xref: /dpdk/drivers/net/mlx5/mlx5_ethdev.c (revision 40d9f906f4e2b6028bd61d68d32956c6a809be03)
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 firmware version of a device.
573  *
574  * @param dev
575  *   Ethernet device port.
576  * @param fw_ver
577  *   String output allocated by caller.
578  * @param fw_size
579  *   Size of the output string, including terminating null byte.
580  *
581  * @return
582  *   0 on success, or the size of the non truncated string if too big.
583  */
584 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
585 {
586 	struct mlx5_priv *priv = dev->data->dev_private;
587 	struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
588 	size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
589 
590 	if (fw_size < size)
591 		return size;
592 	if (fw_ver != NULL)
593 		strlcpy(fw_ver, attr->fw_ver, fw_size);
594 	return 0;
595 }
596 
597 /**
598  * Get supported packet types.
599  *
600  * @param dev
601  *   Pointer to Ethernet device structure.
602  *
603  * @return
604  *   A pointer to the supported Packet types array.
605  */
606 const uint32_t *
607 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
608 {
609 	static const uint32_t ptypes[] = {
610 		/* refers to rxq_cq_to_pkt_type() */
611 		RTE_PTYPE_L2_ETHER,
612 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
613 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
614 		RTE_PTYPE_L4_NONFRAG,
615 		RTE_PTYPE_L4_FRAG,
616 		RTE_PTYPE_L4_TCP,
617 		RTE_PTYPE_L4_UDP,
618 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
619 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
620 		RTE_PTYPE_INNER_L4_NONFRAG,
621 		RTE_PTYPE_INNER_L4_FRAG,
622 		RTE_PTYPE_INNER_L4_TCP,
623 		RTE_PTYPE_INNER_L4_UDP,
624 		RTE_PTYPE_UNKNOWN
625 	};
626 
627 	if (dev->rx_pkt_burst == mlx5_rx_burst ||
628 	    dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
629 	    dev->rx_pkt_burst == mlx5_rx_burst_vec)
630 		return ptypes;
631 	return NULL;
632 }
633 
634 /**
635  * Retrieve the master device for representor in the same switch domain.
636  *
637  * @param dev
638  *   Pointer to representor Ethernet device structure.
639  *
640  * @return
641  *   Master device structure  on success, NULL otherwise.
642  */
643 
644 static struct rte_eth_dev *
645 mlx5_find_master_dev(struct rte_eth_dev *dev)
646 {
647 	struct mlx5_priv *priv;
648 	uint16_t port_id;
649 	uint16_t domain_id;
650 
651 	priv = dev->data->dev_private;
652 	domain_id = priv->domain_id;
653 	assert(priv->representor);
654 	RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
655 		priv = rte_eth_devices[port_id].data->dev_private;
656 		if (priv &&
657 		    priv->master &&
658 		    priv->domain_id == domain_id)
659 			return &rte_eth_devices[port_id];
660 	}
661 	return NULL;
662 }
663 
664 /**
665  * DPDK callback to retrieve physical link information.
666  *
667  * @param dev
668  *   Pointer to Ethernet device structure.
669  * @param[out] link
670  *   Storage for current link status.
671  *
672  * @return
673  *   0 on success, a negative errno value otherwise and rte_errno is set.
674  */
675 static int
676 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
677 			       struct rte_eth_link *link)
678 {
679 	struct mlx5_priv *priv = dev->data->dev_private;
680 	struct ethtool_cmd edata = {
681 		.cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
682 	};
683 	struct ifreq ifr;
684 	struct rte_eth_link dev_link;
685 	int link_speed = 0;
686 	int ret;
687 
688 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
689 	if (ret) {
690 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
691 			dev->data->port_id, strerror(rte_errno));
692 		return ret;
693 	}
694 	dev_link = (struct rte_eth_link) {
695 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
696 				(ifr.ifr_flags & IFF_RUNNING)),
697 	};
698 	ifr = (struct ifreq) {
699 		.ifr_data = (void *)&edata,
700 	};
701 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
702 	if (ret) {
703 		if (ret == -ENOTSUP && priv->representor) {
704 			struct rte_eth_dev *master;
705 
706 			/*
707 			 * For representors we can try to inherit link
708 			 * settings from the master device. Actually
709 			 * link settings do not make a lot of sense
710 			 * for representors due to missing physical
711 			 * link. The old kernel drivers supported
712 			 * emulated settings query for representors,
713 			 * the new ones do not, so we have to add
714 			 * this code for compatibility issues.
715 			 */
716 			master = mlx5_find_master_dev(dev);
717 			if (master) {
718 				ifr = (struct ifreq) {
719 					.ifr_data = (void *)&edata,
720 				};
721 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
722 			}
723 		}
724 		if (ret) {
725 			DRV_LOG(WARNING,
726 				"port %u ioctl(SIOCETHTOOL,"
727 				" ETHTOOL_GSET) failed: %s",
728 				dev->data->port_id, strerror(rte_errno));
729 			return ret;
730 		}
731 	}
732 	link_speed = ethtool_cmd_speed(&edata);
733 	if (link_speed == -1)
734 		dev_link.link_speed = ETH_SPEED_NUM_NONE;
735 	else
736 		dev_link.link_speed = link_speed;
737 	priv->link_speed_capa = 0;
738 	if (edata.supported & SUPPORTED_Autoneg)
739 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
740 	if (edata.supported & (SUPPORTED_1000baseT_Full |
741 			       SUPPORTED_1000baseKX_Full))
742 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
743 	if (edata.supported & SUPPORTED_10000baseKR_Full)
744 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
745 	if (edata.supported & (SUPPORTED_40000baseKR4_Full |
746 			       SUPPORTED_40000baseCR4_Full |
747 			       SUPPORTED_40000baseSR4_Full |
748 			       SUPPORTED_40000baseLR4_Full))
749 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
750 	dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
751 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
752 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
753 			ETH_LINK_SPEED_FIXED);
754 	if (((dev_link.link_speed && !dev_link.link_status) ||
755 	     (!dev_link.link_speed && dev_link.link_status))) {
756 		rte_errno = EAGAIN;
757 		return -rte_errno;
758 	}
759 	*link = dev_link;
760 	return 0;
761 }
762 
763 /**
764  * Retrieve physical link information (unlocked version using new ioctl).
765  *
766  * @param dev
767  *   Pointer to Ethernet device structure.
768  * @param[out] link
769  *   Storage for current link status.
770  *
771  * @return
772  *   0 on success, a negative errno value otherwise and rte_errno is set.
773  */
774 static int
775 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
776 			     struct rte_eth_link *link)
777 
778 {
779 	struct mlx5_priv *priv = dev->data->dev_private;
780 	struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
781 	struct ifreq ifr;
782 	struct rte_eth_link dev_link;
783 	struct rte_eth_dev *master = NULL;
784 	uint64_t sc;
785 	int ret;
786 
787 	ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
788 	if (ret) {
789 		DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
790 			dev->data->port_id, strerror(rte_errno));
791 		return ret;
792 	}
793 	dev_link = (struct rte_eth_link) {
794 		.link_status = ((ifr.ifr_flags & IFF_UP) &&
795 				(ifr.ifr_flags & IFF_RUNNING)),
796 	};
797 	ifr = (struct ifreq) {
798 		.ifr_data = (void *)&gcmd,
799 	};
800 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
801 	if (ret) {
802 		if (ret == -ENOTSUP && priv->representor) {
803 			/*
804 			 * For representors we can try to inherit link
805 			 * settings from the master device. Actually
806 			 * link settings do not make a lot of sense
807 			 * for representors due to missing physical
808 			 * link. The old kernel drivers supported
809 			 * emulated settings query for representors,
810 			 * the new ones do not, so we have to add
811 			 * this code for compatibility issues.
812 			 */
813 			master = mlx5_find_master_dev(dev);
814 			if (master) {
815 				ifr = (struct ifreq) {
816 					.ifr_data = (void *)&gcmd,
817 				};
818 				ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
819 			}
820 		}
821 		if (ret) {
822 			DRV_LOG(DEBUG,
823 				"port %u ioctl(SIOCETHTOOL,"
824 				" ETHTOOL_GLINKSETTINGS) failed: %s",
825 				dev->data->port_id, strerror(rte_errno));
826 			return ret;
827 		}
828 
829 	}
830 	gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
831 
832 	alignas(struct ethtool_link_settings)
833 	uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
834 		     sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
835 	struct ethtool_link_settings *ecmd = (void *)data;
836 
837 	*ecmd = gcmd;
838 	ifr.ifr_data = (void *)ecmd;
839 	ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
840 	if (ret) {
841 		DRV_LOG(DEBUG,
842 			"port %u ioctl(SIOCETHTOOL,"
843 			"ETHTOOL_GLINKSETTINGS) failed: %s",
844 			dev->data->port_id, strerror(rte_errno));
845 		return ret;
846 	}
847 	dev_link.link_speed = ecmd->speed;
848 	sc = ecmd->link_mode_masks[0] |
849 		((uint64_t)ecmd->link_mode_masks[1] << 32);
850 	priv->link_speed_capa = 0;
851 	if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
852 		priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
853 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
854 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
855 		priv->link_speed_capa |= ETH_LINK_SPEED_1G;
856 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
857 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
858 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
859 		priv->link_speed_capa |= ETH_LINK_SPEED_10G;
860 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
861 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
862 		priv->link_speed_capa |= ETH_LINK_SPEED_20G;
863 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
864 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
865 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
866 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
867 		priv->link_speed_capa |= ETH_LINK_SPEED_40G;
868 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
869 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
870 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
871 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
872 		priv->link_speed_capa |= ETH_LINK_SPEED_56G;
873 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
874 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
875 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
876 		priv->link_speed_capa |= ETH_LINK_SPEED_25G;
877 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
878 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
879 		priv->link_speed_capa |= ETH_LINK_SPEED_50G;
880 	if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
881 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
882 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
883 		  MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
884 		priv->link_speed_capa |= ETH_LINK_SPEED_100G;
885 	dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
886 				ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
887 	dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
888 				  ETH_LINK_SPEED_FIXED);
889 	if (((dev_link.link_speed && !dev_link.link_status) ||
890 	     (!dev_link.link_speed && dev_link.link_status))) {
891 		rte_errno = EAGAIN;
892 		return -rte_errno;
893 	}
894 	*link = dev_link;
895 	return 0;
896 }
897 
898 /**
899  * DPDK callback to retrieve physical link information.
900  *
901  * @param dev
902  *   Pointer to Ethernet device structure.
903  * @param wait_to_complete
904  *   Wait for request completion.
905  *
906  * @return
907  *   0 if link status was not updated, positive if it was, a negative errno
908  *   value otherwise and rte_errno is set.
909  */
910 int
911 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
912 {
913 	int ret;
914 	struct rte_eth_link dev_link;
915 	time_t start_time = time(NULL);
916 
917 	do {
918 		ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
919 		if (ret)
920 			ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
921 		if (ret == 0)
922 			break;
923 		/* Handle wait to complete situation. */
924 		if (wait_to_complete && ret == -EAGAIN) {
925 			if (abs((int)difftime(time(NULL), start_time)) <
926 			    MLX5_LINK_STATUS_TIMEOUT) {
927 				usleep(0);
928 				continue;
929 			} else {
930 				rte_errno = EBUSY;
931 				return -rte_errno;
932 			}
933 		} else if (ret < 0) {
934 			return ret;
935 		}
936 	} while (wait_to_complete);
937 	ret = !!memcmp(&dev->data->dev_link, &dev_link,
938 		       sizeof(struct rte_eth_link));
939 	dev->data->dev_link = dev_link;
940 	return ret;
941 }
942 
943 /**
944  * DPDK callback to change the MTU.
945  *
946  * @param dev
947  *   Pointer to Ethernet device structure.
948  * @param in_mtu
949  *   New MTU.
950  *
951  * @return
952  *   0 on success, a negative errno value otherwise and rte_errno is set.
953  */
954 int
955 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
956 {
957 	struct mlx5_priv *priv = dev->data->dev_private;
958 	uint16_t kern_mtu = 0;
959 	int ret;
960 
961 	ret = mlx5_get_mtu(dev, &kern_mtu);
962 	if (ret)
963 		return ret;
964 	/* Set kernel interface MTU first. */
965 	ret = mlx5_set_mtu(dev, mtu);
966 	if (ret)
967 		return ret;
968 	ret = mlx5_get_mtu(dev, &kern_mtu);
969 	if (ret)
970 		return ret;
971 	if (kern_mtu == mtu) {
972 		priv->mtu = mtu;
973 		DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
974 			dev->data->port_id, mtu);
975 		return 0;
976 	}
977 	rte_errno = EAGAIN;
978 	return -rte_errno;
979 }
980 
981 /**
982  * DPDK callback to get flow control status.
983  *
984  * @param dev
985  *   Pointer to Ethernet device structure.
986  * @param[out] fc_conf
987  *   Flow control output buffer.
988  *
989  * @return
990  *   0 on success, a negative errno value otherwise and rte_errno is set.
991  */
992 int
993 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
994 {
995 	struct ifreq ifr;
996 	struct ethtool_pauseparam ethpause = {
997 		.cmd = ETHTOOL_GPAUSEPARAM
998 	};
999 	int ret;
1000 
1001 	ifr.ifr_data = (void *)&ethpause;
1002 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1003 	if (ret) {
1004 		DRV_LOG(WARNING,
1005 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1006 			" %s",
1007 			dev->data->port_id, strerror(rte_errno));
1008 		return ret;
1009 	}
1010 	fc_conf->autoneg = ethpause.autoneg;
1011 	if (ethpause.rx_pause && ethpause.tx_pause)
1012 		fc_conf->mode = RTE_FC_FULL;
1013 	else if (ethpause.rx_pause)
1014 		fc_conf->mode = RTE_FC_RX_PAUSE;
1015 	else if (ethpause.tx_pause)
1016 		fc_conf->mode = RTE_FC_TX_PAUSE;
1017 	else
1018 		fc_conf->mode = RTE_FC_NONE;
1019 	return 0;
1020 }
1021 
1022 /**
1023  * DPDK callback to modify flow control parameters.
1024  *
1025  * @param dev
1026  *   Pointer to Ethernet device structure.
1027  * @param[in] fc_conf
1028  *   Flow control parameters.
1029  *
1030  * @return
1031  *   0 on success, a negative errno value otherwise and rte_errno is set.
1032  */
1033 int
1034 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1035 {
1036 	struct ifreq ifr;
1037 	struct ethtool_pauseparam ethpause = {
1038 		.cmd = ETHTOOL_SPAUSEPARAM
1039 	};
1040 	int ret;
1041 
1042 	ifr.ifr_data = (void *)&ethpause;
1043 	ethpause.autoneg = fc_conf->autoneg;
1044 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1045 	    (fc_conf->mode & RTE_FC_RX_PAUSE))
1046 		ethpause.rx_pause = 1;
1047 	else
1048 		ethpause.rx_pause = 0;
1049 
1050 	if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1051 	    (fc_conf->mode & RTE_FC_TX_PAUSE))
1052 		ethpause.tx_pause = 1;
1053 	else
1054 		ethpause.tx_pause = 0;
1055 	ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1056 	if (ret) {
1057 		DRV_LOG(WARNING,
1058 			"port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1059 			" failed: %s",
1060 			dev->data->port_id, strerror(rte_errno));
1061 		return ret;
1062 	}
1063 	return 0;
1064 }
1065 
1066 /**
1067  * Get PCI information from struct ibv_device.
1068  *
1069  * @param device
1070  *   Pointer to Ethernet device structure.
1071  * @param[out] pci_addr
1072  *   PCI bus address output buffer.
1073  *
1074  * @return
1075  *   0 on success, a negative errno value otherwise and rte_errno is set.
1076  */
1077 int
1078 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1079 			    struct rte_pci_addr *pci_addr)
1080 {
1081 	FILE *file;
1082 	char line[32];
1083 	MKSTR(path, "%s/device/uevent", device->ibdev_path);
1084 
1085 	file = fopen(path, "rb");
1086 	if (file == NULL) {
1087 		rte_errno = errno;
1088 		return -rte_errno;
1089 	}
1090 	while (fgets(line, sizeof(line), file) == line) {
1091 		size_t len = strlen(line);
1092 		int ret;
1093 
1094 		/* Truncate long lines. */
1095 		if (len == (sizeof(line) - 1))
1096 			while (line[(len - 1)] != '\n') {
1097 				ret = fgetc(file);
1098 				if (ret == EOF)
1099 					break;
1100 				line[(len - 1)] = ret;
1101 			}
1102 		/* Extract information. */
1103 		if (sscanf(line,
1104 			   "PCI_SLOT_NAME="
1105 			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1106 			   &pci_addr->domain,
1107 			   &pci_addr->bus,
1108 			   &pci_addr->devid,
1109 			   &pci_addr->function) == 4) {
1110 			ret = 0;
1111 			break;
1112 		}
1113 	}
1114 	fclose(file);
1115 	return 0;
1116 }
1117 
1118 /**
1119  * Handle asynchronous removal event for entire multiport device.
1120  *
1121  * @param sh
1122  *   Infiniband device shared context.
1123  */
1124 static void
1125 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1126 {
1127 	uint32_t i;
1128 
1129 	for (i = 0; i < sh->max_port; ++i) {
1130 		struct rte_eth_dev *dev;
1131 
1132 		if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1133 			/*
1134 			 * Or not existing port either no
1135 			 * handler installed for this port.
1136 			 */
1137 			continue;
1138 		}
1139 		dev = &rte_eth_devices[sh->port[i].ih_port_id];
1140 		assert(dev);
1141 		if (dev->data->dev_conf.intr_conf.rmv)
1142 			_rte_eth_dev_callback_process
1143 				(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1144 	}
1145 }
1146 
1147 /**
1148  * Handle shared asynchronous events the NIC (removal event
1149  * and link status change). Supports multiport IB device.
1150  *
1151  * @param cb_arg
1152  *   Callback argument.
1153  */
1154 void
1155 mlx5_dev_interrupt_handler(void *cb_arg)
1156 {
1157 	struct mlx5_ibv_shared *sh = cb_arg;
1158 	struct ibv_async_event event;
1159 
1160 	/* Read all message from the IB device and acknowledge them. */
1161 	for (;;) {
1162 		struct rte_eth_dev *dev;
1163 		uint32_t tmp;
1164 
1165 		if (mlx5_glue->get_async_event(sh->ctx, &event))
1166 			break;
1167 		/* Retrieve and check IB port index. */
1168 		tmp = (uint32_t)event.element.port_num;
1169 		if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1170 			/*
1171 			 * The DEVICE_FATAL event is called once for
1172 			 * entire device without port specifying.
1173 			 * We should notify all existing ports.
1174 			 */
1175 			mlx5_glue->ack_async_event(&event);
1176 			mlx5_dev_interrupt_device_fatal(sh);
1177 			continue;
1178 		}
1179 		assert(tmp && (tmp <= sh->max_port));
1180 		if (!tmp) {
1181 			/* Unsupported devive level event. */
1182 			mlx5_glue->ack_async_event(&event);
1183 			DRV_LOG(DEBUG,
1184 				"unsupported common event (type %d)",
1185 				event.event_type);
1186 			continue;
1187 		}
1188 		if (tmp > sh->max_port) {
1189 			/* Invalid IB port index. */
1190 			mlx5_glue->ack_async_event(&event);
1191 			DRV_LOG(DEBUG,
1192 				"cannot handle an event (type %d)"
1193 				"due to invalid IB port index (%u)",
1194 				event.event_type, tmp);
1195 			continue;
1196 		}
1197 		if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1198 			/* No handler installed. */
1199 			mlx5_glue->ack_async_event(&event);
1200 			DRV_LOG(DEBUG,
1201 				"cannot handle an event (type %d)"
1202 				"due to no handler installed for port %u",
1203 				event.event_type, tmp);
1204 			continue;
1205 		}
1206 		/* Retrieve ethernet device descriptor. */
1207 		tmp = sh->port[tmp - 1].ih_port_id;
1208 		dev = &rte_eth_devices[tmp];
1209 		assert(dev);
1210 		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1211 		     event.event_type == IBV_EVENT_PORT_ERR) &&
1212 			dev->data->dev_conf.intr_conf.lsc) {
1213 			mlx5_glue->ack_async_event(&event);
1214 			if (mlx5_link_update(dev, 0) == -EAGAIN) {
1215 				usleep(0);
1216 				continue;
1217 			}
1218 			_rte_eth_dev_callback_process
1219 				(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1220 			continue;
1221 		}
1222 		DRV_LOG(DEBUG,
1223 			"port %u cannot handle an unknown event (type %d)",
1224 			dev->data->port_id, event.event_type);
1225 		mlx5_glue->ack_async_event(&event);
1226 	}
1227 }
1228 
1229 /**
1230  * Uninstall shared asynchronous device events handler.
1231  * This function is implemeted to support event sharing
1232  * between multiple ports of single IB device.
1233  *
1234  * @param dev
1235  *   Pointer to Ethernet device.
1236  */
1237 static void
1238 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1239 {
1240 	struct mlx5_priv *priv = dev->data->dev_private;
1241 	struct mlx5_ibv_shared *sh = priv->sh;
1242 
1243 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1244 		return;
1245 	pthread_mutex_lock(&sh->intr_mutex);
1246 	assert(priv->ibv_port);
1247 	assert(priv->ibv_port <= sh->max_port);
1248 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1249 	if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1250 		goto exit;
1251 	assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1252 					(uint32_t)dev->data->port_id);
1253 	assert(sh->intr_cnt);
1254 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1255 	if (!sh->intr_cnt || --sh->intr_cnt)
1256 		goto exit;
1257 	rte_intr_callback_unregister(&sh->intr_handle,
1258 				     mlx5_dev_interrupt_handler, sh);
1259 	sh->intr_handle.fd = 0;
1260 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1261 exit:
1262 	pthread_mutex_unlock(&sh->intr_mutex);
1263 }
1264 
1265 /**
1266  * Install shared asyncronous device events handler.
1267  * This function is implemeted to support event sharing
1268  * between multiple ports of single IB device.
1269  *
1270  * @param dev
1271  *   Pointer to Ethernet device.
1272  */
1273 static void
1274 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1275 {
1276 	struct mlx5_priv *priv = dev->data->dev_private;
1277 	struct mlx5_ibv_shared *sh = priv->sh;
1278 	int ret;
1279 	int flags;
1280 
1281 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1282 		return;
1283 	pthread_mutex_lock(&sh->intr_mutex);
1284 	assert(priv->ibv_port);
1285 	assert(priv->ibv_port <= sh->max_port);
1286 	assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1287 	if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1288 		/* The handler is already installed for this port. */
1289 		assert(sh->intr_cnt);
1290 		goto exit;
1291 	}
1292 	sh->port[priv->ibv_port - 1].ih_port_id = (uint32_t)dev->data->port_id;
1293 	if (sh->intr_cnt) {
1294 		sh->intr_cnt++;
1295 		goto exit;
1296 	}
1297 	/* No shared handler installed. */
1298 	assert(sh->ctx->async_fd > 0);
1299 	flags = fcntl(sh->ctx->async_fd, F_GETFL);
1300 	ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1301 	if (ret) {
1302 		DRV_LOG(INFO, "failed to change file descriptor"
1303 			      " async event queue");
1304 		/* Indicate there will be no interrupts. */
1305 		dev->data->dev_conf.intr_conf.lsc = 0;
1306 		dev->data->dev_conf.intr_conf.rmv = 0;
1307 		sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1308 		goto exit;
1309 	}
1310 	sh->intr_handle.fd = sh->ctx->async_fd;
1311 	sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1312 	rte_intr_callback_register(&sh->intr_handle,
1313 				   mlx5_dev_interrupt_handler, sh);
1314 	sh->intr_cnt++;
1315 exit:
1316 	pthread_mutex_unlock(&sh->intr_mutex);
1317 }
1318 
1319 /**
1320  * Uninstall interrupt handler.
1321  *
1322  * @param dev
1323  *   Pointer to Ethernet device.
1324  */
1325 void
1326 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1327 {
1328 	mlx5_dev_shared_handler_uninstall(dev);
1329 }
1330 
1331 /**
1332  * Install interrupt handler.
1333  *
1334  * @param dev
1335  *   Pointer to Ethernet device.
1336  */
1337 void
1338 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1339 {
1340 	mlx5_dev_shared_handler_install(dev);
1341 }
1342 
1343 /**
1344  * DPDK callback to bring the link DOWN.
1345  *
1346  * @param dev
1347  *   Pointer to Ethernet device structure.
1348  *
1349  * @return
1350  *   0 on success, a negative errno value otherwise and rte_errno is set.
1351  */
1352 int
1353 mlx5_set_link_down(struct rte_eth_dev *dev)
1354 {
1355 	return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1356 }
1357 
1358 /**
1359  * DPDK callback to bring the link UP.
1360  *
1361  * @param dev
1362  *   Pointer to Ethernet device structure.
1363  *
1364  * @return
1365  *   0 on success, a negative errno value otherwise and rte_errno is set.
1366  */
1367 int
1368 mlx5_set_link_up(struct rte_eth_dev *dev)
1369 {
1370 	return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1371 }
1372 
1373 /**
1374  * Configure the TX function to use.
1375  *
1376  * @param dev
1377  *   Pointer to private data structure.
1378  *
1379  * @return
1380  *   Pointer to selected Tx burst function.
1381  */
1382 eth_tx_burst_t
1383 mlx5_select_tx_function(struct rte_eth_dev *dev)
1384 {
1385 	struct mlx5_priv *priv = dev->data->dev_private;
1386 	eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1387 	struct mlx5_dev_config *config = &priv->config;
1388 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1389 	int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1390 				    DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1391 				    DEV_TX_OFFLOAD_GRE_TNL_TSO |
1392 				    DEV_TX_OFFLOAD_IP_TNL_TSO |
1393 				    DEV_TX_OFFLOAD_UDP_TNL_TSO));
1394 	int swp = !!(tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
1395 				    DEV_TX_OFFLOAD_UDP_TNL_TSO |
1396 				    DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM));
1397 	int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1398 
1399 	assert(priv != NULL);
1400 	/* Select appropriate TX function. */
1401 	if (vlan_insert || tso || swp)
1402 		return tx_pkt_burst;
1403 	if (config->mps == MLX5_MPW_ENHANCED) {
1404 		if (mlx5_check_vec_tx_support(dev) > 0) {
1405 			if (mlx5_check_raw_vec_tx_support(dev) > 0)
1406 				tx_pkt_burst = mlx5_tx_burst_raw_vec;
1407 			else
1408 				tx_pkt_burst = mlx5_tx_burst_vec;
1409 			DRV_LOG(DEBUG,
1410 				"port %u selected enhanced MPW Tx vectorized"
1411 				" function",
1412 				dev->data->port_id);
1413 		} else {
1414 			tx_pkt_burst = mlx5_tx_burst_empw;
1415 			DRV_LOG(DEBUG,
1416 				"port %u selected enhanced MPW Tx function",
1417 				dev->data->port_id);
1418 		}
1419 	} else if (config->mps && (config->txq_inline > 0)) {
1420 		tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1421 		DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1422 			dev->data->port_id);
1423 	} else if (config->mps) {
1424 		tx_pkt_burst = mlx5_tx_burst_mpw;
1425 		DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1426 			dev->data->port_id);
1427 	}
1428 	return tx_pkt_burst;
1429 }
1430 
1431 /**
1432  * Configure the RX function to use.
1433  *
1434  * @param dev
1435  *   Pointer to private data structure.
1436  *
1437  * @return
1438  *   Pointer to selected Rx burst function.
1439  */
1440 eth_rx_burst_t
1441 mlx5_select_rx_function(struct rte_eth_dev *dev)
1442 {
1443 	eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1444 
1445 	assert(dev != NULL);
1446 	if (mlx5_check_vec_rx_support(dev) > 0) {
1447 		rx_pkt_burst = mlx5_rx_burst_vec;
1448 		DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1449 			dev->data->port_id);
1450 	} else if (mlx5_mprq_enabled(dev)) {
1451 		rx_pkt_burst = mlx5_rx_burst_mprq;
1452 	}
1453 	return rx_pkt_burst;
1454 }
1455 
1456 /**
1457  * Check if mlx5 device was removed.
1458  *
1459  * @param dev
1460  *   Pointer to Ethernet device structure.
1461  *
1462  * @return
1463  *   1 when device is removed, otherwise 0.
1464  */
1465 int
1466 mlx5_is_removed(struct rte_eth_dev *dev)
1467 {
1468 	struct ibv_device_attr device_attr;
1469 	struct mlx5_priv *priv = dev->data->dev_private;
1470 
1471 	if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1472 		return 1;
1473 	return 0;
1474 }
1475 
1476 /**
1477  * Get port ID list of mlx5 instances sharing a common device.
1478  *
1479  * @param[in] dev
1480  *   Device to look for.
1481  * @param[out] port_list
1482  *   Result buffer for collected port IDs.
1483  * @param port_list_n
1484  *   Maximum number of entries in result buffer. If 0, @p port_list can be
1485  *   NULL.
1486  *
1487  * @return
1488  *   Number of matching instances regardless of the @p port_list_n
1489  *   parameter, 0 if none were found.
1490  */
1491 unsigned int
1492 mlx5_dev_to_port_id(const struct rte_device *dev, uint16_t *port_list,
1493 		    unsigned int port_list_n)
1494 {
1495 	uint16_t id;
1496 	unsigned int n = 0;
1497 
1498 	RTE_ETH_FOREACH_DEV_OF(id, dev) {
1499 		if (n < port_list_n)
1500 			port_list[n] = id;
1501 		n++;
1502 	}
1503 	return n;
1504 }
1505 
1506 /**
1507  * Get the E-Switch domain id this port belongs to.
1508  *
1509  * @param[in] port
1510  *   Device port id.
1511  * @param[out] es_domain_id
1512  *   E-Switch domain id.
1513  * @param[out] es_port_id
1514  *   The port id of the port in the E-Switch.
1515  *
1516  * @return
1517  *   0 on success, a negative errno value otherwise and rte_errno is set.
1518  */
1519 int
1520 mlx5_port_to_eswitch_info(uint16_t port,
1521 			  uint16_t *es_domain_id, uint16_t *es_port_id)
1522 {
1523 	struct rte_eth_dev *dev;
1524 	struct mlx5_priv *priv;
1525 
1526 	if (port >= RTE_MAX_ETHPORTS) {
1527 		rte_errno = EINVAL;
1528 		return -rte_errno;
1529 	}
1530 	if (!rte_eth_dev_is_valid_port(port)) {
1531 		rte_errno = ENODEV;
1532 		return -rte_errno;
1533 	}
1534 	dev = &rte_eth_devices[port];
1535 	priv = dev->data->dev_private;
1536 	if (!(priv->representor || priv->master)) {
1537 		rte_errno = EINVAL;
1538 		return -rte_errno;
1539 	}
1540 	if (es_domain_id)
1541 		*es_domain_id = priv->domain_id;
1542 	if (es_port_id)
1543 		*es_port_id = priv->vport_id;
1544 	return 0;
1545 }
1546 
1547 /**
1548  * Get switch information associated with network interface.
1549  *
1550  * @param ifindex
1551  *   Network interface index.
1552  * @param[out] info
1553  *   Switch information object, populated in case of success.
1554  *
1555  * @return
1556  *   0 on success, a negative errno value otherwise and rte_errno is set.
1557  */
1558 int
1559 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1560 {
1561 	char ifname[IF_NAMESIZE];
1562 	char port_name[IF_NAMESIZE];
1563 	FILE *file;
1564 	struct mlx5_switch_info data = {
1565 		.master = 0,
1566 		.representor = 0,
1567 		.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1568 		.port_name = 0,
1569 		.switch_id = 0,
1570 	};
1571 	DIR *dir;
1572 	bool port_switch_id_set = false;
1573 	bool device_dir = false;
1574 	char c;
1575 	int ret;
1576 
1577 	if (!if_indextoname(ifindex, ifname)) {
1578 		rte_errno = errno;
1579 		return -rte_errno;
1580 	}
1581 
1582 	MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1583 	      ifname);
1584 	MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1585 	      ifname);
1586 	MKSTR(pci_device, "/sys/class/net/%s/device",
1587 	      ifname);
1588 
1589 	file = fopen(phys_port_name, "rb");
1590 	if (file != NULL) {
1591 		ret = fscanf(file, "%s", port_name);
1592 		fclose(file);
1593 		if (ret == 1)
1594 			mlx5_translate_port_name(port_name, &data);
1595 	}
1596 	file = fopen(phys_switch_id, "rb");
1597 	if (file == NULL) {
1598 		rte_errno = errno;
1599 		return -rte_errno;
1600 	}
1601 	port_switch_id_set =
1602 		fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1603 		c == '\n';
1604 	fclose(file);
1605 	dir = opendir(pci_device);
1606 	if (dir != NULL) {
1607 		closedir(dir);
1608 		device_dir = true;
1609 	}
1610 	if (port_switch_id_set) {
1611 		/* We have some E-Switch configuration. */
1612 		mlx5_sysfs_check_switch_info(device_dir, &data);
1613 	}
1614 	*info = data;
1615 	assert(!(data.master && data.representor));
1616 	if (data.master && data.representor) {
1617 		DRV_LOG(ERR, "ifindex %u device is recognized as master"
1618 			     " and as representor", ifindex);
1619 		rte_errno = ENODEV;
1620 		return -rte_errno;
1621 	}
1622 	return 0;
1623 }
1624 
1625 /**
1626  * Analyze gathered port parameters via Netlink to recognize master
1627  * and representor devices for E-Switch configuration.
1628  *
1629  * @param[in] num_vf_set
1630  *   flag of presence of number of VFs port attribute.
1631  * @param[inout] switch_info
1632  *   Port information, including port name as a number and port name
1633  *   type if recognized
1634  *
1635  * @return
1636  *   master and representor flags are set in switch_info according to
1637  *   recognized parameters (if any).
1638  */
1639 void
1640 mlx5_nl_check_switch_info(bool num_vf_set,
1641 			  struct mlx5_switch_info *switch_info)
1642 {
1643 	switch (switch_info->name_type) {
1644 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1645 		/*
1646 		 * Name is not recognized, assume the master,
1647 		 * check the number of VFs key presence.
1648 		 */
1649 		switch_info->master = num_vf_set;
1650 		break;
1651 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1652 		/*
1653 		 * Name is not set, this assumes the legacy naming
1654 		 * schema for master, just check if there is a
1655 		 * number of VFs key.
1656 		 */
1657 		switch_info->master = num_vf_set;
1658 		break;
1659 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1660 		/* New uplink naming schema recognized. */
1661 		switch_info->master = 1;
1662 		break;
1663 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1664 		/* Legacy representors naming schema. */
1665 		switch_info->representor = !num_vf_set;
1666 		break;
1667 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1668 		/* New representors naming schema. */
1669 		switch_info->representor = 1;
1670 		break;
1671 	}
1672 }
1673 
1674 /**
1675  * Analyze gathered port parameters via sysfs to recognize master
1676  * and representor devices for E-Switch configuration.
1677  *
1678  * @param[in] device_dir
1679  *   flag of presence of "device" directory under port device key.
1680  * @param[inout] switch_info
1681  *   Port information, including port name as a number and port name
1682  *   type if recognized
1683  *
1684  * @return
1685  *   master and representor flags are set in switch_info according to
1686  *   recognized parameters (if any).
1687  */
1688 void
1689 mlx5_sysfs_check_switch_info(bool device_dir,
1690 			     struct mlx5_switch_info *switch_info)
1691 {
1692 	switch (switch_info->name_type) {
1693 	case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1694 		/*
1695 		 * Name is not recognized, assume the master,
1696 		 * check the device directory presence.
1697 		 */
1698 		switch_info->master = device_dir;
1699 		break;
1700 	case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1701 		/*
1702 		 * Name is not set, this assumes the legacy naming
1703 		 * schema for master, just check if there is
1704 		 * a device directory.
1705 		 */
1706 		switch_info->master = device_dir;
1707 		break;
1708 	case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1709 		/* New uplink naming schema recognized. */
1710 		switch_info->master = 1;
1711 		break;
1712 	case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1713 		/* Legacy representors naming schema. */
1714 		switch_info->representor = !device_dir;
1715 		break;
1716 	case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1717 		/* New representors naming schema. */
1718 		switch_info->representor = 1;
1719 		break;
1720 	}
1721 }
1722 
1723 /**
1724  * Extract port name, as a number, from sysfs or netlink information.
1725  *
1726  * @param[in] port_name_in
1727  *   String representing the port name.
1728  * @param[out] port_info_out
1729  *   Port information, including port name as a number and port name
1730  *   type if recognized
1731  *
1732  * @return
1733  *   port_name field set according to recognized name format.
1734  */
1735 void
1736 mlx5_translate_port_name(const char *port_name_in,
1737 			 struct mlx5_switch_info *port_info_out)
1738 {
1739 	char pf_c1, pf_c2, vf_c1, vf_c2;
1740 	char *end;
1741 	int sc_items;
1742 
1743 	/*
1744 	 * Check for port-name as a string of the form pf0vf0
1745 	 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
1746 	 */
1747 	sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
1748 			  &pf_c1, &pf_c2, &port_info_out->pf_num,
1749 			  &vf_c1, &vf_c2, &port_info_out->port_name);
1750 	if (sc_items == 6 &&
1751 	    pf_c1 == 'p' && pf_c2 == 'f' &&
1752 	    vf_c1 == 'v' && vf_c2 == 'f') {
1753 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
1754 		return;
1755 	}
1756 	/*
1757 	 * Check for port-name as a string of the form p0
1758 	 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
1759 	 */
1760 	sc_items = sscanf(port_name_in, "%c%d",
1761 			  &pf_c1, &port_info_out->port_name);
1762 	if (sc_items == 2 && pf_c1 == 'p') {
1763 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1764 		return;
1765 	}
1766 	/* Check for port-name as a number (support kernel ver < 5.0 */
1767 	errno = 0;
1768 	port_info_out->port_name = strtol(port_name_in, &end, 0);
1769 	if (!errno &&
1770 	    (size_t)(end - port_name_in) == strlen(port_name_in)) {
1771 		port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
1772 		return;
1773 	}
1774 	port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
1775 	return;
1776 }
1777