr_data = (void *)&edata,
847 	};
848 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
849 	if (ret) {
850 		if (ret == -ENOTSUP && priv->representor) {
851 			struct rte_eth_dev *master;
852 
853 			/*
854 			 * For representors we can try to inherit link
855 			 * settings from the master device. Actually
856 			 * link settings do not make a lot of sense
857 			 * for representors due to missing physical
858 			 * link. The old kernel drivers supported
859 			 * emulated settings query for representors,
860 			 * the new ones do not, so we have to add
861 			 * this code for compatibility issues.
862 			 */
863 			master = mlx5_find_master_dev(dev);
864 			if (master) {
865 				ifr = (struct ifreq) {
866 					.ifr_data = (void *)&edata,
867 				};
868 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
869 			}
870 		}
871 		if (ret) {
872 			DRV_LOG(WARNING,
873 				"port %u ioctl(SIOCETHTOOL,"
874 				" ETHTOOL_GSET) failed: %s",
875 				dev->data->port_id, strerror(rte_errno));
876 			return ret;
877 		}
878 	}
879 	link_speed = ethtool_cmd_speed(&edata);
880 	if (link_speed == -1)
881 		dev_link.link_speed = ETH_SPEED_NUM_NONE;
882 	else
883 		dev_link.link_speed = link_speed;
884 	priv->link_speed_capa = 0;
885 	if (edata.supported & SUPPORTED_Autoneg)
886 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
887 	if (edata.supported & (SUPPORTED_1000baseT_Full |
888 			       SUPPORTED_1000baseKX_Full))
889 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
890 	if (edata.supported & SUPPORTED_10000baseKR_Full)
891 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
892 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
893 			       SUPPORTED_40000baseCR4_Full |
894 			       SUPPORTED_40000baseSR4_Full |
895 			       SUPPORTED_40000baseLR4_Full))
896 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
897 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
898 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
899 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
900 			ETH_LINK_SPEED_FIXED);
901 	if (((dev_link.link_speed && !dev_link.link_status) ||
902 	     (!dev_link.link_speed && dev_link.link_status))) {
903 		rte_errno = EAGAIN;
904 		return -rte_errno;
905 	}
906 	*link = dev_link;
907 	return 0;
908 }
909 
910 /**
911  * Retrieve physical link information (unlocked version using new ioctl).
912  *
913  * @param dev
914  *   Pointer to Ethernet device structure.
915  * @param[out] link
916  *   Storage for current link status.
917  *
918  * @return
919  *   0 on success, a negative errno value otherwise and rte_errno is set.
920  */
921 static int
922 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
923 			     struct rte_eth_link *link)
924 
925 {
926 	struct mlx5_priv *priv = dev->data->dev_private;
927 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
928 	struct ifreq ifr;
929 	struct rte_eth_link dev_link;
930 	struct rte_eth_dev *master = NULL;
931 	uint64_t sc;
932 	int ret;
933 
934 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
935 	if (ret) {
936 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
937 			dev->data->port_id, strerror(rte_errno));
938 		return ret;
939 	}
940 	dev_link = (struct rte_eth_link) {
941 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
942 				(ifr.ifr_flags & IFF_RUNNING)),
943 	};
944 	ifr = (struct ifreq) {
945 		.ifr_data = (void *)&gcmd,
946 	};
947 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
948 	if (ret) {
949 		if (ret == -ENOTSUP && priv->representor) {
950 			/*
951 			 * For representors we can try to inherit link
952 			 * settings from the master device. Actually
953 			 * link settings do not make a lot of sense
954 			 * for representors due to missing physical
955 			 * link. The old kernel drivers supported
956 			 * emulated settings query for representors,
957 			 * the new ones do not, so we have to add
958 			 * this code for compatibility issues.
959 			 */
960 			master = mlx5_find_master_dev(dev);
961 			if (master) {
962 				ifr = (struct ifreq) {
963 					.ifr_data = (void *)&gcmd,
964 				};
965 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
966 			}
967 		}
968 		if (ret) {
969 			DRV_LOG(DEBUG,
970 				"port %u ioctl(SIOCETHTOOL,"
971 				" ETHTOOL_GLINKSETTINGS) failed: %s",
972 				dev->data->port_id, strerror(rte_errno));
973 			return ret;
974 		}
975 
976 	}
977 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
978 
979 	alignas(struct ethtool_link_settings)
980 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
981 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
982 	struct ethtool_link_settings *ecmd = (void *)data;
983 
984 	*ecmd = gcmd;
985 	ifr.ifr_data = (void *)ecmd;
986 	ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
987 	if (ret) {
988 		DRV_LOG(DEBUG,
989 			"port %u ioctl(SIOCETHTOOL,"
990 			"ETHTOOL_GLINKSETTINGS) failed: %s",
991 			dev->data->port_id, strerror(rte_errno));
992 		return ret;
993 	}
994 	dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
995 							    ecmd->speed;
996 	sc = ecmd->link_mode_masks[0] |
997 		((uint64_t)ecmd->link_mode_masks[1] << 32);
998 	priv->link_speed_capa = 0;
999 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
1000 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
1001 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
1002 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
1003 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
1004 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
1005 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
1006 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
1007 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
1008 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
1009 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
1010 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
1011 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
1012 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
1013 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
1014 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
1015 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
1016 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
1017 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
1018 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
1019 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
1020 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
1021 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
1022 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
1023 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
1024 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
1025 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
1026 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
1027 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
1028 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
1029 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
1030 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
1031 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
1032 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
1033 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
1034 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
1035 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
1036 				  ETH_LINK_SPEED_FIXED);
1037 	if (((dev_link.link_speed && !dev_link.link_status) ||
1038 	     (!dev_link.link_speed && dev_link.link_status))) {
1039 		rte_errno = EAGAIN;
1040 		return -rte_errno;
1041 	}
1042 	*link = dev_link;
1043 	return 0;
1044 }
1045 
1046 /**
1047  * DPDK callback to retrieve physical link information.
1048  *
1049  * @param dev
1050  *   Pointer to Ethernet device structure.
1051  * @param wait_to_complete
1052  *   Wait for request completion.
1053  *
1054  * @return
1055  *   0 if link status was not updated, positive if it was, a negative errno
1056  *   value otherwise and rte_errno is set.
1057  */
1058 int
1059 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1060 {
1061 	int ret;
1062 	struct rte_eth_link dev_link;
1063 	time_t start_time = time(NULL);
1064 	int retry = MLX5_GET_LINK_STATUS_RETRY_COUNT;
1065 
1066 	do {
1067 		ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1068 		if (ret == -ENOTSUP)
1069 			ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1070 		if (ret == 0)
1071 			break;
1072 		/* Handle wait to complete situation. */
1073 		if ((wait_to_complete || retry) && ret == -EAGAIN) {
1074 			if (abs((int)difftime(time(NULL), start_time)) <
1075 			    MLX5_LINK_STATUS_TIMEOUT) {
1076 				usleep(0);
1077 				continue;
1078 			} else {
1079 				rte_errno = EBUSY;
1080 				return -rte_errno;
1081 			}
1082 		} else if (ret < 0) {
1083 			return ret;
1084 		}
1085 	} while (wait_to_complete || retry-- > 0);
1086 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
1087 		       sizeof(struct rte_eth_link));
1088 	dev->data->dev_link = dev_link;
1089 	return ret;
1090 }
1091 
1092 /**
1093  * DPDK callback to change the MTU.
1094  *
1095  * @param dev
1096  *   Pointer to Ethernet device structure.
1097  * @param in_mtu
1098  *   New MTU.
1099  *
1100  * @return
1101  *   0 on success, a negative errno value otherwise and rte_errno is set.
1102  */
1103 int
1104 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1105 {
1106 	struct mlx5_priv *priv = dev->data->dev_private;
1107 	uint16_t kern_mtu = 0;
1108 	int ret;
1109 
1110 	ret = mlx5_get_mtu(dev, &kern_mtu);
1111 	if (ret)
1112 		return ret;
1113 	/* Set kernel interface MTU first. */
1114 	ret = mlx5_set_mtu(dev, mtu);
1115 	if (ret)
1116 		return ret;
1117 	ret = mlx5_get_mtu(dev, &kern_mtu);
1118 	if (ret)
1119 		return ret;
1120 	if (kern_mtu == mtu) {
1121 		priv->mtu = mtu;
1122 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1123 			dev->data->port_id, mtu);
1124 		return 0;
1125 	}
1126 	rte_errno = EAGAIN;
1127 	return -rte_errno;
1128 }
1129 
1130 /**
1131  * DPDK callback to get flow control status.
1132  *
1133  * @param dev
1134  *   Pointer to Ethernet device structure.
1135  * @param[out] fc_conf
1136  *   Flow control output buffer.
1137  *
1138  * @return
1139  *   0 on success, a negative errno value otherwise and rte_errno is set.
1140  */
1141 int
1142 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1143 {
1144 	struct ifreq ifr;
1145 	struct ethtool_pauseparam ethpause = {
1146 		.cmd = ETHTOOL_GPAUSEPARAM
1147 	};
1148 	int ret;
1149 
1150 	ifr.ifr_data = (void *)&ethpause;
1151 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1152 	if (ret) {
1153 		DRV_LOG(WARNING,
1154 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1155 			" %s",
1156 			dev->data->port_id, strerror(rte_errno));
1157 		return ret;
1158 	}
1159 	fc_conf->autoneg = ethpause.autoneg;
1160 	if (ethpause.rx_pause && ethpause.tx_pause)
1161 		fc_conf->mode = RTE_FC_FULL;
1162 	else if (ethpause.rx_pause)
1163 		fc_conf->mode = RTE_FC_RX_PAUSE;
1164 	else if (ethpause.tx_pause)
1165 		fc_conf->mode = RTE_FC_TX_PAUSE;
1166 	else
1167 		fc_conf->mode = RTE_FC_NONE;
1168 	return 0;
1169 }
1170 
1171 /**
1172  * DPDK callback to modify flow control parameters.
1173  *
1174  * @param dev
1175  *   Pointer to Ethernet device structure.
1176  * @param[in] fc_conf
1177  *   Flow control parameters.
1178  *
1179  * @return
1180  *   0 on success, a negative errno value otherwise and rte_errno is set.
1181  */
1182 int
1183 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1184 {
1185 	struct ifreq ifr;
1186 	struct ethtool_pauseparam ethpause = {
1187 		.cmd = ETHTOOL_SPAUSEPARAM
1188 	};
1189 	int ret;
1190 
1191 	ifr.ifr_data = (void *)&ethpause;
1192 	ethpause.autoneg = fc_conf->autoneg;
1193 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1194 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1195 		ethpause.rx_pause = 1;
1196 	else
1197 		ethpause.rx_pause = 0;
1198 
1199 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1200 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1201 		ethpause.tx_pause = 1;
1202 	else
1203 		ethpause.tx_pause = 0;
1204 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1205 	if (ret) {
1206 		DRV_LOG(WARNING,
1207 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1208 			" failed: %s",
1209 			dev->data->port_id, strerror(rte_errno));
1210 		return ret;
1211 	}
1212 	return 0;
1213 }
1214 
1215 /**
1216  * Handle asynchronous removal event for entire multiport device.
1217  *
1218  * @param sh
1219  *   Infiniband device shared context.
1220  */
1221 static void
1222 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1223 {
1224 	uint32_t i;
1225 
1226 	for (i = 0; i < sh->max_port; ++i) {
1227 		struct rte_eth_dev *dev;
1228 
1229 		if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1230 			/*
1231 			 * Or not existing port either no
1232 			 * handler installed for this port.
1233 			 */
1234 			continue;
1235 		}
1236 		dev = &rte_eth_devices[sh->port[i].ih_port_id];
1237 		assert(dev);
1238 		if (dev->data->dev_conf.intr_conf.rmv)
1239 			_rte_eth_dev_callback_process
1240 				(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1241 	}
1242 }
1243 
1244 /**
1245  * Handle shared asynchronous events the NIC (removal event
1246  * and link status change). Supports multiport IB device.
1247  *
1248  * @param cb_arg
1249  *   Callback argument.
1250  */
1251 void
1252 mlx5_dev_interrupt_handler(void *cb_arg)
1253 {
1254 	struct mlx5_ibv_shared *sh = cb_arg;
1255 	struct ibv_async_event event;
1256 
1257 	/* Read all message from the IB device and acknowledge them. */
1258 	for (;;) {
1259 		struct rte_eth_dev *dev;
1260 		uint32_t tmp;
1261 
1262 		if (mlx5_glue->get_async_event(sh->ctx, &event))
1263 			break;
1264 		/* Retrieve and check IB port index. */
1265 		tmp = (uint32_t)event.element.port_num;
1266 		if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1267 			/*
1268 			 * The DEVICE_FATAL event is called once for
1269 			 * entire device without port specifying.
1270 			 * We should notify all existing ports.
1271 			 */
1272 			mlx5_glue->ack_async_event(&event);
1273 			mlx5_dev_interrupt_device_fatal(sh);
1274 			continue;
1275 		}
1276 		assert(tmp && (tmp <= sh->max_port));
1277 		if (!tmp) {
1278 			/* Unsupported devive level event. */
1279 			mlx5_glue->ack_async_event(&event);
1280 			DRV_LOG(DEBUG,
1281 				"unsupported common event (type %d)",
1282 				event.event_type);
1283 			continue;
1284 		}
1285 		if (tmp > sh->max_port) {
1286 			/* Invalid IB port index. */
1287 			mlx5_glue->ack_async_event(&event);
1288 			DRV_LOG(DEBUG,
1289 				"cannot handle an event (type %d)"
1290 				"due to invalid IB port index (%u)",
1291 				event.event_type, tmp);
1292 			continue;
1293 		}
1294 		if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1295 			/* No handler installed. */
1296 			mlx5_glue->ack_async_event(&event);
1297 			DRV_LOG(DEBUG,
1298 				"cannot handle an event (type %d)"
1299 				"due to no handler installed for port %u",
1300 				event.event_type, tmp);
1301 			continue;
1302 		}
1303 		/* Retrieve ethernet device descriptor. */
1304 		tmp = sh->port[tmp - 1].ih_port_id;
1305 		dev = &rte_eth_devices[tmp];
1306 		assert(dev);
1307 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1308 		     event.event_type == IBV_EVENT_PORT_ERR) &&
1309 			dev->data->dev_conf.intr_conf.lsc) {
1310 			mlx5_glue->ack_async_event(&event);
1311 			if (mlx5_link_update(dev, 0) == -EAGAIN) {
1312 				usleep(0);
1313 				continue;
1314 			}
1315 			_rte_eth_dev_callback_process
1316 				(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1317 			continue;
1318 		}
1319 		DRV_LOG(DEBUG,
1320 			"port %u cannot handle an unknown event (type %d)",
1321 			dev->data->port_id, event.event_type);
1322 		mlx5_glue->ack_async_event(&event);
1323 	}
1324 }
1325 
1326 /*
1327  * Unregister callback handler safely. The handler may be active
1328  * while we are trying to unregister it, in this case code -EAGAIN
1329  * is returned by rte_intr_callback_unregister(). This routine checks
1330  * the return code and tries to unregister handler again.
1331  *
1332  * @param handle
1333  *   interrupt handle
1334  * @param cb_fn
1335  *   pointer to callback routine
1336  * @cb_arg
1337  *   opaque callback parameter
1338  */
1339 void
1340 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1341 			      rte_intr_callback_fn cb_fn, void *cb_arg)
1342 {
1343 	/*
1344 	 * Try to reduce timeout management overhead by not calling
1345 	 * the timer related routines on the first iteration. If the
1346 	 * unregistering succeeds on first call there will be no
1347 	 * timer calls at all.
1348 	 */
1349 	uint64_t twait = 0;
1350 	uint64_t start = 0;
1351 
1352 	do {
1353 		int ret;
1354 
1355 		ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1356 		if (ret >= 0)
1357 			return;
1358 		if (ret != -EAGAIN) {
1359 			DRV_LOG(INFO, "failed to unregister interrupt"
1360 				      " handler (error: %d)", ret);
1361 			assert(false);
1362 			return;
1363 		}
1364 		if (twait) {
1365 			struct timespec onems;
1366 
1367 			/* Wait one millisecond and try again. */
1368 			onems.tv_sec = 0;
1369 			onems.tv_nsec = NS_PER_S / MS_PER_S;
1370 			nanosleep(&onems, 0);
1371 			/* Check whether one second elapsed. */
1372 			if ((rte_get_timer_cycles() - start) <= twait)
1373 				continue;
1374 		} else {
1375 			/*
1376 			 * We get the amount of timer ticks for one second.
1377 			 * If this amount elapsed it means we spent one
1378 			 * second in waiting. This branch is executed once
1379 			 * on first iteration.
1380 			 */
1381 			twait = rte_get_timer_hz();
1382 			assert(twait);
1383 		}
1384 		/*
1385 		 * Timeout elapsed, show message (once a second) and retry.
1386 		 * We have no other acceptable option here, if we ignore
1387 		 * the unregistering return code the handler will not
1388 		 * be unregistered, fd will be closed and we may get the
1389 		 * crush. Hanging and messaging in the loop seems not to be
1390 		 * the worst choice.
1391 		 */
1392 		DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1393 		start = rte_get_timer_cycles();
1394 	} while (true);
1395 }
1396 
1397 /**
1398  * Handle DEVX interrupts from the NIC.
1399  * This function is probably called from the DPDK host thread.
1400  *
1401  * @param cb_arg
1402  *   Callback argument.
1403  */
1404 void
1405 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1406 {
1407 #ifndef HAVE_IBV_DEVX_ASYNC
1408 	(void)cb_arg;
1409 	return;
1410 #else
1411 	struct mlx5_ibv_shared *sh = cb_arg;
1412 	union {
1413 		struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1414 		uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1415 			    MLX5_ST_SZ_BYTES(traffic_counter) +
1416 			    sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1417 	} out;
1418 	uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1419 
1420 	while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1421 						   &out.cmd_resp,
1422 						   sizeof(out.buf)))
1423 		mlx5_flow_async_pool_query_handle
1424 			(sh, (uint64_t)out.cmd_resp.wr_id,
1425 			 mlx5_devx_get_out_command_status(buf));
1426 #endif /* HAVE_IBV_DEVX_ASYNC */
1427 }
1428 
1429 /**
1430  * Uninstall shared asynchronous device events handler.
1431  * This function is implemented to support event sharing
1432  * between multiple ports of single IB device.
1433  *
1434  * @param dev
1435  *   Pointer to Ethernet device.
1436  */
1437 static void
1438 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1439 {
1440 	struct mlx5_priv *priv = dev->data->dev_private;
1441 	struct mlx5_ibv_shared *sh = priv->sh;
1442 
1443 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1444 		return;
1445 	pthread_mutex_lock(&sh->intr_mutex);
1446 	assert(priv->ibv_port);
1447 	assert(priv->ibv_port <= sh->max_port);
1448 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1449 	if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1450 		goto exit;
1451 	assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1452 					(uint32_t)dev->data->port_id);
1453 	assert(sh->intr_cnt);
1454 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1455 	if (!sh->intr_cnt || --sh->intr_cnt)
1456 		goto exit;
1457 	mlx5_intr_callback_unregister(&sh->intr_handle,
1458 				     mlx5_dev_interrupt_handler, sh);
1459 	sh->intr_handle.fd = 0;
1460 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1461 exit:
1462 	pthread_mutex_unlock(&sh->intr_mutex);
1463 }
1464 
1465 /**
1466  * Uninstall devx shared asynchronous device events handler.
1467  * This function is implemeted to support event sharing
1468  * between multiple ports of single IB device.
1469  *
1470  * @param dev
1471  *   Pointer to Ethernet device.
1472  */
1473 static void
1474 mlx5_dev_shared_handler_devx_uninstall(struct rte_eth_dev *dev)
1475 {
1476 	struct mlx5_priv *priv = dev->data->dev_private;
1477 	struct mlx5_ibv_shared *sh = priv->sh;
1478 
1479 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1480 		return;
1481 	pthread_mutex_lock(&sh->intr_mutex);
1482 	assert(priv->ibv_port);
1483 	assert(priv->ibv_port <= sh->max_port);
1484 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1485 	if (sh->port[priv->ibv_port - 1].devx_ih_port_id >= RTE_MAX_ETHPORTS)
1486 		goto exit;
1487 	assert(sh->port[priv->ibv_port - 1].devx_ih_port_id ==
1488 					(uint32_t)dev->data->port_id);
1489 	sh->port[priv->ibv_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1490 	if (!sh->devx_intr_cnt || --sh->devx_intr_cnt)
1491 		goto exit;
1492 	if (sh->intr_handle_devx.fd) {
1493 		rte_intr_callback_unregister(&sh->intr_handle_devx,
1494 					     mlx5_dev_interrupt_handler_devx,
1495 					     sh);
1496 		sh->intr_handle_devx.fd = 0;
1497 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1498 	}
1499 	if (sh->devx_comp) {
1500 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1501 		sh->devx_comp = NULL;
1502 	}
1503 exit:
1504 	pthread_mutex_unlock(&sh->intr_mutex);
1505 }
1506 
1507 /**
1508  * Install shared asynchronous device events handler.
1509  * This function is implemented to support event sharing
1510  * between multiple ports of single IB device.
1511  *
1512  * @param dev
1513  *   Pointer to Ethernet device.
1514  */
1515 static void
1516 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1517 {
1518 	struct mlx5_priv *priv = dev->data->dev_private;
1519 	struct mlx5_ibv_shared *sh = priv->sh;
1520 	int ret;
1521 	int flags;
1522 
1523 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1524 		return;
1525 	pthread_mutex_lock(&sh->intr_mutex);
1526 	assert(priv->ibv_port);
1527 	assert(priv->ibv_port <= sh->max_port);
1528 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1529 	if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1530 		/* The handler is already installed for this port. */
1531 		assert(sh->intr_cnt);
1532 		goto exit;
1533 	}
1534 	if (sh->intr_cnt) {
1535 		sh->port[priv->ibv_port - 1].ih_port_id =
1536 						(uint32_t)dev->data->port_id;
1537 		sh->intr_cnt++;
1538 		goto exit;
1539 	}
1540 	/* No shared handler installed. */
1541 	assert(sh->ctx->async_fd > 0);
1542 	flags = fcntl(sh->ctx->async_fd, F_GETFL);
1543 	ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1544 	if (ret) {
1545 		DRV_LOG(INFO, "failed to change file descriptor async event"
1546 			" queue");
1547 		/* Indicate there will be no interrupts. */
1548 		dev->data->dev_conf.intr_conf.lsc = 0;
1549 		dev->data->dev_conf.intr_conf.rmv = 0;
1550 	} else {
1551 		sh->intr_handle.fd = sh->ctx->async_fd;
1552 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1553 		rte_intr_callback_register(&sh->intr_handle,
1554 					   mlx5_dev_interrupt_handler, sh);
1555 		sh->intr_cnt++;
1556 		sh->port[priv->ibv_port - 1].ih_port_id =
1557 						(uint32_t)dev->data->port_id;
1558 	}
1559 exit:
1560 	pthread_mutex_unlock(&sh->intr_mutex);
1561 }
1562 
1563 /**
1564  * Install devx shared asyncronous device events handler.
1565  * This function is implemeted to support event sharing
1566  * between multiple ports of single IB device.
1567  *
1568  * @param dev
1569  *   Pointer to Ethernet device.
1570  */
1571 static void
1572 mlx5_dev_shared_handler_devx_install(struct rte_eth_dev *dev)
1573 {
1574 	struct mlx5_priv *priv = dev->data->dev_private;
1575 	struct mlx5_ibv_shared *sh = priv->sh;
1576 
1577 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1578 		return;
1579 	pthread_mutex_lock(&sh->intr_mutex);
1580 	assert(priv->ibv_port);
1581 	assert(priv->ibv_port <= sh->max_port);
1582 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1583 	if (sh->port[priv->ibv_port - 1].devx_ih_port_id < RTE_MAX_ETHPORTS) {
1584 		/* The handler is already installed for this port. */
1585 		assert(sh->devx_intr_cnt);
1586 		goto exit;
1587 	}
1588 	if (sh->devx_intr_cnt) {
1589 		sh->devx_intr_cnt++;
1590 		sh->port[priv->ibv_port - 1].devx_ih_port_id =
1591 					(uint32_t)dev->data->port_id;
1592 		goto exit;
1593 	}
1594 	if (priv->config.devx) {
1595 #ifndef HAVE_IBV_DEVX_ASYNC
1596 		goto exit;
1597 #else
1598 		sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1599 		if (sh->devx_comp) {
1600 			int flags = fcntl(sh->devx_comp->fd, F_GETFL);
1601 			int ret = fcntl(sh->devx_comp->fd, F_SETFL,
1602 				    flags | O_NONBLOCK);
1603 
1604 			if (ret) {
1605 				DRV_LOG(INFO, "failed to change file descriptor"
1606 					" devx async event queue");
1607 			} else {
1608 				sh->intr_handle_devx.fd = sh->devx_comp->fd;
1609 				sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1610 				rte_intr_callback_register
1611 					(&sh->intr_handle_devx,
1612 					 mlx5_dev_interrupt_handler_devx, sh);
1613 				sh->devx_intr_cnt++;
1614 				sh->port[priv->ibv_port - 1].devx_ih_port_id =
1615 						(uint32_t)dev->data->port_id;
1616 			}
1617 		}
1618 #endif /* HAVE_IBV_DEVX_ASYNC */
1619 	}
1620 exit:
1621 	pthread_mutex_unlock(&sh->intr_mutex);
1622 }
1623 
1624 /**
1625  * Uninstall interrupt handler.
1626  *
1627  * @param dev
1628  *   Pointer to Ethernet device.
1629  */
1630 void
1631 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1632 {
1633 	mlx5_dev_shared_handler_uninstall(dev);
1634 }
1635 
1636 /**
1637  * Install interrupt handler.
1638  *
1639  * @param dev
1640  *   Pointer to Ethernet device.
1641  */
1642 void
1643 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1644 {
1645 	mlx5_dev_shared_handler_install(dev);
1646 }
1647 
1648 /**
1649  * Devx uninstall interrupt handler.
1650  *
1651  * @param dev
1652  *   Pointer to Ethernet device.
1653  */
1654 void
1655 mlx5_dev_interrupt_handler_devx_uninstall(struct rte_eth_dev *dev)
1656 {
1657 	mlx5_dev_shared_handler_devx_uninstall(dev);
1658 }
1659 
1660 /**
1661  * Devx install interrupt handler.
1662  *
1663  * @param dev
1664  *   Pointer to Ethernet device.
1665  */
1666 void
1667 mlx5_dev_interrupt_handler_devx_install(struct rte_eth_dev *dev)
1668 {
1669 	mlx5_dev_shared_handler_devx_install(dev);
1670 }
1671 
1672 /**
1673  * DPDK callback to bring the link DOWN.
1674  *
1675  * @param dev
1676  *   Pointer to Ethernet device structure.
1677  *
1678  * @return
1679  *   0 on success, a negative errno value otherwise and rte_errno is set.
1680  */
1681 int
1682 mlx5_set_link_down(struct rte_eth_dev *dev)
1683 {
1684 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1685 }
1686 
1687 /**
1688  * DPDK callback to bring the link UP.
1689  *
1690  * @param dev
1691  *   Pointer to Ethernet device structure.
1692  *
1693  * @return
1694  *   0 on success, a negative errno value otherwise and rte_errno is set.
1695  */
1696 int
1697 mlx5_set_link_up(struct rte_eth_dev *dev)
1698 {
1699 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1700 }
1701 
1702 /**
1703  * Configure the RX function to use.
1704  *
1705  * @param dev
1706  *   Pointer to private data structure.
1707  *
1708  * @return
1709  *   Pointer to selected Rx burst function.
1710  */
1711 eth_rx_burst_t
1712 mlx5_select_rx_function(struct rte_eth_dev *dev)
1713 {
1714 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1715 
1716 	assert(dev != NULL);
1717 	if (mlx5_check_vec_rx_support(dev) > 0) {
1718 		rx_pkt_burst = mlx5_rx_burst_vec;
1719 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1720 			dev->data->port_id);
1721 	} else if (mlx5_mprq_enabled(dev)) {
1722 		rx_pkt_burst = mlx5_rx_burst_mprq;
1723 	}
1724 	return rx_pkt_burst;
1725 }
1726 
1727 /**
1728  * Check if mlx5 device was removed.
1729  *
1730  * @param dev
1731  *   Pointer to Ethernet device structure.
1732  *
1733  * @return
1734  *   1 when device is removed, otherwise 0.
1735  */
1736 int
1737 mlx5_is_removed(struct rte_eth_dev *dev)
1738 {
1739 	struct ibv_device_attr device_attr;
1740 	struct mlx5_priv *priv = dev->data->dev_private;
1741 
1742 	if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1743 		return 1;
1744 	return 0;
1745 }
1746 
1747 /**
1748  * Get the E-Switch parameters by port id.
1749  *
1750  * @param[in] port
1751  *   Device port id.
1752  * @param[in] valid
1753  *   Device port id is valid, skip check. This flag is useful
1754  *   when trials are performed from probing and device is not
1755  *   flagged as valid yet (in attaching process).
1756  * @param[out] es_domain_id
1757  *   E-Switch domain id.
1758  * @param[out] es_port_id
1759  *   The port id of the port in the E-Switch.
1760  *
1761  * @return
1762  *   pointer to device private data structure containing data needed
1763  *   on success, NULL otherwise and rte_errno is set.
1764  */
1765 struct mlx5_priv *
1766 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
1767 {
1768 	struct rte_eth_dev *dev;
1769 	struct mlx5_priv *priv;
1770 
1771 	if (port >= RTE_MAX_ETHPORTS) {
1772 		rte_errno = EINVAL;
1773 		return NULL;
1774 	}
1775 	if (!valid && !rte_eth_dev_is_valid_port(port)) {
1776 		rte_errno = ENODEV;
1777 		return NULL;
1778 	}
1779 	dev = &rte_eth_devices[port];
1780 	priv = dev->data->dev_private;
1781 	if (!(priv->representor || priv->master)) {
1782 		rte_errno = EINVAL;
1783 		return NULL;
1784 	}
1785 	return priv;
1786 }
1787 
1788 /**
1789  * Get the E-Switch parameters by device instance.
1790  *
1791  * @param[in] port
1792  *   Device port id.
1793  * @param[out] es_domain_id
1794  *   E-Switch domain id.
1795  * @param[out] es_port_id
1796  *   The port id of the port in the E-Switch.
1797  *
1798  * @return
1799  *   pointer to device private data structure containing data needed
1800  *   on success, NULL otherwise and rte_errno is set.
1801  */
1802 struct mlx5_priv *
1803 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1804 {
1805 	struct mlx5_priv *priv;
1806 
1807 	priv = dev->data->dev_private;
1808 	if (!(priv->representor || priv->master)) {
1809 		rte_errno = EINVAL;
1810 		return NULL;
1811 	}
1812 	return priv;
1813 }
1814 
1815 /**
1816  * Get switch information associated with network interface.
1817  *
1818  * @param ifindex
1819  *   Network interface index.
1820  * @param[out] info
1821  *   Switch information object, populated in case of success.
1822  *
1823  * @return
1824  *   0 on success, a negative errno value otherwise and rte_errno is set.
1825  */
1826 int
1827 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1828 {
1829 	char ifname[IF_NAMESIZE];
1830 	char port_name[IF_NAMESIZE];
1831 	FILE *file;
1832 	struct mlx5_switch_info data = {
1833 		.master = 0,
1834 		.representor = 0,
1835 		.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1836 		.port_name = 0,
1837 		.switch_id = 0,
1838 	};
1839 	DIR *dir;
1840 	bool port_switch_id_set = false;
1841 	bool device_dir = false;
1842 	char c;
1843 	int ret;
1844 
1845 	if (!if_indextoname(ifindex, ifname)) {
1846 		rte_errno = errno;
1847 		return -rte_errno;
1848 	}
1849 
1850 	MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1851 	      ifname);
1852 	MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1853 	      ifname);
1854 	MKSTR(pci_device, "/sys/class/net/%s/device",
1855 	      ifname);
1856 
1857 	file = fopen(phys_port_name, "rb");
1858 	if (file != NULL) {
1859 		ret = fscanf(file, "%s", port_name);
1860 		fclose(file);
1861 		if (ret == 1)
1862 			mlx5_translate_port_name(port_name, &data);
1863 	}
1864 	file = fopen(phys_switch_id, "rb");
1865 	if (file == NULL) {
1866 		rte_errno = errno;
1867 		return -rte_errno;
1868 	}
1869 	port_switch_id_set =
1870 		fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1871 		c == '\n';
1872 	fclose(file);
1873 	dir = opendir(pci_device);
1874 	if (dir != NULL) {
1875 		closedir(dir);
1876 		device_dir = true;
1877 	}
1878 	if (port_switch_id_set) {
1879 		/* We have some E-Switch configuration. */
1880 		mlx5_sysfs_check_switch_info(device_dir, &data);
1881 	}
1882 	*info = data;
1883 	assert(!(data.master && data.representor));
1884 	if (data.master && data.representor) {
1885 		DRV_LOG(ERR, "ifindex %u device is recognized as master"
1886 			     " and as representor", ifindex);
1887 		rte_errno = ENODEV;
1888 		return -rte_errno;
1889 	}
1890 	return 0;
1891 }
1892 
1893 /**
1894  * Analyze gathered port parameters via Netlink to recognize master
1895  * and representor devices for E-Switch configuration.
1896  *
1897  * @param[in] num_vf_set
1898  *   flag of presence of number of VFs port attribute.
1899  * @param[inout] switch_info
1900  *   Port information, including port name as a number and port name
1901  *   type if recognized
1902  *
1903  * @return
1904  *   master and representor flags are set in switch_info according to
1905  *   recognized parameters (if any).
1906  */
1907 void
1908 mlx5_nl_check_switch_info(bool num_vf_set,
1909 			  struct mlx5_switch_info *switch_info)
1910 {
1911 	switch (switch_info->name_type) {
1912 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1913 		/*
1914 		 * Name is not recognized, assume the master,
1915 		 * check the number of VFs key presence.
1916 		 */
1917 		switch_info->master = num_vf_set;
1918 		break;
1919 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1920 		/*
1921 		 * Name is not set, this assumes the legacy naming
1922 		 * schema for master, just check if there is a
1923 		 * number of VFs key.
1924 		 */
1925 		switch_info->master = num_vf_set;
1926 		break;
1927 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1928 		/* New uplink naming schema recognized. */
1929 		switch_info->master = 1;
1930 		break;
1931 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1932 		/* Legacy representors naming schema. */
1933 		switch_info->representor = !num_vf_set;
1934 		break;
1935 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1936 		/* New representors naming schema. */
1937 		switch_info->representor = 1;
1938 		break;
1939 	}
1940 }
1941 
1942 /**
1943  * Analyze gathered port parameters via sysfs to recognize master
1944  * and representor devices for E-Switch configuration.
1945  *
1946  * @param[in] device_dir
1947  *   flag of presence of "device" directory under port device key.
1948  * @param[inout] switch_info
1949  *   Port information, including port name as a number and port name
1950  *   type if recognized
1951  *
1952  * @return
1953  *   master and representor flags are set in switch_info according to
1954  *   recognized parameters (if any).
1955  */
1956 void
1957 mlx5_sysfs_check_switch_info(bool device_dir,
1958 			     struct mlx5_switch_info *switch_info)
1959 {
1960 	switch (switch_info->name_type) {
1961 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1962 		/*
1963 		 * Name is not recognized, assume the master,
1964 		 * check the device directory presence.
1965 		 */
1966 		switch_info->master = device_dir;
1967 		break;
1968 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1969 		/*
1970 		 * Name is not set, this assumes the legacy naming
1971 		 * schema for master, just check if there is
1972 		 * a device directory.
1973 		 */
1974 		switch_info->master = device_dir;
1975 		break;
1976 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1977 		/* New uplink naming schema recognized. */
1978 		switch_info->master = 1;
1979 		break;
1980 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1981 		/* Legacy representors naming schema. */
1982 		switch_info->representor = !device_dir;
1983 		break;
1984 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1985 		/* New representors naming schema. */
1986 		switch_info->representor = 1;
1987 		break;
1988 	}
1989 }
1990 
1991 /**
1992  * Extract port name, as a number, from sysfs or netlink information.
1993  *
1994  * @param[in] port_name_in
1995  *   String representing the port name.
1996  * @param[out] port_info_out
1997  *   Port information, including port name as a number and port name
1998  *   type if recognized
1999  *
2000  * @return
2001  *   port_name field set according to recognized name format.
2002  */
2003 void
2004 mlx5_translate_port_name(const char *port_name_in,
2005 			 struct mlx5_switch_info *port_info_out)
2006 {
2007 	char pf_c1, pf_c2, vf_c1, vf_c2;
2008 	char *end;
2009 	int sc_items;
2010 
2011 	/*
2012 	 * Check for port-name as a string of the form pf0vf0
2013 	 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
2014 	 */
2015 	sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
2016 			  &pf_c1, &pf_c2, &port_info_out->pf_num,
2017 			  &vf_c1, &vf_c2, &port_info_out->port_name);
2018 	if (sc_items == 6 &&
2019 	    pf_c1 == 'p' && pf_c2 == 'f' &&
2020 	    vf_c1 == 'v' && vf_c2 == 'f') {
2021 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
2022 		return;
2023 	}
2024 	/*
2025 	 * Check for port-name as a string of the form p0
2026 	 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
2027 	 */
2028 	sc_items = sscanf(port_name_in, "%c%d",
2029 			  &pf_c1, &port_info_out->port_name);
2030 	if (sc_items == 2 && pf_c1 == 'p') {
2031 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
2032 		return;
2033 	}
2034 	/* Check for port-name as a number (support kernel ver < 5.0 */
2035 	errno = 0;
2036 	port_info_out->port_name = strtol(port_name_in, &end, 0);
2037 	if (!errno &&
2038 	    (size_t)(end - port_name_in) == strlen(port_name_in)) {
2039 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
2040 		return;
2041 	}
2042 	port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
2043 	return;
2044 }
2045 
2046 /**
2047  * DPDK callback to retrieve plug-in module EEPROM information (type and size).
2048  *
2049  * @param dev
2050  *   Pointer to Ethernet device structure.
2051  * @param[out] modinfo
2052  *   Storage for plug-in module EEPROM information.
2053  *
2054  * @return
2055  *   0 on success, a negative errno value otherwise and rte_errno is set.
2056  */
2057 int
2058 mlx5_get_module_info(struct rte_eth_dev *dev,
2059 		     struct rte_eth_dev_module_info *modinfo)
2060 {
2061 	struct ethtool_modinfo info = {
2062 		.cmd = ETHTOOL_GMODULEINFO,
2063 	};
2064 	struct ifreq ifr = (struct ifreq) {
2065 		.ifr_data = (void *)&info,
2066 	};
2067 	int ret = 0;
2068 
2069 	if (!dev || !modinfo) {
2070 		DRV_LOG(WARNING, "missing argument, cannot get module info");
2071 		rte_errno = EINVAL;
2072 		return -rte_errno;
2073 	}
2074 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2075 	if (ret) {
2076 		DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2077 			dev->data->port_id, strerror(rte_errno));
2078 		return ret;
2079 	}
2080 	modinfo->type = info.type;
2081 	modinfo->eeprom_len = info.eeprom_len;
2082 	return ret;
2083 }
2084 
2085 /**
2086  * DPDK callback to retrieve plug-in module EEPROM data.
2087  *
2088  * @param dev
2089  *   Pointer to Ethernet device structure.
2090  * @param[out] info
2091  *   Storage for plug-in module EEPROM data.
2092  *
2093  * @return
2094  *   0 on success, a negative errno value otherwise and rte_errno is set.
2095  */
2096 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
2097 			   struct rte_dev_eeprom_info *info)
2098 {
2099 	struct ethtool_eeprom *eeprom;
2100 	struct ifreq ifr;
2101 	int ret = 0;
2102 
2103 	if (!dev || !info) {
2104 		DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
2105 		rte_errno = EINVAL;
2106 		return -rte_errno;
2107 	}
2108 	eeprom = rte_calloc(__func__, 1,
2109 			    (sizeof(struct ethtool_eeprom) + info->length), 0);
2110 	if (!eeprom) {
2111 		DRV_LOG(WARNING, "port %u cannot allocate memory for "
2112 			"eeprom data", dev->data->port_id);
2113 		rte_errno = ENOMEM;
2114 		return -rte_errno;
2115 	}
2116 	eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2117 	eeprom->offset = info->offset;
2118 	eeprom->len = info->length;
2119 	ifr = (struct ifreq) {
2120 		.ifr_data = (void *)eeprom,
2121 	};
2122 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2123 	if (ret)
2124 		DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2125 			dev->data->port_id, strerror(rte_errno));
2126 	else
2127 		rte_memcpy(info->data, eeprom->data, info->length);
2128 	rte_free(eeprom);
2129 	return ret;
2130 }
2131 
2132 /**
2133  * DPDK callback to retrieve hairpin capabilities.
2134  *
2135  * @param dev
2136  *   Pointer to Ethernet device structure.
2137  * @param[out] cap
2138  *   Storage for hairpin capability data.
2139  *
2140  * @return
2141  *   0 on success, a negative errno value otherwise and rte_errno is set.
2142  */
2143 int mlx5_hairpin_cap_get(struct rte_eth_dev *dev,
2144 			 struct rte_eth_hairpin_cap *cap)
2145 {
2146 	struct mlx5_priv *priv = dev->data->dev_private;
2147 
2148 	if (priv->sh->devx == 0) {
2149 		rte_errno = ENOTSUP;
2150 		return -rte_errno;
2151 	}
2152 	cap->max_nb_queues = UINT16_MAX;
2153 	cap->max_rx_2_tx = 1;
2154 	cap->max_tx_2_rx = 1;
2155 	cap->max_nb_desc = 8192;
2156 	return 0;
2157 }
2158