xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision e571ad5541457e5741d84babb7d2301f2cd18c2f)
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 <stdint.h>
11 #include <stdio.h>
12 #include <string.h>
13 #include <stdlib.h>
14 #include <errno.h>
15 #include <dirent.h>
16 #include <net/if.h>
17 #include <sys/ioctl.h>
18 #include <sys/socket.h>
19 #include <netinet/in.h>
20 #include <linux/ethtool.h>
21 #include <linux/sockios.h>
22 #include <fcntl.h>
23 #include <stdalign.h>
24 #include <sys/un.h>
25 #include <time.h>
26 
27 #include <rte_atomic.h>
28 #include <rte_ethdev_driver.h>
29 #include <rte_bus_pci.h>
30 #include <rte_mbuf.h>
31 #include <rte_common.h>
32 #include <rte_interrupts.h>
33 #include <rte_malloc.h>
34 #include <rte_string_fns.h>
35 #include <rte_rwlock.h>
36 
37 #include "mlx5.h"
38 #include "mlx5_glue.h"
39 #include "mlx5_rxtx.h"
40 #include "mlx5_utils.h"
41 
42 /* Supported speed values found in /usr/include/linux/ethtool.h */
43 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
44 #define SUPPORTED_40000baseKR4_Full (1 << 23)
45 #endif
46 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
47 #define SUPPORTED_40000baseCR4_Full (1 << 24)
48 #endif
49 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
50 #define SUPPORTED_40000baseSR4_Full (1 << 25)
51 #endif
52 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
53 #define SUPPORTED_40000baseLR4_Full (1 << 26)
54 #endif
55 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
56 #define SUPPORTED_56000baseKR4_Full (1 << 27)
57 #endif
58 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
59 #define SUPPORTED_56000baseCR4_Full (1 << 28)
60 #endif
61 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
62 #define SUPPORTED_56000baseSR4_Full (1 << 29)
63 #endif
64 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
65 #define SUPPORTED_56000baseLR4_Full (1 << 30)
66 #endif
67 
68 /* Add defines in case the running kernel is not the same as user headers. */
69 #ifndef ETHTOOL_GLINKSETTINGS
70 struct ethtool_link_settings {
71 	uint32_t cmd;
72 	uint32_t speed;
73 	uint8_t duplex;
74 	uint8_t port;
75 	uint8_t phy_address;
76 	uint8_t autoneg;
77 	uint8_t mdio_support;
78 	uint8_t eth_to_mdix;
79 	uint8_t eth_tp_mdix_ctrl;
80 	int8_t link_mode_masks_nwords;
81 	uint32_t reserved[8];
82 	uint32_t link_mode_masks[];
83 };
84 
85 #define ETHTOOL_GLINKSETTINGS 0x0000004c
86 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
87 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
88 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
89 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
90 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
91 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
92 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
93 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
94 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
95 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
96 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
97 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
98 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
99 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
100 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
101 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
102 #endif
103 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
104 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
105 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
106 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
107 #endif
108 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
109 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
110 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
111 #endif
112 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
113 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
114 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
115 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
116 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
117 #endif
118 
119 /**
120  * Get master interface name from private structure.
121  *
122  * @param[in] dev
123  *   Pointer to Ethernet device.
124  * @param[out] ifname
125  *   Interface name output buffer.
126  *
127  * @return
128  *   0 on success, a negative errno value otherwise and rte_errno is set.
129  */
130 int
131 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
132 {
133 	DIR *dir;
134 	struct dirent *dent;
135 	unsigned int dev_type = 0;
136 	unsigned int dev_port_prev = ~0u;
137 	char match[IF_NAMESIZE] = "";
138 
139 	assert(ibdev_path);
140 	{
141 		MKSTR(path, "%s/device/net", ibdev_path);
142 
143 		dir = opendir(path);
144 		if (dir == NULL) {
145 			rte_errno = errno;
146 			return -rte_errno;
147 		}
148 	}
149 	while ((dent = readdir(dir)) != NULL) {
150 		char *name = dent->d_name;
151 		FILE *file;
152 		unsigned int dev_port;
153 		int r;
154 
155 		if ((name[0] == '.') &&
156 		    ((name[1] == '\0') ||
157 		     ((name[1] == '.') && (name[2] == '\0'))))
158 			continue;
159 
160 		MKSTR(path, "%s/device/net/%s/%s",
161 		      ibdev_path, name,
162 		      (dev_type ? "dev_id" : "dev_port"));
163 
164 		file = fopen(path, "rb");
165 		if (file == NULL) {
166 			if (errno != ENOENT)
167 				continue;
168 			/*
169 			 * Switch to dev_id when dev_port does not exist as
170 			 * is the case with Linux kernel versions < 3.15.
171 			 */
172 try_dev_id:
173 			match[0] = '\0';
174 			if (dev_type)
175 				break;
176 			dev_type = 1;
177 			dev_port_prev = ~0u;
178 			rewinddir(dir);
179 			continue;
180 		}
181 		r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
182 		fclose(file);
183 		if (r != 1)
184 			continue;
185 		/*
186 		 * Switch to dev_id when dev_port returns the same value for
187 		 * all ports. May happen when using a MOFED release older than
188 		 * 3.0 with a Linux kernel >= 3.15.
189 		 */
190 		if (dev_port == dev_port_prev)
191 			goto try_dev_id;
192 		dev_port_prev = dev_port;
193 		if (dev_port == 0)
194 			strlcpy(match, name, sizeof(match));
195 	}
196 	closedir(dir);
197 	if (match[0] == '\0') {
198 		rte_errno = ENOENT;
199 		return -rte_errno;
200 	}
201 	strncpy(*ifname, match, sizeof(*ifname));
202 	return 0;
203 }
204 
205 /**
206  * Get interface name from private structure.
207  *
208  * This is a port representor-aware version of mlx5_get_master_ifname().
209  *
210  * @param[in] dev
211  *   Pointer to Ethernet device.
212  * @param[out] ifname
213  *   Interface name output buffer.
214  *
215  * @return
216  *   0 on success, a negative errno value otherwise and rte_errno is set.
217  */
218 int
219 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
220 {
221 	struct mlx5_priv *priv = dev->data->dev_private;
222 	unsigned int ifindex;
223 
224 	assert(priv);
225 	assert(priv->sh);
226 	ifindex = priv->nl_socket_rdma >= 0 ?
227 		  mlx5_nl_ifindex(priv->nl_socket_rdma,
228 				  priv->sh->ibdev_name,
229 				  priv->ibv_port) : 0;
230 	if (!ifindex) {
231 		if (!priv->representor)
232 			return mlx5_get_master_ifname(priv->sh->ibdev_path,
233 						      ifname);
234 		rte_errno = ENXIO;
235 		return -rte_errno;
236 	}
237 	if (if_indextoname(ifindex, &(*ifname)[0]))
238 		return 0;
239 	rte_errno = errno;
240 	return -rte_errno;
241 }
242 
243 /**
244  * Get the interface index from device name.
245  *
246  * @param[in] dev
247  *   Pointer to Ethernet device.
248  *
249  * @return
250  *   Nonzero interface index on success, zero otherwise and rte_errno is set.
251  */
252 unsigned int
253 mlx5_ifindex(const struct rte_eth_dev *dev)
254 {
255 	char ifname[IF_NAMESIZE];
256 	unsigned int ifindex;
257 
258 	if (mlx5_get_ifname(dev, &ifname))
259 		return 0;
260 	ifindex = if_nametoindex(ifname);
261 	if (!ifindex)
262 		rte_errno = errno;
263 	return ifindex;
264 }
265 
266 /**
267  * Perform ifreq ioctl() on associated Ethernet device.
268  *
269  * @param[in] dev
270  *   Pointer to Ethernet device.
271  * @param req
272  *   Request number to pass to ioctl().
273  * @param[out] ifr
274  *   Interface request structure output buffer.
275  *
276  * @return
277  *   0 on success, a negative errno value otherwise and rte_errno is set.
278  */
279 int
280 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
281 {
282 	int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
283 	int ret = 0;
284 
285 	if (sock == -1) {
286 		rte_errno = errno;
287 		return -rte_errno;
288 	}
289 	ret = mlx5_get_ifname(dev, &ifr->ifr_name);
290 	if (ret)
291 		goto error;
292 	ret = ioctl(sock, req, ifr);
293 	if (ret == -1) {
294 		rte_errno = errno;
295 		goto error;
296 	}
297 	close(sock);
298 	return 0;
299 error:
300 	close(sock);
301 	return -rte_errno;
302 }
303 
304 /**
305  * Get device MTU.
306  *
307  * @param dev
308  *   Pointer to Ethernet device.
309  * @param[out] mtu
310  *   MTU value output buffer.
311  *
312  * @return
313  *   0 on success, a negative errno value otherwise and rte_errno is set.
314  */
315 int
316 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
317 {
318 	struct ifreq request;
319 	int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
320 
321 	if (ret)
322 		return ret;
323 	*mtu = request.ifr_mtu;
324 	return 0;
325 }
326 
327 /**
328  * Set device MTU.
329  *
330  * @param dev
331  *   Pointer to Ethernet device.
332  * @param mtu
333  *   MTU value to set.
334  *
335  * @return
336  *   0 on success, a negative errno value otherwise and rte_errno is set.
337  */
338 static int
339 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
340 {
341 	struct ifreq request = { .ifr_mtu = mtu, };
342 
343 	return mlx5_ifreq(dev, SIOCSIFMTU, &request);
344 }
345 
346 /**
347  * Set device flags.
348  *
349  * @param dev
350  *   Pointer to Ethernet device.
351  * @param keep
352  *   Bitmask for flags that must remain untouched.
353  * @param flags
354  *   Bitmask for flags to modify.
355  *
356  * @return
357  *   0 on success, a negative errno value otherwise and rte_errno is set.
358  */
359 int
360 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
361 {
362 	struct ifreq request;
363 	int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
364 
365 	if (ret)
366 		return ret;
367 	request.ifr_flags &= keep;
368 	request.ifr_flags |= flags & ~keep;
369 	return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
370 }
371 
372 /**
373  * DPDK callback for Ethernet device configuration.
374  *
375  * @param dev
376  *   Pointer to Ethernet device structure.
377  *
378  * @return
379  *   0 on success, a negative errno value otherwise and rte_errno is set.
380  */
381 int
382 mlx5_dev_configure(struct rte_eth_dev *dev)
383 {
384 	struct mlx5_priv *priv = dev->data->dev_private;
385 	unsigned int rxqs_n = dev->data->nb_rx_queues;
386 	unsigned int txqs_n = dev->data->nb_tx_queues;
387 	unsigned int i;
388 	unsigned int j;
389 	unsigned int reta_idx_n;
390 	const uint8_t use_app_rss_key =
391 		!!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
392 	int ret = 0;
393 
394 	if (use_app_rss_key &&
395 	    (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
396 	     MLX5_RSS_HASH_KEY_LEN)) {
397 		DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
398 			dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
399 		rte_errno = EINVAL;
400 		return -rte_errno;
401 	}
402 	priv->rss_conf.rss_key =
403 		rte_realloc(priv->rss_conf.rss_key,
404 			    MLX5_RSS_HASH_KEY_LEN, 0);
405 	if (!priv->rss_conf.rss_key) {
406 		DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
407 			dev->data->port_id, rxqs_n);
408 		rte_errno = ENOMEM;
409 		return -rte_errno;
410 	}
411 	memcpy(priv->rss_conf.rss_key,
412 	       use_app_rss_key ?
413 	       dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
414 	       rss_hash_default_key,
415 	       MLX5_RSS_HASH_KEY_LEN);
416 	priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
417 	priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
418 	priv->rxqs = (void *)dev->data->rx_queues;
419 	priv->txqs = (void *)dev->data->tx_queues;
420 	if (txqs_n != priv->txqs_n) {
421 		DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
422 			dev->data->port_id, priv->txqs_n, txqs_n);
423 		priv->txqs_n = txqs_n;
424 	}
425 	if (rxqs_n > priv->config.ind_table_max_size) {
426 		DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
427 			dev->data->port_id, rxqs_n);
428 		rte_errno = EINVAL;
429 		return -rte_errno;
430 	}
431 	if (rxqs_n != priv->rxqs_n) {
432 		DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
433 			dev->data->port_id, priv->rxqs_n, rxqs_n);
434 		priv->rxqs_n = rxqs_n;
435 		/*
436 		 * If the requested number of RX queues is not a power of two,
437 		 * use the maximum indirection table size for better balancing.
438 		 * The result is always rounded to the next power of two.
439 		 */
440 		reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
441 					     priv->config.ind_table_max_size :
442 					     rxqs_n));
443 		ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
444 		if (ret)
445 			return ret;
446 		/*
447 		 * When the number of RX queues is not a power of two,
448 		 * the remaining table entries are padded with reused WQs
449 		 * and hashes are not spread uniformly.
450 		 */
451 		for (i = 0, j = 0; (i != reta_idx_n); ++i) {
452 			(*priv->reta_idx)[i] = j;
453 			if (++j == rxqs_n)
454 				j = 0;
455 		}
456 	}
457 	ret = mlx5_proc_priv_init(dev);
458 	if (ret)
459 		return ret;
460 	return 0;
461 }
462 
463 /**
464  * Sets default tuning parameters.
465  *
466  * @param dev
467  *   Pointer to Ethernet device.
468  * @param[out] info
469  *   Info structure output buffer.
470  */
471 static void
472 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
473 {
474 	struct mlx5_priv *priv = dev->data->dev_private;
475 
476 	/* Minimum CPU utilization. */
477 	info->default_rxportconf.ring_size = 256;
478 	info->default_txportconf.ring_size = 256;
479 	info->default_rxportconf.burst_size = 64;
480 	info->default_txportconf.burst_size = 64;
481 	if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
482 		info->default_rxportconf.nb_queues = 16;
483 		info->default_txportconf.nb_queues = 16;
484 		if (dev->data->nb_rx_queues > 2 ||
485 		    dev->data->nb_tx_queues > 2) {
486 			/* Max Throughput. */
487 			info->default_rxportconf.ring_size = 2048;
488 			info->default_txportconf.ring_size = 2048;
489 		}
490 	} else {
491 		info->default_rxportconf.nb_queues = 8;
492 		info->default_txportconf.nb_queues = 8;
493 		if (dev->data->nb_rx_queues > 2 ||
494 		    dev->data->nb_tx_queues > 2) {
495 			/* Max Throughput. */
496 			info->default_rxportconf.ring_size = 4096;
497 			info->default_txportconf.ring_size = 4096;
498 		}
499 	}
500 }
501 
502 /**
503  * DPDK callback to get information about the device.
504  *
505  * @param dev
506  *   Pointer to Ethernet device structure.
507  * @param[out] info
508  *   Info structure output buffer.
509  */
510 void
511 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
512 {
513 	struct mlx5_priv *priv = dev->data->dev_private;
514 	struct mlx5_dev_config *config = &priv->config;
515 	unsigned int max;
516 	char ifname[IF_NAMESIZE];
517 
518 	/* FIXME: we should ask the device for these values. */
519 	info->min_rx_bufsize = 32;
520 	info->max_rx_pktlen = 65536;
521 	/*
522 	 * Since we need one CQ per QP, the limit is the minimum number
523 	 * between the two values.
524 	 */
525 	max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
526 		      priv->sh->device_attr.orig_attr.max_qp);
527 	/* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
528 	if (max >= 65535)
529 		max = 65535;
530 	info->max_rx_queues = max;
531 	info->max_tx_queues = max;
532 	info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
533 	info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
534 	info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
535 				 info->rx_queue_offload_capa);
536 	info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
537 	if (mlx5_get_ifname(dev, &ifname) == 0)
538 		info->if_index = if_nametoindex(ifname);
539 	info->reta_size = priv->reta_idx_n ?
540 		priv->reta_idx_n : config->ind_table_max_size;
541 	info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
542 	info->speed_capa = priv->link_speed_capa;
543 	info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
544 	mlx5_set_default_params(dev, info);
545 	info->switch_info.name = dev->data->name;
546 	info->switch_info.domain_id = priv->domain_id;
547 	info->switch_info.port_id = priv->representor_id;
548 	if (priv->representor) {
549 		unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
550 		uint16_t port_id[i];
551 
552 		i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
553 		while (i--) {
554 			struct mlx5_priv *opriv =
555 				rte_eth_devices[port_id[i]].data->dev_private;
556 
557 			if (!opriv ||
558 			    opriv->representor ||
559 			    opriv->domain_id != priv->domain_id)
560 				continue;
561 			/*
562 			 * Override switch name with that of the master
563 			 * device.
564 			 */
565 			info->switch_info.name = opriv->dev_data->name;
566 			break;
567 		}
568 	}
569 }
570 
571 /**
572  * Get device current raw clock counter
573  *
574  * @param dev
575  *   Pointer to Ethernet device structure.
576  * @param[out] time
577  *   Current raw clock counter of the device.
578  *
579  * @return
580  *   0 if the clock has correctly been read
581  *   The value of errno in case of error
582  */
583 int
584 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
585 {
586 	struct mlx5_priv *priv = dev->data->dev_private;
587 	struct ibv_context *ctx = priv->sh->ctx;
588 	struct ibv_values_ex values;
589 	int err = 0;
590 
591 	values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
592 	err = mlx5_glue->query_rt_values_ex(ctx, &values);
593 	if (err != 0) {
594 		DRV_LOG(WARNING, "Could not query the clock !");
595 		return err;
596 	}
597 	*clock = values.raw_clock.tv_nsec;
598 	return 0;
599 }
600 
601 /**
602  * Get firmware version of a device.
603  *
604  * @param dev
605  *   Ethernet device port.
606  * @param fw_ver
607  *   String output allocated by caller.
608  * @param fw_size
609  *   Size of the output string, including terminating null byte.
610  *
611  * @return
612  *   0 on success, or the size of the non truncated string if too big.
613  */
614 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
615 {
616 	struct mlx5_priv *priv = dev->data->dev_private;
617 	struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
618 	size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
619 
620 	if (fw_size < size)
621 		return size;
622 	if (fw_ver != NULL)
623 		strlcpy(fw_ver, attr->fw_ver, fw_size);
624 	return 0;
625 }
626 
627 /**
628  * Get supported packet types.
629  *
630  * @param dev
631  *   Pointer to Ethernet device structure.
632  *
633  * @return
634  *   A pointer to the supported Packet types array.
635  */
636 const uint32_t *
637 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
638 {
639 	static const uint32_t ptypes[] = {
640 		/* refers to rxq_cq_to_pkt_type() */
641 		RTE_PTYPE_L2_ETHER,
642 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
643 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
644 		RTE_PTYPE_L4_NONFRAG,
645 		RTE_PTYPE_L4_FRAG,
646 		RTE_PTYPE_L4_TCP,
647 		RTE_PTYPE_L4_UDP,
648 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
649 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
650 		RTE_PTYPE_INNER_L4_NONFRAG,
651 		RTE_PTYPE_INNER_L4_FRAG,
652 		RTE_PTYPE_INNER_L4_TCP,
653 		RTE_PTYPE_INNER_L4_UDP,
654 		RTE_PTYPE_UNKNOWN
655 	};
656 
657 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
658 	    dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
659 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
660 		return ptypes;
661 	return NULL;
662 }
663 
664 /**
665  * Retrieve the master device for representor in the same switch domain.
666  *
667  * @param dev
668  *   Pointer to representor Ethernet device structure.
669  *
670  * @return
671  *   Master device structure  on success, NULL otherwise.
672  */
673 
674 static struct rte_eth_dev *
675 mlx5_find_master_dev(struct rte_eth_dev *dev)
676 {
677 	struct mlx5_priv *priv;
678 	uint16_t port_id;
679 	uint16_t domain_id;
680 
681 	priv = dev->data->dev_private;
682 	domain_id = priv->domain_id;
683 	assert(priv->representor);
684 	RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
685 		priv = rte_eth_devices[port_id].data->dev_private;
686 		if (priv &&
687 		    priv->master &&
688 		    priv->domain_id == domain_id)
689 			return &rte_eth_devices[port_id];
690 	}
691 	return NULL;
692 }
693 
694 /**
695  * DPDK callback to retrieve physical link information.
696  *
697  * @param dev
698  *   Pointer to Ethernet device structure.
699  * @param[out] link
700  *   Storage for current link status.
701  *
702  * @return
703  *   0 on success, a negative errno value otherwise and rte_errno is set.
704  */
705 static int
706 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
707 			       struct rte_eth_link *link)
708 {
709 	struct mlx5_priv *priv = dev->data->dev_private;
710 	struct ethtool_cmd edata = {
711 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
712 	};
713 	struct ifreq ifr;
714 	struct rte_eth_link dev_link;
715 	int link_speed = 0;
716 	int ret;
717 
718 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
719 	if (ret) {
720 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
721 			dev->data->port_id, strerror(rte_errno));
722 		return ret;
723 	}
724 	dev_link = (struct rte_eth_link) {
725 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
726 				(ifr.ifr_flags & IFF_RUNNING)),
727 	};
728 	ifr = (struct ifreq) {
729 		.ifr_data = (void *)&edata,
730 	};
731 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
732 	if (ret) {
733 		if (ret == -ENOTSUP && priv->representor) {
734 			struct rte_eth_dev *master;
735 
736 			/*
737 			 * For representors we can try to inherit link
738 			 * settings from the master device. Actually
739 			 * link settings do not make a lot of sense
740 			 * for representors due to missing physical
741 			 * link. The old kernel drivers supported
742 			 * emulated settings query for representors,
743 			 * the new ones do not, so we have to add
744 			 * this code for compatibility issues.
745 			 */
746 			master = mlx5_find_master_dev(dev);
747 			if (master) {
748 				ifr = (struct ifreq) {
749 					.ifr_data = (void *)&edata,
750 				};
751 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
752 			}
753 		}
754 		if (ret) {
755 			DRV_LOG(WARNING,
756 				"port %u ioctl(SIOCETHTOOL,"
757 				" ETHTOOL_GSET) failed: %s",
758 				dev->data->port_id, strerror(rte_errno));
759 			return ret;
760 		}
761 	}
762 	link_speed = ethtool_cmd_speed(&edata);
763 	if (link_speed == -1)
764 		dev_link.link_speed = ETH_SPEED_NUM_NONE;
765 	else
766 		dev_link.link_speed = link_speed;
767 	priv->link_speed_capa = 0;
768 	if (edata.supported & SUPPORTED_Autoneg)
769 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
770 	if (edata.supported & (SUPPORTED_1000baseT_Full |
771 			       SUPPORTED_1000baseKX_Full))
772 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
773 	if (edata.supported & SUPPORTED_10000baseKR_Full)
774 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
775 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
776 			       SUPPORTED_40000baseCR4_Full |
777 			       SUPPORTED_40000baseSR4_Full |
778 			       SUPPORTED_40000baseLR4_Full))
779 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
780 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
781 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
782 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
783 			ETH_LINK_SPEED_FIXED);
784 	if (((dev_link.link_speed && !dev_link.link_status) ||
785 	     (!dev_link.link_speed && dev_link.link_status))) {
786 		rte_errno = EAGAIN;
787 		return -rte_errno;
788 	}
789 	*link = dev_link;
790 	return 0;
791 }
792 
793 /**
794  * Retrieve physical link information (unlocked version using new ioctl).
795  *
796  * @param dev
797  *   Pointer to Ethernet device structure.
798  * @param[out] link
799  *   Storage for current link status.
800  *
801  * @return
802  *   0 on success, a negative errno value otherwise and rte_errno is set.
803  */
804 static int
805 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
806 			     struct rte_eth_link *link)
807 
808 {
809 	struct mlx5_priv *priv = dev->data->dev_private;
810 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
811 	struct ifreq ifr;
812 	struct rte_eth_link dev_link;
813 	struct rte_eth_dev *master = NULL;
814 	uint64_t sc;
815 	int ret;
816 
817 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
818 	if (ret) {
819 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
820 			dev->data->port_id, strerror(rte_errno));
821 		return ret;
822 	}
823 	dev_link = (struct rte_eth_link) {
824 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
825 				(ifr.ifr_flags & IFF_RUNNING)),
826 	};
827 	ifr = (struct ifreq) {
828 		.ifr_data = (void *)&gcmd,
829 	};
830 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
831 	if (ret) {
832 		if (ret == -ENOTSUP && priv->representor) {
833 			/*
834 			 * For representors we can try to inherit link
835 			 * settings from the master device. Actually
836 			 * link settings do not make a lot of sense
837 			 * for representors due to missing physical
838 			 * link. The old kernel drivers supported
839 			 * emulated settings query for representors,
840 			 * the new ones do not, so we have to add
841 			 * this code for compatibility issues.
842 			 */
843 			master = mlx5_find_master_dev(dev);
844 			if (master) {
845 				ifr = (struct ifreq) {
846 					.ifr_data = (void *)&gcmd,
847 				};
848 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
849 			}
850 		}
851 		if (ret) {
852 			DRV_LOG(DEBUG,
853 				"port %u ioctl(SIOCETHTOOL,"
854 				" ETHTOOL_GLINKSETTINGS) failed: %s",
855 				dev->data->port_id, strerror(rte_errno));
856 			return ret;
857 		}
858 
859 	}
860 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
861 
862 	alignas(struct ethtool_link_settings)
863 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
864 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
865 	struct ethtool_link_settings *ecmd = (void *)data;
866 
867 	*ecmd = gcmd;
868 	ifr.ifr_data = (void *)ecmd;
869 	ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
870 	if (ret) {
871 		DRV_LOG(DEBUG,
872 			"port %u ioctl(SIOCETHTOOL,"
873 			"ETHTOOL_GLINKSETTINGS) failed: %s",
874 			dev->data->port_id, strerror(rte_errno));
875 		return ret;
876 	}
877 	dev_link.link_speed = ecmd->speed;
878 	sc = ecmd->link_mode_masks[0] |
879 		((uint64_t)ecmd->link_mode_masks[1] << 32);
880 	priv->link_speed_capa = 0;
881 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
882 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
883 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
884 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
885 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
886 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
887 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
888 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
889 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
890 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
891 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
892 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
893 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
894 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
895 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
896 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
897 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
898 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
899 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
900 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
901 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
902 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
903 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
904 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
905 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
906 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
907 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
908 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
909 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
910 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
911 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
912 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
913 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
914 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
915 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
916 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
917 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
918 				  ETH_LINK_SPEED_FIXED);
919 	if (((dev_link.link_speed && !dev_link.link_status) ||
920 	     (!dev_link.link_speed && dev_link.link_status))) {
921 		rte_errno = EAGAIN;
922 		return -rte_errno;
923 	}
924 	*link = dev_link;
925 	return 0;
926 }
927 
928 /**
929  * DPDK callback to retrieve physical link information.
930  *
931  * @param dev
932  *   Pointer to Ethernet device structure.
933  * @param wait_to_complete
934  *   Wait for request completion.
935  *
936  * @return
937  *   0 if link status was not updated, positive if it was, a negative errno
938  *   value otherwise and rte_errno is set.
939  */
940 int
941 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
942 {
943 	int ret;
944 	struct rte_eth_link dev_link;
945 	time_t start_time = time(NULL);
946 
947 	do {
948 		ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
949 		if (ret)
950 			ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
951 		if (ret == 0)
952 			break;
953 		/* Handle wait to complete situation. */
954 		if (wait_to_complete && ret == -EAGAIN) {
955 			if (abs((int)difftime(time(NULL), start_time)) <
956 			    MLX5_LINK_STATUS_TIMEOUT) {
957 				usleep(0);
958 				continue;
959 			} else {
960 				rte_errno = EBUSY;
961 				return -rte_errno;
962 			}
963 		} else if (ret < 0) {
964 			return ret;
965 		}
966 	} while (wait_to_complete);
967 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
968 		       sizeof(struct rte_eth_link));
969 	dev->data->dev_link = dev_link;
970 	return ret;
971 }
972 
973 /**
974  * DPDK callback to change the MTU.
975  *
976  * @param dev
977  *   Pointer to Ethernet device structure.
978  * @param in_mtu
979  *   New MTU.
980  *
981  * @return
982  *   0 on success, a negative errno value otherwise and rte_errno is set.
983  */
984 int
985 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
986 {
987 	struct mlx5_priv *priv = dev->data->dev_private;
988 	uint16_t kern_mtu = 0;
989 	int ret;
990 
991 	ret = mlx5_get_mtu(dev, &kern_mtu);
992 	if (ret)
993 		return ret;
994 	/* Set kernel interface MTU first. */
995 	ret = mlx5_set_mtu(dev, mtu);
996 	if (ret)
997 		return ret;
998 	ret = mlx5_get_mtu(dev, &kern_mtu);
999 	if (ret)
1000 		return ret;
1001 	if (kern_mtu == mtu) {
1002 		priv->mtu = mtu;
1003 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1004 			dev->data->port_id, mtu);
1005 		return 0;
1006 	}
1007 	rte_errno = EAGAIN;
1008 	return -rte_errno;
1009 }
1010 
1011 /**
1012  * DPDK callback to get flow control status.
1013  *
1014  * @param dev
1015  *   Pointer to Ethernet device structure.
1016  * @param[out] fc_conf
1017  *   Flow control output buffer.
1018  *
1019  * @return
1020  *   0 on success, a negative errno value otherwise and rte_errno is set.
1021  */
1022 int
1023 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1024 {
1025 	struct ifreq ifr;
1026 	struct ethtool_pauseparam ethpause = {
1027 		.cmd = ETHTOOL_GPAUSEPARAM
1028 	};
1029 	int ret;
1030 
1031 	ifr.ifr_data = (void *)&ethpause;
1032 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1033 	if (ret) {
1034 		DRV_LOG(WARNING,
1035 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1036 			" %s",
1037 			dev->data->port_id, strerror(rte_errno));
1038 		return ret;
1039 	}
1040 	fc_conf->autoneg = ethpause.autoneg;
1041 	if (ethpause.rx_pause && ethpause.tx_pause)
1042 		fc_conf->mode = RTE_FC_FULL;
1043 	else if (ethpause.rx_pause)
1044 		fc_conf->mode = RTE_FC_RX_PAUSE;
1045 	else if (ethpause.tx_pause)
1046 		fc_conf->mode = RTE_FC_TX_PAUSE;
1047 	else
1048 		fc_conf->mode = RTE_FC_NONE;
1049 	return 0;
1050 }
1051 
1052 /**
1053  * DPDK callback to modify flow control parameters.
1054  *
1055  * @param dev
1056  *   Pointer to Ethernet device structure.
1057  * @param[in] fc_conf
1058  *   Flow control parameters.
1059  *
1060  * @return
1061  *   0 on success, a negative errno value otherwise and rte_errno is set.
1062  */
1063 int
1064 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1065 {
1066 	struct ifreq ifr;
1067 	struct ethtool_pauseparam ethpause = {
1068 		.cmd = ETHTOOL_SPAUSEPARAM
1069 	};
1070 	int ret;
1071 
1072 	ifr.ifr_data = (void *)&ethpause;
1073 	ethpause.autoneg = fc_conf->autoneg;
1074 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1075 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1076 		ethpause.rx_pause = 1;
1077 	else
1078 		ethpause.rx_pause = 0;
1079 
1080 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1081 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1082 		ethpause.tx_pause = 1;
1083 	else
1084 		ethpause.tx_pause = 0;
1085 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1086 	if (ret) {
1087 		DRV_LOG(WARNING,
1088 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1089 			" failed: %s",
1090 			dev->data->port_id, strerror(rte_errno));
1091 		return ret;
1092 	}
1093 	return 0;
1094 }
1095 
1096 /**
1097  * Get PCI information from struct ibv_device.
1098  *
1099  * @param device
1100  *   Pointer to Ethernet device structure.
1101  * @param[out] pci_addr
1102  *   PCI bus address output buffer.
1103  *
1104  * @return
1105  *   0 on success, a negative errno value otherwise and rte_errno is set.
1106  */
1107 int
1108 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1109 			    struct rte_pci_addr *pci_addr)
1110 {
1111 	FILE *file;
1112 	char line[32];
1113 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
1114 
1115 	file = fopen(path, "rb");
1116 	if (file == NULL) {
1117 		rte_errno = errno;
1118 		return -rte_errno;
1119 	}
1120 	while (fgets(line, sizeof(line), file) == line) {
1121 		size_t len = strlen(line);
1122 		int ret;
1123 
1124 		/* Truncate long lines. */
1125 		if (len == (sizeof(line) - 1))
1126 			while (line[(len - 1)] != '\n') {
1127 				ret = fgetc(file);
1128 				if (ret == EOF)
1129 					break;
1130 				line[(len - 1)] = ret;
1131 			}
1132 		/* Extract information. */
1133 		if (sscanf(line,
1134 			   "PCI_SLOT_NAME="
1135 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1136 			   &pci_addr->domain,
1137 			   &pci_addr->bus,
1138 			   &pci_addr->devid,
1139 			   &pci_addr->function) == 4) {
1140 			ret = 0;
1141 			break;
1142 		}
1143 	}
1144 	fclose(file);
1145 	return 0;
1146 }
1147 
1148 /**
1149  * Handle asynchronous removal event for entire multiport device.
1150  *
1151  * @param sh
1152  *   Infiniband device shared context.
1153  */
1154 static void
1155 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1156 {
1157 	uint32_t i;
1158 
1159 	for (i = 0; i < sh->max_port; ++i) {
1160 		struct rte_eth_dev *dev;
1161 
1162 		if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1163 			/*
1164 			 * Or not existing port either no
1165 			 * handler installed for this port.
1166 			 */
1167 			continue;
1168 		}
1169 		dev = &rte_eth_devices[sh->port[i].ih_port_id];
1170 		assert(dev);
1171 		if (dev->data->dev_conf.intr_conf.rmv)
1172 			_rte_eth_dev_callback_process
1173 				(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1174 	}
1175 }
1176 
1177 /**
1178  * Handle shared asynchronous events the NIC (removal event
1179  * and link status change). Supports multiport IB device.
1180  *
1181  * @param cb_arg
1182  *   Callback argument.
1183  */
1184 void
1185 mlx5_dev_interrupt_handler(void *cb_arg)
1186 {
1187 	struct mlx5_ibv_shared *sh = cb_arg;
1188 	struct ibv_async_event event;
1189 
1190 	/* Read all message from the IB device and acknowledge them. */
1191 	for (;;) {
1192 		struct rte_eth_dev *dev;
1193 		uint32_t tmp;
1194 
1195 		if (mlx5_glue->get_async_event(sh->ctx, &event))
1196 			break;
1197 		/* Retrieve and check IB port index. */
1198 		tmp = (uint32_t)event.element.port_num;
1199 		if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1200 			/*
1201 			 * The DEVICE_FATAL event is called once for
1202 			 * entire device without port specifying.
1203 			 * We should notify all existing ports.
1204 			 */
1205 			mlx5_glue->ack_async_event(&event);
1206 			mlx5_dev_interrupt_device_fatal(sh);
1207 			continue;
1208 		}
1209 		assert(tmp && (tmp <= sh->max_port));
1210 		if (!tmp) {
1211 			/* Unsupported devive level event. */
1212 			mlx5_glue->ack_async_event(&event);
1213 			DRV_LOG(DEBUG,
1214 				"unsupported common event (type %d)",
1215 				event.event_type);
1216 			continue;
1217 		}
1218 		if (tmp > sh->max_port) {
1219 			/* Invalid IB port index. */
1220 			mlx5_glue->ack_async_event(&event);
1221 			DRV_LOG(DEBUG,
1222 				"cannot handle an event (type %d)"
1223 				"due to invalid IB port index (%u)",
1224 				event.event_type, tmp);
1225 			continue;
1226 		}
1227 		if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1228 			/* No handler installed. */
1229 			mlx5_glue->ack_async_event(&event);
1230 			DRV_LOG(DEBUG,
1231 				"cannot handle an event (type %d)"
1232 				"due to no handler installed for port %u",
1233 				event.event_type, tmp);
1234 			continue;
1235 		}
1236 		/* Retrieve ethernet device descriptor. */
1237 		tmp = sh->port[tmp - 1].ih_port_id;
1238 		dev = &rte_eth_devices[tmp];
1239 		assert(dev);
1240 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1241 		     event.event_type == IBV_EVENT_PORT_ERR) &&
1242 			dev->data->dev_conf.intr_conf.lsc) {
1243 			mlx5_glue->ack_async_event(&event);
1244 			if (mlx5_link_update(dev, 0) == -EAGAIN) {
1245 				usleep(0);
1246 				continue;
1247 			}
1248 			_rte_eth_dev_callback_process
1249 				(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1250 			continue;
1251 		}
1252 		DRV_LOG(DEBUG,
1253 			"port %u cannot handle an unknown event (type %d)",
1254 			dev->data->port_id, event.event_type);
1255 		mlx5_glue->ack_async_event(&event);
1256 	}
1257 }
1258 
1259 /**
1260  * Uninstall shared asynchronous device events handler.
1261  * This function is implemeted to support event sharing
1262  * between multiple ports of single IB device.
1263  *
1264  * @param dev
1265  *   Pointer to Ethernet device.
1266  */
1267 static void
1268 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1269 {
1270 	struct mlx5_priv *priv = dev->data->dev_private;
1271 	struct mlx5_ibv_shared *sh = priv->sh;
1272 
1273 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1274 		return;
1275 	pthread_mutex_lock(&sh->intr_mutex);
1276 	assert(priv->ibv_port);
1277 	assert(priv->ibv_port <= sh->max_port);
1278 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1279 	if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1280 		goto exit;
1281 	assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1282 					(uint32_t)dev->data->port_id);
1283 	assert(sh->intr_cnt);
1284 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1285 	if (!sh->intr_cnt || --sh->intr_cnt)
1286 		goto exit;
1287 	rte_intr_callback_unregister(&sh->intr_handle,
1288 				     mlx5_dev_interrupt_handler, sh);
1289 	sh->intr_handle.fd = 0;
1290 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1291 exit:
1292 	pthread_mutex_unlock(&sh->intr_mutex);
1293 }
1294 
1295 /**
1296  * Install shared asyncronous device events handler.
1297  * This function is implemeted to support event sharing
1298  * between multiple ports of single IB device.
1299  *
1300  * @param dev
1301  *   Pointer to Ethernet device.
1302  */
1303 static void
1304 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1305 {
1306 	struct mlx5_priv *priv = dev->data->dev_private;
1307 	struct mlx5_ibv_shared *sh = priv->sh;
1308 	int ret;
1309 	int flags;
1310 
1311 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1312 		return;
1313 	pthread_mutex_lock(&sh->intr_mutex);
1314 	assert(priv->ibv_port);
1315 	assert(priv->ibv_port <= sh->max_port);
1316 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1317 	if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1318 		/* The handler is already installed for this port. */
1319 		assert(sh->intr_cnt);
1320 		goto exit;
1321 	}
1322 	sh->port[priv->ibv_port - 1].ih_port_id = (uint32_t)dev->data->port_id;
1323 	if (sh->intr_cnt) {
1324 		sh->intr_cnt++;
1325 		goto exit;
1326 	}
1327 	/* No shared handler installed. */
1328 	assert(sh->ctx->async_fd > 0);
1329 	flags = fcntl(sh->ctx->async_fd, F_GETFL);
1330 	ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1331 	if (ret) {
1332 		DRV_LOG(INFO, "failed to change file descriptor"
1333 			      " async event queue");
1334 		/* Indicate there will be no interrupts. */
1335 		dev->data->dev_conf.intr_conf.lsc = 0;
1336 		dev->data->dev_conf.intr_conf.rmv = 0;
1337 		sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1338 		goto exit;
1339 	}
1340 	sh->intr_handle.fd = sh->ctx->async_fd;
1341 	sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1342 	rte_intr_callback_register(&sh->intr_handle,
1343 				   mlx5_dev_interrupt_handler, sh);
1344 	sh->intr_cnt++;
1345 exit:
1346 	pthread_mutex_unlock(&sh->intr_mutex);
1347 }
1348 
1349 /**
1350  * Uninstall interrupt handler.
1351  *
1352  * @param dev
1353  *   Pointer to Ethernet device.
1354  */
1355 void
1356 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1357 {
1358 	mlx5_dev_shared_handler_uninstall(dev);
1359 }
1360 
1361 /**
1362  * Install interrupt handler.
1363  *
1364  * @param dev
1365  *   Pointer to Ethernet device.
1366  */
1367 void
1368 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1369 {
1370 	mlx5_dev_shared_handler_install(dev);
1371 }
1372 
1373 /**
1374  * DPDK callback to bring the link DOWN.
1375  *
1376  * @param dev
1377  *   Pointer to Ethernet device structure.
1378  *
1379  * @return
1380  *   0 on success, a negative errno value otherwise and rte_errno is set.
1381  */
1382 int
1383 mlx5_set_link_down(struct rte_eth_dev *dev)
1384 {
1385 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1386 }
1387 
1388 /**
1389  * DPDK callback to bring the link UP.
1390  *
1391  * @param dev
1392  *   Pointer to Ethernet device structure.
1393  *
1394  * @return
1395  *   0 on success, a negative errno value otherwise and rte_errno is set.
1396  */
1397 int
1398 mlx5_set_link_up(struct rte_eth_dev *dev)
1399 {
1400 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1401 }
1402 
1403 /**
1404  * Configure the TX function to use.
1405  *
1406  * @param dev
1407  *   Pointer to private data structure.
1408  *
1409  * @return
1410  *   Pointer to selected Tx burst function.
1411  */
1412 eth_tx_burst_t
1413 mlx5_select_tx_function(struct rte_eth_dev *dev)
1414 {
1415 	struct mlx5_priv *priv = dev->data->dev_private;
1416 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1417 	struct mlx5_dev_config *config = &priv->config;
1418 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1419 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1420 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1421 				    DEV_TX_OFFLOAD_GRE_TNL_TSO |
1422 				    DEV_TX_OFFLOAD_IP_TNL_TSO |
1423 				    DEV_TX_OFFLOAD_UDP_TNL_TSO));
1424 	int swp = !!(tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
1425 				    DEV_TX_OFFLOAD_UDP_TNL_TSO |
1426 				    DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM));
1427 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1428 
1429 	assert(priv != NULL);
1430 	/* Select appropriate TX function. */
1431 	if (vlan_insert || tso || swp)
1432 		return tx_pkt_burst;
1433 	if (config->mps == MLX5_MPW_ENHANCED) {
1434 		if (mlx5_check_vec_tx_support(dev) > 0) {
1435 			if (mlx5_check_raw_vec_tx_support(dev) > 0)
1436 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1437 			else
1438 				tx_pkt_burst = mlx5_tx_burst_vec;
1439 			DRV_LOG(DEBUG,
1440 				"port %u selected enhanced MPW Tx vectorized"
1441 				" function",
1442 				dev->data->port_id);
1443 		} else {
1444 			tx_pkt_burst = mlx5_tx_burst_empw;
1445 			DRV_LOG(DEBUG,
1446 				"port %u selected enhanced MPW Tx function",
1447 				dev->data->port_id);
1448 		}
1449 	} else if (config->mps && (config->txq_inline > 0)) {
1450 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1451 		DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1452 			dev->data->port_id);
1453 	} else if (config->mps) {
1454 		tx_pkt_burst = mlx5_tx_burst_mpw;
1455 		DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1456 			dev->data->port_id);
1457 	}
1458 	return tx_pkt_burst;
1459 }
1460 
1461 /**
1462  * Configure the RX function to use.
1463  *
1464  * @param dev
1465  *   Pointer to private data structure.
1466  *
1467  * @return
1468  *   Pointer to selected Rx burst function.
1469  */
1470 eth_rx_burst_t
1471 mlx5_select_rx_function(struct rte_eth_dev *dev)
1472 {
1473 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1474 
1475 	assert(dev != NULL);
1476 	if (mlx5_check_vec_rx_support(dev) > 0) {
1477 		rx_pkt_burst = mlx5_rx_burst_vec;
1478 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1479 			dev->data->port_id);
1480 	} else if (mlx5_mprq_enabled(dev)) {
1481 		rx_pkt_burst = mlx5_rx_burst_mprq;
1482 	}
1483 	return rx_pkt_burst;
1484 }
1485 
1486 /**
1487  * Check if mlx5 device was removed.
1488  *
1489  * @param dev
1490  *   Pointer to Ethernet device structure.
1491  *
1492  * @return
1493  *   1 when device is removed, otherwise 0.
1494  */
1495 int
1496 mlx5_is_removed(struct rte_eth_dev *dev)
1497 {
1498 	struct ibv_device_attr device_attr;
1499 	struct mlx5_priv *priv = dev->data->dev_private;
1500 
1501 	if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1502 		return 1;
1503 	return 0;
1504 }
1505 
1506 /**
1507  * Get port ID list of mlx5 instances sharing a common device.
1508  *
1509  * @param[in] dev
1510  *   Device to look for.
1511  * @param[out] port_list
1512  *   Result buffer for collected port IDs.
1513  * @param port_list_n
1514  *   Maximum number of entries in result buffer. If 0, @p port_list can be
1515  *   NULL.
1516  *
1517  * @return
1518  *   Number of matching instances regardless of the @p port_list_n
1519  *   parameter, 0 if none were found.
1520  */
1521 unsigned int
1522 mlx5_dev_to_port_id(const struct rte_device *dev, uint16_t *port_list,
1523 		    unsigned int port_list_n)
1524 {
1525 	uint16_t id;
1526 	unsigned int n = 0;
1527 
1528 	RTE_ETH_FOREACH_DEV_OF(id, dev) {
1529 		if (n < port_list_n)
1530 			port_list[n] = id;
1531 		n++;
1532 	}
1533 	return n;
1534 }
1535 
1536 /**
1537  * Get the E-Switch domain id this port belongs to.
1538  *
1539  * @param[in] port
1540  *   Device port id.
1541  * @param[out] es_domain_id
1542  *   E-Switch domain id.
1543  * @param[out] es_port_id
1544  *   The port id of the port in the E-Switch.
1545  *
1546  * @return
1547  *   0 on success, a negative errno value otherwise and rte_errno is set.
1548  */
1549 int
1550 mlx5_port_to_eswitch_info(uint16_t port,
1551 			  uint16_t *es_domain_id, uint16_t *es_port_id)
1552 {
1553 	struct rte_eth_dev *dev;
1554 	struct mlx5_priv *priv;
1555 
1556 	if (port >= RTE_MAX_ETHPORTS) {
1557 		rte_errno = EINVAL;
1558 		return -rte_errno;
1559 	}
1560 	if (!rte_eth_dev_is_valid_port(port)) {
1561 		rte_errno = ENODEV;
1562 		return -rte_errno;
1563 	}
1564 	dev = &rte_eth_devices[port];
1565 	priv = dev->data->dev_private;
1566 	if (!(priv->representor || priv->master)) {
1567 		rte_errno = EINVAL;
1568 		return -rte_errno;
1569 	}
1570 	if (es_domain_id)
1571 		*es_domain_id = priv->domain_id;
1572 	if (es_port_id)
1573 		*es_port_id = priv->vport_id;
1574 	return 0;
1575 }
1576 
1577 /**
1578  * Get switch information associated with network interface.
1579  *
1580  * @param ifindex
1581  *   Network interface index.
1582  * @param[out] info
1583  *   Switch information object, populated in case of success.
1584  *
1585  * @return
1586  *   0 on success, a negative errno value otherwise and rte_errno is set.
1587  */
1588 int
1589 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1590 {
1591 	char ifname[IF_NAMESIZE];
1592 	char port_name[IF_NAMESIZE];
1593 	FILE *file;
1594 	struct mlx5_switch_info data = {
1595 		.master = 0,
1596 		.representor = 0,
1597 		.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1598 		.port_name = 0,
1599 		.switch_id = 0,
1600 	};
1601 	DIR *dir;
1602 	bool port_switch_id_set = false;
1603 	bool device_dir = false;
1604 	char c;
1605 	int ret;
1606 
1607 	if (!if_indextoname(ifindex, ifname)) {
1608 		rte_errno = errno;
1609 		return -rte_errno;
1610 	}
1611 
1612 	MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1613 	      ifname);
1614 	MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1615 	      ifname);
1616 	MKSTR(pci_device, "/sys/class/net/%s/device",
1617 	      ifname);
1618 
1619 	file = fopen(phys_port_name, "rb");
1620 	if (file != NULL) {
1621 		ret = fscanf(file, "%s", port_name);
1622 		fclose(file);
1623 		if (ret == 1)
1624 			mlx5_translate_port_name(port_name, &data);
1625 	}
1626 	file = fopen(phys_switch_id, "rb");
1627 	if (file == NULL) {
1628 		rte_errno = errno;
1629 		return -rte_errno;
1630 	}
1631 	port_switch_id_set =
1632 		fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1633 		c == '\n';
1634 	fclose(file);
1635 	dir = opendir(pci_device);
1636 	if (dir != NULL) {
1637 		closedir(dir);
1638 		device_dir = true;
1639 	}
1640 	if (port_switch_id_set) {
1641 		/* We have some E-Switch configuration. */
1642 		mlx5_sysfs_check_switch_info(device_dir, &data);
1643 	}
1644 	*info = data;
1645 	assert(!(data.master && data.representor));
1646 	if (data.master && data.representor) {
1647 		DRV_LOG(ERR, "ifindex %u device is recognized as master"
1648 			     " and as representor", ifindex);
1649 		rte_errno = ENODEV;
1650 		return -rte_errno;
1651 	}
1652 	return 0;
1653 }
1654 
1655 /**
1656  * Analyze gathered port parameters via Netlink to recognize master
1657  * and representor devices for E-Switch configuration.
1658  *
1659  * @param[in] num_vf_set
1660  *   flag of presence of number of VFs port attribute.
1661  * @param[inout] switch_info
1662  *   Port information, including port name as a number and port name
1663  *   type if recognized
1664  *
1665  * @return
1666  *   master and representor flags are set in switch_info according to
1667  *   recognized parameters (if any).
1668  */
1669 void
1670 mlx5_nl_check_switch_info(bool num_vf_set,
1671 			  struct mlx5_switch_info *switch_info)
1672 {
1673 	switch (switch_info->name_type) {
1674 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1675 		/*
1676 		 * Name is not recognized, assume the master,
1677 		 * check the number of VFs key presence.
1678 		 */
1679 		switch_info->master = num_vf_set;
1680 		break;
1681 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1682 		/*
1683 		 * Name is not set, this assumes the legacy naming
1684 		 * schema for master, just check if there is a
1685 		 * number of VFs key.
1686 		 */
1687 		switch_info->master = num_vf_set;
1688 		break;
1689 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1690 		/* New uplink naming schema recognized. */
1691 		switch_info->master = 1;
1692 		break;
1693 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1694 		/* Legacy representors naming schema. */
1695 		switch_info->representor = !num_vf_set;
1696 		break;
1697 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1698 		/* New representors naming schema. */
1699 		switch_info->representor = 1;
1700 		break;
1701 	}
1702 }
1703 
1704 /**
1705  * Analyze gathered port parameters via sysfs to recognize master
1706  * and representor devices for E-Switch configuration.
1707  *
1708  * @param[in] device_dir
1709  *   flag of presence of "device" directory under port device key.
1710  * @param[inout] switch_info
1711  *   Port information, including port name as a number and port name
1712  *   type if recognized
1713  *
1714  * @return
1715  *   master and representor flags are set in switch_info according to
1716  *   recognized parameters (if any).
1717  */
1718 void
1719 mlx5_sysfs_check_switch_info(bool device_dir,
1720 			     struct mlx5_switch_info *switch_info)
1721 {
1722 	switch (switch_info->name_type) {
1723 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1724 		/*
1725 		 * Name is not recognized, assume the master,
1726 		 * check the device directory presence.
1727 		 */
1728 		switch_info->master = device_dir;
1729 		break;
1730 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1731 		/*
1732 		 * Name is not set, this assumes the legacy naming
1733 		 * schema for master, just check if there is
1734 		 * a device directory.
1735 		 */
1736 		switch_info->master = device_dir;
1737 		break;
1738 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1739 		/* New uplink naming schema recognized. */
1740 		switch_info->master = 1;
1741 		break;
1742 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1743 		/* Legacy representors naming schema. */
1744 		switch_info->representor = !device_dir;
1745 		break;
1746 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1747 		/* New representors naming schema. */
1748 		switch_info->representor = 1;
1749 		break;
1750 	}
1751 }
1752 
1753 /**
1754  * Extract port name, as a number, from sysfs or netlink information.
1755  *
1756  * @param[in] port_name_in
1757  *   String representing the port name.
1758  * @param[out] port_info_out
1759  *   Port information, including port name as a number and port name
1760  *   type if recognized
1761  *
1762  * @return
1763  *   port_name field set according to recognized name format.
1764  */
1765 void
1766 mlx5_translate_port_name(const char *port_name_in,
1767 			 struct mlx5_switch_info *port_info_out)
1768 {
1769 	char pf_c1, pf_c2, vf_c1, vf_c2;
1770 	char *end;
1771 	int sc_items;
1772 
1773 	/*
1774 	 * Check for port-name as a string of the form pf0vf0
1775 	 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
1776 	 */
1777 	sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
1778 			  &pf_c1, &pf_c2, &port_info_out->pf_num,
1779 			  &vf_c1, &vf_c2, &port_info_out->port_name);
1780 	if (sc_items == 6 &&
1781 	    pf_c1 == 'p' && pf_c2 == 'f' &&
1782 	    vf_c1 == 'v' && vf_c2 == 'f') {
1783 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
1784 		return;
1785 	}
1786 	/*
1787 	 * Check for port-name as a string of the form p0
1788 	 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
1789 	 */
1790 	sc_items = sscanf(port_name_in, "%c%d",
1791 			  &pf_c1, &port_info_out->port_name);
1792 	if (sc_items == 2 && pf_c1 == 'p') {
1793 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1794 		return;
1795 	}
1796 	/* Check for port-name as a number (support kernel ver < 5.0 */
1797 	errno = 0;
1798 	port_info_out->port_name = strtol(port_name_in, &end, 0);
1799 	if (!errno &&
1800 	    (size_t)(end - port_name_in) == strlen(port_name_in)) {
1801 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
1802 		return;
1803 	}
1804 	port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
1805 	return;
1806 }
1807