1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
18 #include <sys/ioctl.h>
19 #include <sys/socket.h>
20 #include <netinet/in.h>
21 #include <linux/ethtool.h>
22 #include <linux/sockios.h>
28 #include <rte_atomic.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_bus_pci.h>
32 #include <rte_common.h>
33 #include <rte_interrupts.h>
34 #include <rte_malloc.h>
35 #include <rte_string_fns.h>
36 #include <rte_rwlock.h>
37 #include <rte_cycles.h>
40 #include "mlx5_glue.h"
41 #include "mlx5_rxtx.h"
42 #include "mlx5_utils.h"
44 /* Supported speed values found in /usr/include/linux/ethtool.h */
45 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
46 #define SUPPORTED_40000baseKR4_Full (1 << 23)
48 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
49 #define SUPPORTED_40000baseCR4_Full (1 << 24)
51 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
52 #define SUPPORTED_40000baseSR4_Full (1 << 25)
54 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
55 #define SUPPORTED_40000baseLR4_Full (1 << 26)
57 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
58 #define SUPPORTED_56000baseKR4_Full (1 << 27)
60 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
61 #define SUPPORTED_56000baseCR4_Full (1 << 28)
63 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
64 #define SUPPORTED_56000baseSR4_Full (1 << 29)
66 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
67 #define SUPPORTED_56000baseLR4_Full (1 << 30)
70 /* Add defines in case the running kernel is not the same as user headers. */
71 #ifndef ETHTOOL_GLINKSETTINGS
72 struct ethtool_link_settings {
81 uint8_t eth_tp_mdix_ctrl;
82 int8_t link_mode_masks_nwords;
84 uint32_t link_mode_masks[];
87 #define ETHTOOL_GLINKSETTINGS 0x0000004c
88 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
89 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
90 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
91 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
92 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
93 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
94 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
95 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
96 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
97 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
98 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
99 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
100 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
101 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
102 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
103 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
105 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
106 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
107 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
108 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
110 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
111 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
112 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
114 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
115 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
116 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
117 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
118 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
122 * Get master interface name from private structure.
125 * Pointer to Ethernet device.
127 * Interface name output buffer.
130 * 0 on success, a negative errno value otherwise and rte_errno is set.
133 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
137 unsigned int dev_type = 0;
138 unsigned int dev_port_prev = ~0u;
139 char match[IF_NAMESIZE] = "";
143 MKSTR(path, "%s/device/net", ibdev_path);
151 while ((dent = readdir(dir)) != NULL) {
152 char *name = dent->d_name;
154 unsigned int dev_port;
157 if ((name[0] == '.') &&
158 ((name[1] == '\0') ||
159 ((name[1] == '.') && (name[2] == '\0'))))
162 MKSTR(path, "%s/device/net/%s/%s",
164 (dev_type ? "dev_id" : "dev_port"));
166 file = fopen(path, "rb");
171 * Switch to dev_id when dev_port does not exist as
172 * is the case with Linux kernel versions < 3.15.
183 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
188 * Switch to dev_id when dev_port returns the same value for
189 * all ports. May happen when using a MOFED release older than
190 * 3.0 with a Linux kernel >= 3.15.
192 if (dev_port == dev_port_prev)
194 dev_port_prev = dev_port;
196 strlcpy(match, name, sizeof(match));
199 if (match[0] == '\0') {
203 strncpy(*ifname, match, sizeof(*ifname));
208 * Get interface name from private structure.
210 * This is a port representor-aware version of mlx5_get_master_ifname().
213 * Pointer to Ethernet device.
215 * Interface name output buffer.
218 * 0 on success, a negative errno value otherwise and rte_errno is set.
221 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
223 struct mlx5_priv *priv = dev->data->dev_private;
224 unsigned int ifindex;
228 ifindex = mlx5_ifindex(dev);
230 if (!priv->representor)
231 return mlx5_get_master_ifname(priv->sh->ibdev_path,
236 if (if_indextoname(ifindex, &(*ifname)[0]))
243 * Get the interface index from device name.
246 * Pointer to Ethernet device.
249 * Nonzero interface index on success, zero otherwise and rte_errno is set.
252 mlx5_ifindex(const struct rte_eth_dev *dev)
254 struct mlx5_priv *priv = dev->data->dev_private;
255 unsigned int ifindex;
258 assert(priv->if_index);
259 ifindex = priv->if_index;
266 * Perform ifreq ioctl() on associated Ethernet device.
269 * Pointer to Ethernet device.
271 * Request number to pass to ioctl().
273 * Interface request structure output buffer.
276 * 0 on success, a negative errno value otherwise and rte_errno is set.
279 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
281 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
288 ret = mlx5_get_ifname(dev, &ifr->ifr_name);
291 ret = ioctl(sock, req, ifr);
307 * Pointer to Ethernet device.
309 * MTU value output buffer.
312 * 0 on success, a negative errno value otherwise and rte_errno is set.
315 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
317 struct ifreq request;
318 int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
322 *mtu = request.ifr_mtu;
330 * Pointer to Ethernet device.
335 * 0 on success, a negative errno value otherwise and rte_errno is set.
338 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
340 struct ifreq request = { .ifr_mtu = mtu, };
342 return mlx5_ifreq(dev, SIOCSIFMTU, &request);
349 * Pointer to Ethernet device.
351 * Bitmask for flags that must remain untouched.
353 * Bitmask for flags to modify.
356 * 0 on success, a negative errno value otherwise and rte_errno is set.
359 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
361 struct ifreq request;
362 int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
366 request.ifr_flags &= keep;
367 request.ifr_flags |= flags & ~keep;
368 return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
372 * DPDK callback for Ethernet device configuration.
375 * Pointer to Ethernet device structure.
378 * 0 on success, a negative errno value otherwise and rte_errno is set.
381 mlx5_dev_configure(struct rte_eth_dev *dev)
383 struct mlx5_priv *priv = dev->data->dev_private;
384 unsigned int rxqs_n = dev->data->nb_rx_queues;
385 unsigned int txqs_n = dev->data->nb_tx_queues;
388 unsigned int reta_idx_n;
389 const uint8_t use_app_rss_key =
390 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
392 unsigned int lro_on = mlx5_lro_on(dev);
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));
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);
411 memcpy(priv->rss_conf.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;
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);
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;
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.
440 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
441 priv->config.ind_table_max_size :
443 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
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.
451 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
452 (*priv->reta_idx)[i] = j;
457 if (lro_on && priv->config.cqe_comp) {
458 /* CQE compressing is not supported for LRO CQEs. */
459 DRV_LOG(WARNING, "Rx CQE compression isn't supported with LRO");
460 priv->config.cqe_comp = 0;
462 ret = mlx5_proc_priv_init(dev);
469 * Sets default tuning parameters.
472 * Pointer to Ethernet device.
474 * Info structure output buffer.
477 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
479 struct mlx5_priv *priv = dev->data->dev_private;
481 /* Minimum CPU utilization. */
482 info->default_rxportconf.ring_size = 256;
483 info->default_txportconf.ring_size = 256;
484 info->default_rxportconf.burst_size = 64;
485 info->default_txportconf.burst_size = 64;
486 if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
487 info->default_rxportconf.nb_queues = 16;
488 info->default_txportconf.nb_queues = 16;
489 if (dev->data->nb_rx_queues > 2 ||
490 dev->data->nb_tx_queues > 2) {
491 /* Max Throughput. */
492 info->default_rxportconf.ring_size = 2048;
493 info->default_txportconf.ring_size = 2048;
496 info->default_rxportconf.nb_queues = 8;
497 info->default_txportconf.nb_queues = 8;
498 if (dev->data->nb_rx_queues > 2 ||
499 dev->data->nb_tx_queues > 2) {
500 /* Max Throughput. */
501 info->default_rxportconf.ring_size = 4096;
502 info->default_txportconf.ring_size = 4096;
508 * Sets tx mbuf limiting parameters.
511 * Pointer to Ethernet device.
513 * Info structure output buffer.
516 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
518 struct mlx5_priv *priv = dev->data->dev_private;
519 struct mlx5_dev_config *config = &priv->config;
523 inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
524 MLX5_SEND_DEF_INLINE_LEN :
525 (unsigned int)config->txq_inline_max;
526 assert(config->txq_inline_min >= 0);
527 inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
528 inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
529 MLX5_ESEG_MIN_INLINE_SIZE -
532 MLX5_WQE_DSEG_SIZE * 2);
533 nb_max = (MLX5_WQE_SIZE_MAX +
534 MLX5_ESEG_MIN_INLINE_SIZE -
538 inlen) / MLX5_WSEG_SIZE;
539 info->tx_desc_lim.nb_seg_max = nb_max;
540 info->tx_desc_lim.nb_mtu_seg_max = nb_max;
544 * DPDK callback to get information about the device.
547 * Pointer to Ethernet device structure.
549 * Info structure output buffer.
552 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
554 struct mlx5_priv *priv = dev->data->dev_private;
555 struct mlx5_dev_config *config = &priv->config;
558 /* FIXME: we should ask the device for these values. */
559 info->min_rx_bufsize = 32;
560 info->max_rx_pktlen = 65536;
562 * Since we need one CQ per QP, the limit is the minimum number
563 * between the two values.
565 max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
566 priv->sh->device_attr.orig_attr.max_qp);
567 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
570 info->max_rx_queues = max;
571 info->max_tx_queues = max;
572 info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
573 info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
574 info->rx_offload_capa = (mlx5_get_rx_port_offloads(dev) |
575 info->rx_queue_offload_capa);
576 info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
577 info->if_index = mlx5_ifindex(dev);
578 info->reta_size = priv->reta_idx_n ?
579 priv->reta_idx_n : config->ind_table_max_size;
580 info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
581 info->speed_capa = priv->link_speed_capa;
582 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
583 mlx5_set_default_params(dev, info);
584 mlx5_set_txlimit_params(dev, info);
585 info->switch_info.name = dev->data->name;
586 info->switch_info.domain_id = priv->domain_id;
587 info->switch_info.port_id = priv->representor_id;
588 if (priv->representor) {
589 unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
592 i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
594 struct mlx5_priv *opriv =
595 rte_eth_devices[port_id[i]].data->dev_private;
598 opriv->representor ||
599 opriv->domain_id != priv->domain_id)
602 * Override switch name with that of the master
605 info->switch_info.name = opriv->dev_data->name;
612 * Get device current raw clock counter
615 * Pointer to Ethernet device structure.
617 * Current raw clock counter of the device.
620 * 0 if the clock has correctly been read
621 * The value of errno in case of error
624 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
626 struct mlx5_priv *priv = dev->data->dev_private;
627 struct ibv_context *ctx = priv->sh->ctx;
628 struct ibv_values_ex values;
631 values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
632 err = mlx5_glue->query_rt_values_ex(ctx, &values);
634 DRV_LOG(WARNING, "Could not query the clock !");
637 *clock = values.raw_clock.tv_nsec;
642 * Get firmware version of a device.
645 * Ethernet device port.
647 * String output allocated by caller.
649 * Size of the output string, including terminating null byte.
652 * 0 on success, or the size of the non truncated string if too big.
654 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
656 struct mlx5_priv *priv = dev->data->dev_private;
657 struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
658 size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
663 strlcpy(fw_ver, attr->fw_ver, fw_size);
668 * Get supported packet types.
671 * Pointer to Ethernet device structure.
674 * A pointer to the supported Packet types array.
677 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
679 static const uint32_t ptypes[] = {
680 /* refers to rxq_cq_to_pkt_type() */
682 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
683 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
684 RTE_PTYPE_L4_NONFRAG,
688 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
689 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
690 RTE_PTYPE_INNER_L4_NONFRAG,
691 RTE_PTYPE_INNER_L4_FRAG,
692 RTE_PTYPE_INNER_L4_TCP,
693 RTE_PTYPE_INNER_L4_UDP,
697 if (dev->rx_pkt_burst == mlx5_rx_burst ||
698 dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
699 dev->rx_pkt_burst == mlx5_rx_burst_vec)
705 * Retrieve the master device for representor in the same switch domain.
708 * Pointer to representor Ethernet device structure.
711 * Master device structure on success, NULL otherwise.
714 static struct rte_eth_dev *
715 mlx5_find_master_dev(struct rte_eth_dev *dev)
717 struct mlx5_priv *priv;
721 priv = dev->data->dev_private;
722 domain_id = priv->domain_id;
723 assert(priv->representor);
724 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
725 priv = rte_eth_devices[port_id].data->dev_private;
728 priv->domain_id == domain_id)
729 return &rte_eth_devices[port_id];
735 * DPDK callback to retrieve physical link information.
738 * Pointer to Ethernet device structure.
740 * Storage for current link status.
743 * 0 on success, a negative errno value otherwise and rte_errno is set.
746 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
747 struct rte_eth_link *link)
749 struct mlx5_priv *priv = dev->data->dev_private;
750 struct ethtool_cmd edata = {
751 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
754 struct rte_eth_link dev_link;
758 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
760 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
761 dev->data->port_id, strerror(rte_errno));
764 dev_link = (struct rte_eth_link) {
765 .link_status = ((ifr.ifr_flags & IFF_UP) &&
766 (ifr.ifr_flags & IFF_RUNNING)),
768 ifr = (struct ifreq) {
769 .ifr_data = (void *)&edata,
771 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
773 if (ret == -ENOTSUP && priv->representor) {
774 struct rte_eth_dev *master;
777 * For representors we can try to inherit link
778 * settings from the master device. Actually
779 * link settings do not make a lot of sense
780 * for representors due to missing physical
781 * link. The old kernel drivers supported
782 * emulated settings query for representors,
783 * the new ones do not, so we have to add
784 * this code for compatibility issues.
786 master = mlx5_find_master_dev(dev);
788 ifr = (struct ifreq) {
789 .ifr_data = (void *)&edata,
791 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
796 "port %u ioctl(SIOCETHTOOL,"
797 " ETHTOOL_GSET) failed: %s",
798 dev->data->port_id, strerror(rte_errno));
802 link_speed = ethtool_cmd_speed(&edata);
803 if (link_speed == -1)
804 dev_link.link_speed = ETH_SPEED_NUM_NONE;
806 dev_link.link_speed = link_speed;
807 priv->link_speed_capa = 0;
808 if (edata.supported & SUPPORTED_Autoneg)
809 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
810 if (edata.supported & (SUPPORTED_1000baseT_Full |
811 SUPPORTED_1000baseKX_Full))
812 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
813 if (edata.supported & SUPPORTED_10000baseKR_Full)
814 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
815 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
816 SUPPORTED_40000baseCR4_Full |
817 SUPPORTED_40000baseSR4_Full |
818 SUPPORTED_40000baseLR4_Full))
819 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
820 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
821 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
822 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
823 ETH_LINK_SPEED_FIXED);
824 if (((dev_link.link_speed && !dev_link.link_status) ||
825 (!dev_link.link_speed && dev_link.link_status))) {
834 * Retrieve physical link information (unlocked version using new ioctl).
837 * Pointer to Ethernet device structure.
839 * Storage for current link status.
842 * 0 on success, a negative errno value otherwise and rte_errno is set.
845 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
846 struct rte_eth_link *link)
849 struct mlx5_priv *priv = dev->data->dev_private;
850 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
852 struct rte_eth_link dev_link;
853 struct rte_eth_dev *master = NULL;
857 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
859 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
860 dev->data->port_id, strerror(rte_errno));
863 dev_link = (struct rte_eth_link) {
864 .link_status = ((ifr.ifr_flags & IFF_UP) &&
865 (ifr.ifr_flags & IFF_RUNNING)),
867 ifr = (struct ifreq) {
868 .ifr_data = (void *)&gcmd,
870 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
872 if (ret == -ENOTSUP && priv->representor) {
874 * For representors we can try to inherit link
875 * settings from the master device. Actually
876 * link settings do not make a lot of sense
877 * for representors due to missing physical
878 * link. The old kernel drivers supported
879 * emulated settings query for representors,
880 * the new ones do not, so we have to add
881 * this code for compatibility issues.
883 master = mlx5_find_master_dev(dev);
885 ifr = (struct ifreq) {
886 .ifr_data = (void *)&gcmd,
888 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
893 "port %u ioctl(SIOCETHTOOL,"
894 " ETHTOOL_GLINKSETTINGS) failed: %s",
895 dev->data->port_id, strerror(rte_errno));
900 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
902 alignas(struct ethtool_link_settings)
903 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
904 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
905 struct ethtool_link_settings *ecmd = (void *)data;
908 ifr.ifr_data = (void *)ecmd;
909 ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
912 "port %u ioctl(SIOCETHTOOL,"
913 "ETHTOOL_GLINKSETTINGS) failed: %s",
914 dev->data->port_id, strerror(rte_errno));
917 dev_link.link_speed = ecmd->speed;
918 sc = ecmd->link_mode_masks[0] |
919 ((uint64_t)ecmd->link_mode_masks[1] << 32);
920 priv->link_speed_capa = 0;
921 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
922 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
923 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
924 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
925 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
926 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
927 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
928 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
929 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
930 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
931 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
932 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
933 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
934 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
935 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
936 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
937 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
938 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
939 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
940 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
941 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
942 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
943 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
944 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
945 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
946 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
947 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
948 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
949 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
950 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
951 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
952 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
953 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
954 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
955 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
956 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
957 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
958 ETH_LINK_SPEED_FIXED);
959 if (((dev_link.link_speed && !dev_link.link_status) ||
960 (!dev_link.link_speed && dev_link.link_status))) {
969 * DPDK callback to retrieve physical link information.
972 * Pointer to Ethernet device structure.
973 * @param wait_to_complete
974 * Wait for request completion.
977 * 0 if link status was not updated, positive if it was, a negative errno
978 * value otherwise and rte_errno is set.
981 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
984 struct rte_eth_link dev_link;
985 time_t start_time = time(NULL);
988 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
990 ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
993 /* Handle wait to complete situation. */
994 if (wait_to_complete && ret == -EAGAIN) {
995 if (abs((int)difftime(time(NULL), start_time)) <
996 MLX5_LINK_STATUS_TIMEOUT) {
1003 } else if (ret < 0) {
1006 } while (wait_to_complete);
1007 ret = !!memcmp(&dev->data->dev_link, &dev_link,
1008 sizeof(struct rte_eth_link));
1009 dev->data->dev_link = dev_link;
1014 * DPDK callback to change the MTU.
1017 * Pointer to Ethernet device structure.
1022 * 0 on success, a negative errno value otherwise and rte_errno is set.
1025 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1027 struct mlx5_priv *priv = dev->data->dev_private;
1028 uint16_t kern_mtu = 0;
1031 ret = mlx5_get_mtu(dev, &kern_mtu);
1034 /* Set kernel interface MTU first. */
1035 ret = mlx5_set_mtu(dev, mtu);
1038 ret = mlx5_get_mtu(dev, &kern_mtu);
1041 if (kern_mtu == mtu) {
1043 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1044 dev->data->port_id, mtu);
1052 * DPDK callback to get flow control status.
1055 * Pointer to Ethernet device structure.
1056 * @param[out] fc_conf
1057 * Flow control output buffer.
1060 * 0 on success, a negative errno value otherwise and rte_errno is set.
1063 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1066 struct ethtool_pauseparam ethpause = {
1067 .cmd = ETHTOOL_GPAUSEPARAM
1071 ifr.ifr_data = (void *)ðpause;
1072 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1075 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1077 dev->data->port_id, strerror(rte_errno));
1080 fc_conf->autoneg = ethpause.autoneg;
1081 if (ethpause.rx_pause && ethpause.tx_pause)
1082 fc_conf->mode = RTE_FC_FULL;
1083 else if (ethpause.rx_pause)
1084 fc_conf->mode = RTE_FC_RX_PAUSE;
1085 else if (ethpause.tx_pause)
1086 fc_conf->mode = RTE_FC_TX_PAUSE;
1088 fc_conf->mode = RTE_FC_NONE;
1093 * DPDK callback to modify flow control parameters.
1096 * Pointer to Ethernet device structure.
1097 * @param[in] fc_conf
1098 * Flow control parameters.
1101 * 0 on success, a negative errno value otherwise and rte_errno is set.
1104 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1107 struct ethtool_pauseparam ethpause = {
1108 .cmd = ETHTOOL_SPAUSEPARAM
1112 ifr.ifr_data = (void *)ðpause;
1113 ethpause.autoneg = fc_conf->autoneg;
1114 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1115 (fc_conf->mode & RTE_FC_RX_PAUSE))
1116 ethpause.rx_pause = 1;
1118 ethpause.rx_pause = 0;
1120 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1121 (fc_conf->mode & RTE_FC_TX_PAUSE))
1122 ethpause.tx_pause = 1;
1124 ethpause.tx_pause = 0;
1125 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1128 "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1130 dev->data->port_id, strerror(rte_errno));
1137 * Get PCI information from struct ibv_device.
1140 * Pointer to Ethernet device structure.
1141 * @param[out] pci_addr
1142 * PCI bus address output buffer.
1145 * 0 on success, a negative errno value otherwise and rte_errno is set.
1148 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1149 struct rte_pci_addr *pci_addr)
1153 MKSTR(path, "%s/device/uevent", device->ibdev_path);
1155 file = fopen(path, "rb");
1160 while (fgets(line, sizeof(line), file) == line) {
1161 size_t len = strlen(line);
1164 /* Truncate long lines. */
1165 if (len == (sizeof(line) - 1))
1166 while (line[(len - 1)] != '\n') {
1170 line[(len - 1)] = ret;
1172 /* Extract information. */
1175 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1179 &pci_addr->function) == 4) {
1189 * Handle asynchronous removal event for entire multiport device.
1192 * Infiniband device shared context.
1195 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1199 for (i = 0; i < sh->max_port; ++i) {
1200 struct rte_eth_dev *dev;
1202 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1204 * Or not existing port either no
1205 * handler installed for this port.
1209 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1211 if (dev->data->dev_conf.intr_conf.rmv)
1212 _rte_eth_dev_callback_process
1213 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1218 * Handle shared asynchronous events the NIC (removal event
1219 * and link status change). Supports multiport IB device.
1222 * Callback argument.
1225 mlx5_dev_interrupt_handler(void *cb_arg)
1227 struct mlx5_ibv_shared *sh = cb_arg;
1228 struct ibv_async_event event;
1230 /* Read all message from the IB device and acknowledge them. */
1232 struct rte_eth_dev *dev;
1235 if (mlx5_glue->get_async_event(sh->ctx, &event))
1237 /* Retrieve and check IB port index. */
1238 tmp = (uint32_t)event.element.port_num;
1239 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1241 * The DEVICE_FATAL event is called once for
1242 * entire device without port specifying.
1243 * We should notify all existing ports.
1245 mlx5_glue->ack_async_event(&event);
1246 mlx5_dev_interrupt_device_fatal(sh);
1249 assert(tmp && (tmp <= sh->max_port));
1251 /* Unsupported devive level event. */
1252 mlx5_glue->ack_async_event(&event);
1254 "unsupported common event (type %d)",
1258 if (tmp > sh->max_port) {
1259 /* Invalid IB port index. */
1260 mlx5_glue->ack_async_event(&event);
1262 "cannot handle an event (type %d)"
1263 "due to invalid IB port index (%u)",
1264 event.event_type, tmp);
1267 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1268 /* No handler installed. */
1269 mlx5_glue->ack_async_event(&event);
1271 "cannot handle an event (type %d)"
1272 "due to no handler installed for port %u",
1273 event.event_type, tmp);
1276 /* Retrieve ethernet device descriptor. */
1277 tmp = sh->port[tmp - 1].ih_port_id;
1278 dev = &rte_eth_devices[tmp];
1280 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1281 event.event_type == IBV_EVENT_PORT_ERR) &&
1282 dev->data->dev_conf.intr_conf.lsc) {
1283 mlx5_glue->ack_async_event(&event);
1284 if (mlx5_link_update(dev, 0) == -EAGAIN) {
1288 _rte_eth_dev_callback_process
1289 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1293 "port %u cannot handle an unknown event (type %d)",
1294 dev->data->port_id, event.event_type);
1295 mlx5_glue->ack_async_event(&event);
1300 * Unregister callback handler safely. The handler may be active
1301 * while we are trying to unregister it, in this case code -EAGAIN
1302 * is returned by rte_intr_callback_unregister(). This routine checks
1303 * the return code and tries to unregister handler again.
1308 * pointer to callback routine
1310 * opaque callback parameter
1313 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1314 rte_intr_callback_fn cb_fn, void *cb_arg)
1317 * Try to reduce timeout management overhead by not calling
1318 * the timer related routines on the first iteration. If the
1319 * unregistering succeeds on first call there will be no
1320 * timer calls at all.
1328 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1331 if (ret != -EAGAIN) {
1332 DRV_LOG(INFO, "failed to unregister interrupt"
1333 " handler (error: %d)", ret);
1338 struct timespec onems;
1340 /* Wait one millisecond and try again. */
1342 onems.tv_nsec = NS_PER_S / MS_PER_S;
1343 nanosleep(&onems, 0);
1344 /* Check whether one second elapsed. */
1345 if ((rte_get_timer_cycles() - start) <= twait)
1349 * We get the amount of timer ticks for one second.
1350 * If this amount elapsed it means we spent one
1351 * second in waiting. This branch is executed once
1352 * on first iteration.
1354 twait = rte_get_timer_hz();
1358 * Timeout elapsed, show message (once a second) and retry.
1359 * We have no other acceptable option here, if we ignore
1360 * the unregistering return code the handler will not
1361 * be unregistered, fd will be closed and we may get the
1362 * crush. Hanging and messaging in the loop seems not to be
1365 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1366 start = rte_get_timer_cycles();
1371 * Handle DEVX interrupts from the NIC.
1372 * This function is probably called from the DPDK host thread.
1375 * Callback argument.
1378 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1380 #ifndef HAVE_IBV_DEVX_ASYNC
1384 struct mlx5_ibv_shared *sh = cb_arg;
1386 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1387 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1388 MLX5_ST_SZ_BYTES(traffic_counter) +
1389 sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1391 uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1393 while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1396 mlx5_flow_async_pool_query_handle
1397 (sh, (uint64_t)out.cmd_resp.wr_id,
1398 mlx5_devx_get_out_command_status(buf));
1399 #endif /* HAVE_IBV_DEVX_ASYNC */
1403 * Uninstall shared asynchronous device events handler.
1404 * This function is implemented to support event sharing
1405 * between multiple ports of single IB device.
1408 * Pointer to Ethernet device.
1411 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1413 struct mlx5_priv *priv = dev->data->dev_private;
1414 struct mlx5_ibv_shared *sh = priv->sh;
1416 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1418 pthread_mutex_lock(&sh->intr_mutex);
1419 assert(priv->ibv_port);
1420 assert(priv->ibv_port <= sh->max_port);
1421 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1422 if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1424 assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1425 (uint32_t)dev->data->port_id);
1426 assert(sh->intr_cnt);
1427 sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1428 if (!sh->intr_cnt || --sh->intr_cnt)
1430 mlx5_intr_callback_unregister(&sh->intr_handle,
1431 mlx5_dev_interrupt_handler, sh);
1432 sh->intr_handle.fd = 0;
1433 sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1434 if (sh->intr_handle_devx.fd) {
1435 rte_intr_callback_unregister(&sh->intr_handle_devx,
1436 mlx5_dev_interrupt_handler_devx,
1438 sh->intr_handle_devx.fd = 0;
1439 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1441 if (sh->devx_comp) {
1442 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1443 sh->devx_comp = NULL;
1446 pthread_mutex_unlock(&sh->intr_mutex);
1450 * Install shared asynchronous device events handler.
1451 * This function is implemented to support event sharing
1452 * between multiple ports of single IB device.
1455 * Pointer to Ethernet device.
1458 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1460 struct mlx5_priv *priv = dev->data->dev_private;
1461 struct mlx5_ibv_shared *sh = priv->sh;
1465 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1467 pthread_mutex_lock(&sh->intr_mutex);
1468 assert(priv->ibv_port);
1469 assert(priv->ibv_port <= sh->max_port);
1470 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1471 if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1472 /* The handler is already installed for this port. */
1473 assert(sh->intr_cnt);
1476 sh->port[priv->ibv_port - 1].ih_port_id = (uint32_t)dev->data->port_id;
1481 /* No shared handler installed. */
1482 assert(sh->ctx->async_fd > 0);
1483 flags = fcntl(sh->ctx->async_fd, F_GETFL);
1484 ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1486 DRV_LOG(INFO, "failed to change file descriptor"
1487 " async event queue");
1490 sh->intr_handle.fd = sh->ctx->async_fd;
1491 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1492 rte_intr_callback_register(&sh->intr_handle,
1493 mlx5_dev_interrupt_handler, sh);
1494 if (priv->config.devx) {
1495 #ifndef HAVE_IBV_DEVX_ASYNC
1496 goto error_unregister;
1498 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1499 if (sh->devx_comp) {
1500 flags = fcntl(sh->devx_comp->fd, F_GETFL);
1501 ret = fcntl(sh->devx_comp->fd, F_SETFL,
1502 flags | O_NONBLOCK);
1504 DRV_LOG(INFO, "failed to change file descriptor"
1505 " devx async event queue");
1506 goto error_unregister;
1508 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1509 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1510 rte_intr_callback_register
1511 (&sh->intr_handle_devx,
1512 mlx5_dev_interrupt_handler_devx, sh);
1514 DRV_LOG(INFO, "failed to create devx async command "
1516 goto error_unregister;
1518 #endif /* HAVE_IBV_DEVX_ASYNC */
1523 rte_intr_callback_unregister(&sh->intr_handle,
1524 mlx5_dev_interrupt_handler, sh);
1526 /* Indicate there will be no interrupts. */
1527 dev->data->dev_conf.intr_conf.lsc = 0;
1528 dev->data->dev_conf.intr_conf.rmv = 0;
1529 sh->intr_handle.fd = 0;
1530 sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1531 sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1533 pthread_mutex_unlock(&sh->intr_mutex);
1537 * Uninstall interrupt handler.
1540 * Pointer to Ethernet device.
1543 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1545 mlx5_dev_shared_handler_uninstall(dev);
1549 * Install interrupt handler.
1552 * Pointer to Ethernet device.
1555 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1557 mlx5_dev_shared_handler_install(dev);
1561 * DPDK callback to bring the link DOWN.
1564 * Pointer to Ethernet device structure.
1567 * 0 on success, a negative errno value otherwise and rte_errno is set.
1570 mlx5_set_link_down(struct rte_eth_dev *dev)
1572 return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1576 * DPDK callback to bring the link UP.
1579 * Pointer to Ethernet device structure.
1582 * 0 on success, a negative errno value otherwise and rte_errno is set.
1585 mlx5_set_link_up(struct rte_eth_dev *dev)
1587 return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1591 * Configure the RX function to use.
1594 * Pointer to private data structure.
1597 * Pointer to selected Rx burst function.
1600 mlx5_select_rx_function(struct rte_eth_dev *dev)
1602 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1604 assert(dev != NULL);
1605 if (mlx5_check_vec_rx_support(dev) > 0) {
1606 rx_pkt_burst = mlx5_rx_burst_vec;
1607 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1608 dev->data->port_id);
1609 } else if (mlx5_mprq_enabled(dev)) {
1610 rx_pkt_burst = mlx5_rx_burst_mprq;
1612 return rx_pkt_burst;
1616 * Check if mlx5 device was removed.
1619 * Pointer to Ethernet device structure.
1622 * 1 when device is removed, otherwise 0.
1625 mlx5_is_removed(struct rte_eth_dev *dev)
1627 struct ibv_device_attr device_attr;
1628 struct mlx5_priv *priv = dev->data->dev_private;
1630 if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1636 * Get port ID list of mlx5 instances sharing a common device.
1639 * Device to look for.
1640 * @param[out] port_list
1641 * Result buffer for collected port IDs.
1642 * @param port_list_n
1643 * Maximum number of entries in result buffer. If 0, @p port_list can be
1647 * Number of matching instances regardless of the @p port_list_n
1648 * parameter, 0 if none were found.
1651 mlx5_dev_to_port_id(const struct rte_device *dev, uint16_t *port_list,
1652 unsigned int port_list_n)
1657 RTE_ETH_FOREACH_DEV_OF(id, dev) {
1658 if (n < port_list_n)
1666 * Get the E-Switch domain id this port belongs to.
1670 * @param[out] es_domain_id
1671 * E-Switch domain id.
1672 * @param[out] es_port_id
1673 * The port id of the port in the E-Switch.
1676 * 0 on success, a negative errno value otherwise and rte_errno is set.
1679 mlx5_port_to_eswitch_info(uint16_t port,
1680 uint16_t *es_domain_id, uint16_t *es_port_id)
1682 struct rte_eth_dev *dev;
1683 struct mlx5_priv *priv;
1685 if (port >= RTE_MAX_ETHPORTS) {
1689 if (!rte_eth_dev_is_valid_port(port)) {
1693 dev = &rte_eth_devices[port];
1694 priv = dev->data->dev_private;
1695 if (!(priv->representor || priv->master)) {
1700 *es_domain_id = priv->domain_id;
1702 *es_port_id = priv->vport_id;
1707 * Get switch information associated with network interface.
1710 * Network interface index.
1712 * Switch information object, populated in case of success.
1715 * 0 on success, a negative errno value otherwise and rte_errno is set.
1718 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1720 char ifname[IF_NAMESIZE];
1721 char port_name[IF_NAMESIZE];
1723 struct mlx5_switch_info data = {
1726 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1731 bool port_switch_id_set = false;
1732 bool device_dir = false;
1736 if (!if_indextoname(ifindex, ifname)) {
1741 MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1743 MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1745 MKSTR(pci_device, "/sys/class/net/%s/device",
1748 file = fopen(phys_port_name, "rb");
1750 ret = fscanf(file, "%s", port_name);
1753 mlx5_translate_port_name(port_name, &data);
1755 file = fopen(phys_switch_id, "rb");
1760 port_switch_id_set =
1761 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1764 dir = opendir(pci_device);
1769 if (port_switch_id_set) {
1770 /* We have some E-Switch configuration. */
1771 mlx5_sysfs_check_switch_info(device_dir, &data);
1774 assert(!(data.master && data.representor));
1775 if (data.master && data.representor) {
1776 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1777 " and as representor", ifindex);
1785 * Analyze gathered port parameters via Netlink to recognize master
1786 * and representor devices for E-Switch configuration.
1788 * @param[in] num_vf_set
1789 * flag of presence of number of VFs port attribute.
1790 * @param[inout] switch_info
1791 * Port information, including port name as a number and port name
1792 * type if recognized
1795 * master and representor flags are set in switch_info according to
1796 * recognized parameters (if any).
1799 mlx5_nl_check_switch_info(bool num_vf_set,
1800 struct mlx5_switch_info *switch_info)
1802 switch (switch_info->name_type) {
1803 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1805 * Name is not recognized, assume the master,
1806 * check the number of VFs key presence.
1808 switch_info->master = num_vf_set;
1810 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1812 * Name is not set, this assumes the legacy naming
1813 * schema for master, just check if there is a
1814 * number of VFs key.
1816 switch_info->master = num_vf_set;
1818 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1819 /* New uplink naming schema recognized. */
1820 switch_info->master = 1;
1822 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1823 /* Legacy representors naming schema. */
1824 switch_info->representor = !num_vf_set;
1826 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1827 /* New representors naming schema. */
1828 switch_info->representor = 1;
1834 * Analyze gathered port parameters via sysfs to recognize master
1835 * and representor devices for E-Switch configuration.
1837 * @param[in] device_dir
1838 * flag of presence of "device" directory under port device key.
1839 * @param[inout] switch_info
1840 * Port information, including port name as a number and port name
1841 * type if recognized
1844 * master and representor flags are set in switch_info according to
1845 * recognized parameters (if any).
1848 mlx5_sysfs_check_switch_info(bool device_dir,
1849 struct mlx5_switch_info *switch_info)
1851 switch (switch_info->name_type) {
1852 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1854 * Name is not recognized, assume the master,
1855 * check the device directory presence.
1857 switch_info->master = device_dir;
1859 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1861 * Name is not set, this assumes the legacy naming
1862 * schema for master, just check if there is
1863 * a device directory.
1865 switch_info->master = device_dir;
1867 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1868 /* New uplink naming schema recognized. */
1869 switch_info->master = 1;
1871 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1872 /* Legacy representors naming schema. */
1873 switch_info->representor = !device_dir;
1875 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1876 /* New representors naming schema. */
1877 switch_info->representor = 1;
1883 * Extract port name, as a number, from sysfs or netlink information.
1885 * @param[in] port_name_in
1886 * String representing the port name.
1887 * @param[out] port_info_out
1888 * Port information, including port name as a number and port name
1889 * type if recognized
1892 * port_name field set according to recognized name format.
1895 mlx5_translate_port_name(const char *port_name_in,
1896 struct mlx5_switch_info *port_info_out)
1898 char pf_c1, pf_c2, vf_c1, vf_c2;
1903 * Check for port-name as a string of the form pf0vf0
1904 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
1906 sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
1907 &pf_c1, &pf_c2, &port_info_out->pf_num,
1908 &vf_c1, &vf_c2, &port_info_out->port_name);
1909 if (sc_items == 6 &&
1910 pf_c1 == 'p' && pf_c2 == 'f' &&
1911 vf_c1 == 'v' && vf_c2 == 'f') {
1912 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
1916 * Check for port-name as a string of the form p0
1917 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
1919 sc_items = sscanf(port_name_in, "%c%d",
1920 &pf_c1, &port_info_out->port_name);
1921 if (sc_items == 2 && pf_c1 == 'p') {
1922 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1925 /* Check for port-name as a number (support kernel ver < 5.0 */
1927 port_info_out->port_name = strtol(port_name_in, &end, 0);
1929 (size_t)(end - port_name_in) == strlen(port_name_in)) {
1930 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
1933 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;