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;
393 if (use_app_rss_key &&
394 (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
395 MLX5_RSS_HASH_KEY_LEN)) {
396 DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
397 dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
401 priv->rss_conf.rss_key =
402 rte_realloc(priv->rss_conf.rss_key,
403 MLX5_RSS_HASH_KEY_LEN, 0);
404 if (!priv->rss_conf.rss_key) {
405 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
406 dev->data->port_id, rxqs_n);
410 memcpy(priv->rss_conf.rss_key,
412 dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
413 rss_hash_default_key,
414 MLX5_RSS_HASH_KEY_LEN);
415 priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
416 priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
417 priv->rxqs = (void *)dev->data->rx_queues;
418 priv->txqs = (void *)dev->data->tx_queues;
419 if (txqs_n != priv->txqs_n) {
420 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
421 dev->data->port_id, priv->txqs_n, txqs_n);
422 priv->txqs_n = txqs_n;
424 if (rxqs_n > priv->config.ind_table_max_size) {
425 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
426 dev->data->port_id, rxqs_n);
430 if (rxqs_n != priv->rxqs_n) {
431 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
432 dev->data->port_id, priv->rxqs_n, rxqs_n);
433 priv->rxqs_n = rxqs_n;
435 * If the requested number of RX queues is not a power of two,
436 * use the maximum indirection table size for better balancing.
437 * The result is always rounded to the next power of two.
439 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
440 priv->config.ind_table_max_size :
442 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
446 * When the number of RX queues is not a power of two,
447 * the remaining table entries are padded with reused WQs
448 * and hashes are not spread uniformly.
450 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
451 (*priv->reta_idx)[i] = j;
456 ret = mlx5_proc_priv_init(dev);
463 * Sets default tuning parameters.
466 * Pointer to Ethernet device.
468 * Info structure output buffer.
471 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
473 struct mlx5_priv *priv = dev->data->dev_private;
475 /* Minimum CPU utilization. */
476 info->default_rxportconf.ring_size = 256;
477 info->default_txportconf.ring_size = 256;
478 info->default_rxportconf.burst_size = 64;
479 info->default_txportconf.burst_size = 64;
480 if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
481 info->default_rxportconf.nb_queues = 16;
482 info->default_txportconf.nb_queues = 16;
483 if (dev->data->nb_rx_queues > 2 ||
484 dev->data->nb_tx_queues > 2) {
485 /* Max Throughput. */
486 info->default_rxportconf.ring_size = 2048;
487 info->default_txportconf.ring_size = 2048;
490 info->default_rxportconf.nb_queues = 8;
491 info->default_txportconf.nb_queues = 8;
492 if (dev->data->nb_rx_queues > 2 ||
493 dev->data->nb_tx_queues > 2) {
494 /* Max Throughput. */
495 info->default_rxportconf.ring_size = 4096;
496 info->default_txportconf.ring_size = 4096;
502 * Sets tx mbuf limiting parameters.
505 * Pointer to Ethernet device.
507 * Info structure output buffer.
510 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
512 struct mlx5_priv *priv = dev->data->dev_private;
513 struct mlx5_dev_config *config = &priv->config;
517 inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
518 MLX5_SEND_DEF_INLINE_LEN :
519 (unsigned int)config->txq_inline_max;
520 assert(config->txq_inline_min >= 0);
521 inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
522 inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
523 MLX5_ESEG_MIN_INLINE_SIZE -
526 MLX5_WQE_DSEG_SIZE * 2);
527 nb_max = (MLX5_WQE_SIZE_MAX +
528 MLX5_ESEG_MIN_INLINE_SIZE -
532 inlen) / MLX5_WSEG_SIZE;
533 info->tx_desc_lim.nb_seg_max = nb_max;
534 info->tx_desc_lim.nb_mtu_seg_max = nb_max;
538 * DPDK callback to get information about the device.
541 * Pointer to Ethernet device structure.
543 * Info structure output buffer.
546 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
548 struct mlx5_priv *priv = dev->data->dev_private;
549 struct mlx5_dev_config *config = &priv->config;
552 /* FIXME: we should ask the device for these values. */
553 info->min_rx_bufsize = 32;
554 info->max_rx_pktlen = 65536;
556 * Since we need one CQ per QP, the limit is the minimum number
557 * between the two values.
559 max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
560 priv->sh->device_attr.orig_attr.max_qp);
561 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
564 info->max_rx_queues = max;
565 info->max_tx_queues = max;
566 info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
567 info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
568 info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
569 info->rx_queue_offload_capa);
570 info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
571 info->if_index = mlx5_ifindex(dev);
572 info->reta_size = priv->reta_idx_n ?
573 priv->reta_idx_n : config->ind_table_max_size;
574 info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
575 info->speed_capa = priv->link_speed_capa;
576 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
577 mlx5_set_default_params(dev, info);
578 mlx5_set_txlimit_params(dev, info);
579 info->switch_info.name = dev->data->name;
580 info->switch_info.domain_id = priv->domain_id;
581 info->switch_info.port_id = priv->representor_id;
582 if (priv->representor) {
583 unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
586 i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
588 struct mlx5_priv *opriv =
589 rte_eth_devices[port_id[i]].data->dev_private;
592 opriv->representor ||
593 opriv->domain_id != priv->domain_id)
596 * Override switch name with that of the master
599 info->switch_info.name = opriv->dev_data->name;
608 * Get device current raw clock counter
611 * Pointer to Ethernet device structure.
613 * Current raw clock counter of the device.
616 * 0 if the clock has correctly been read
617 * The value of errno in case of error
620 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
622 struct mlx5_priv *priv = dev->data->dev_private;
623 struct ibv_context *ctx = priv->sh->ctx;
624 struct ibv_values_ex values;
627 values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
628 err = mlx5_glue->query_rt_values_ex(ctx, &values);
630 DRV_LOG(WARNING, "Could not query the clock !");
633 *clock = values.raw_clock.tv_nsec;
638 * Get firmware version of a device.
641 * Ethernet device port.
643 * String output allocated by caller.
645 * Size of the output string, including terminating null byte.
648 * 0 on success, or the size of the non truncated string if too big.
650 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
652 struct mlx5_priv *priv = dev->data->dev_private;
653 struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
654 size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
659 strlcpy(fw_ver, attr->fw_ver, fw_size);
664 * Get supported packet types.
667 * Pointer to Ethernet device structure.
670 * A pointer to the supported Packet types array.
673 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
675 static const uint32_t ptypes[] = {
676 /* refers to rxq_cq_to_pkt_type() */
678 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
679 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
680 RTE_PTYPE_L4_NONFRAG,
684 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
685 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
686 RTE_PTYPE_INNER_L4_NONFRAG,
687 RTE_PTYPE_INNER_L4_FRAG,
688 RTE_PTYPE_INNER_L4_TCP,
689 RTE_PTYPE_INNER_L4_UDP,
693 if (dev->rx_pkt_burst == mlx5_rx_burst ||
694 dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
695 dev->rx_pkt_burst == mlx5_rx_burst_vec)
701 * Retrieve the master device for representor in the same switch domain.
704 * Pointer to representor Ethernet device structure.
707 * Master device structure on success, NULL otherwise.
710 static struct rte_eth_dev *
711 mlx5_find_master_dev(struct rte_eth_dev *dev)
713 struct mlx5_priv *priv;
717 priv = dev->data->dev_private;
718 domain_id = priv->domain_id;
719 assert(priv->representor);
720 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
721 priv = rte_eth_devices[port_id].data->dev_private;
724 priv->domain_id == domain_id)
725 return &rte_eth_devices[port_id];
731 * DPDK callback to retrieve physical link information.
734 * Pointer to Ethernet device structure.
736 * Storage for current link status.
739 * 0 on success, a negative errno value otherwise and rte_errno is set.
742 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
743 struct rte_eth_link *link)
745 struct mlx5_priv *priv = dev->data->dev_private;
746 struct ethtool_cmd edata = {
747 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
750 struct rte_eth_link dev_link;
754 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
756 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
757 dev->data->port_id, strerror(rte_errno));
760 dev_link = (struct rte_eth_link) {
761 .link_status = ((ifr.ifr_flags & IFF_UP) &&
762 (ifr.ifr_flags & IFF_RUNNING)),
764 ifr = (struct ifreq) {
765 .ifr_data = (void *)&edata,
767 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
769 if (ret == -ENOTSUP && priv->representor) {
770 struct rte_eth_dev *master;
773 * For representors we can try to inherit link
774 * settings from the master device. Actually
775 * link settings do not make a lot of sense
776 * for representors due to missing physical
777 * link. The old kernel drivers supported
778 * emulated settings query for representors,
779 * the new ones do not, so we have to add
780 * this code for compatibility issues.
782 master = mlx5_find_master_dev(dev);
784 ifr = (struct ifreq) {
785 .ifr_data = (void *)&edata,
787 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
792 "port %u ioctl(SIOCETHTOOL,"
793 " ETHTOOL_GSET) failed: %s",
794 dev->data->port_id, strerror(rte_errno));
798 link_speed = ethtool_cmd_speed(&edata);
799 if (link_speed == -1)
800 dev_link.link_speed = ETH_SPEED_NUM_NONE;
802 dev_link.link_speed = link_speed;
803 priv->link_speed_capa = 0;
804 if (edata.supported & SUPPORTED_Autoneg)
805 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
806 if (edata.supported & (SUPPORTED_1000baseT_Full |
807 SUPPORTED_1000baseKX_Full))
808 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
809 if (edata.supported & SUPPORTED_10000baseKR_Full)
810 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
811 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
812 SUPPORTED_40000baseCR4_Full |
813 SUPPORTED_40000baseSR4_Full |
814 SUPPORTED_40000baseLR4_Full))
815 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
816 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
817 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
818 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
819 ETH_LINK_SPEED_FIXED);
820 if (((dev_link.link_speed && !dev_link.link_status) ||
821 (!dev_link.link_speed && dev_link.link_status))) {
830 * Retrieve physical link information (unlocked version using new ioctl).
833 * Pointer to Ethernet device structure.
835 * Storage for current link status.
838 * 0 on success, a negative errno value otherwise and rte_errno is set.
841 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
842 struct rte_eth_link *link)
845 struct mlx5_priv *priv = dev->data->dev_private;
846 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
848 struct rte_eth_link dev_link;
849 struct rte_eth_dev *master = NULL;
853 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
855 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
856 dev->data->port_id, strerror(rte_errno));
859 dev_link = (struct rte_eth_link) {
860 .link_status = ((ifr.ifr_flags & IFF_UP) &&
861 (ifr.ifr_flags & IFF_RUNNING)),
863 ifr = (struct ifreq) {
864 .ifr_data = (void *)&gcmd,
866 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
868 if (ret == -ENOTSUP && priv->representor) {
870 * For representors we can try to inherit link
871 * settings from the master device. Actually
872 * link settings do not make a lot of sense
873 * for representors due to missing physical
874 * link. The old kernel drivers supported
875 * emulated settings query for representors,
876 * the new ones do not, so we have to add
877 * this code for compatibility issues.
879 master = mlx5_find_master_dev(dev);
881 ifr = (struct ifreq) {
882 .ifr_data = (void *)&gcmd,
884 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
889 "port %u ioctl(SIOCETHTOOL,"
890 " ETHTOOL_GLINKSETTINGS) failed: %s",
891 dev->data->port_id, strerror(rte_errno));
896 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
898 alignas(struct ethtool_link_settings)
899 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
900 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
901 struct ethtool_link_settings *ecmd = (void *)data;
904 ifr.ifr_data = (void *)ecmd;
905 ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
908 "port %u ioctl(SIOCETHTOOL,"
909 "ETHTOOL_GLINKSETTINGS) failed: %s",
910 dev->data->port_id, strerror(rte_errno));
913 dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
915 sc = ecmd->link_mode_masks[0] |
916 ((uint64_t)ecmd->link_mode_masks[1] << 32);
917 priv->link_speed_capa = 0;
918 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
919 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
920 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
921 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
922 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
923 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
924 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
925 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
926 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
927 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
928 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
929 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
930 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
931 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
932 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
933 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
934 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
935 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
936 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
937 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
938 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
939 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
940 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
941 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
942 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
943 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
944 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
945 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
946 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
947 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
948 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
949 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
950 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
951 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
952 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
953 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
954 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
955 ETH_LINK_SPEED_FIXED);
956 if (((dev_link.link_speed && !dev_link.link_status) ||
957 (!dev_link.link_speed && dev_link.link_status))) {
966 * DPDK callback to retrieve physical link information.
969 * Pointer to Ethernet device structure.
970 * @param wait_to_complete
971 * Wait for request completion.
974 * 0 if link status was not updated, positive if it was, a negative errno
975 * value otherwise and rte_errno is set.
978 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
981 struct rte_eth_link dev_link;
982 time_t start_time = time(NULL);
985 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
987 ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
990 /* Handle wait to complete situation. */
991 if (wait_to_complete && ret == -EAGAIN) {
992 if (abs((int)difftime(time(NULL), start_time)) <
993 MLX5_LINK_STATUS_TIMEOUT) {
1000 } else if (ret < 0) {
1003 } while (wait_to_complete);
1004 ret = !!memcmp(&dev->data->dev_link, &dev_link,
1005 sizeof(struct rte_eth_link));
1006 dev->data->dev_link = dev_link;
1011 * DPDK callback to change the MTU.
1014 * Pointer to Ethernet device structure.
1019 * 0 on success, a negative errno value otherwise and rte_errno is set.
1022 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1024 struct mlx5_priv *priv = dev->data->dev_private;
1025 uint16_t kern_mtu = 0;
1028 ret = mlx5_get_mtu(dev, &kern_mtu);
1031 /* Set kernel interface MTU first. */
1032 ret = mlx5_set_mtu(dev, mtu);
1035 ret = mlx5_get_mtu(dev, &kern_mtu);
1038 if (kern_mtu == mtu) {
1040 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1041 dev->data->port_id, mtu);
1049 * DPDK callback to get flow control status.
1052 * Pointer to Ethernet device structure.
1053 * @param[out] fc_conf
1054 * Flow control output buffer.
1057 * 0 on success, a negative errno value otherwise and rte_errno is set.
1060 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1063 struct ethtool_pauseparam ethpause = {
1064 .cmd = ETHTOOL_GPAUSEPARAM
1068 ifr.ifr_data = (void *)ðpause;
1069 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1072 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1074 dev->data->port_id, strerror(rte_errno));
1077 fc_conf->autoneg = ethpause.autoneg;
1078 if (ethpause.rx_pause && ethpause.tx_pause)
1079 fc_conf->mode = RTE_FC_FULL;
1080 else if (ethpause.rx_pause)
1081 fc_conf->mode = RTE_FC_RX_PAUSE;
1082 else if (ethpause.tx_pause)
1083 fc_conf->mode = RTE_FC_TX_PAUSE;
1085 fc_conf->mode = RTE_FC_NONE;
1090 * DPDK callback to modify flow control parameters.
1093 * Pointer to Ethernet device structure.
1094 * @param[in] fc_conf
1095 * Flow control parameters.
1098 * 0 on success, a negative errno value otherwise and rte_errno is set.
1101 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1104 struct ethtool_pauseparam ethpause = {
1105 .cmd = ETHTOOL_SPAUSEPARAM
1109 ifr.ifr_data = (void *)ðpause;
1110 ethpause.autoneg = fc_conf->autoneg;
1111 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1112 (fc_conf->mode & RTE_FC_RX_PAUSE))
1113 ethpause.rx_pause = 1;
1115 ethpause.rx_pause = 0;
1117 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1118 (fc_conf->mode & RTE_FC_TX_PAUSE))
1119 ethpause.tx_pause = 1;
1121 ethpause.tx_pause = 0;
1122 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1125 "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1127 dev->data->port_id, strerror(rte_errno));
1134 * Get PCI information from struct ibv_device.
1137 * Pointer to Ethernet device structure.
1138 * @param[out] pci_addr
1139 * PCI bus address output buffer.
1142 * 0 on success, a negative errno value otherwise and rte_errno is set.
1145 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1146 struct rte_pci_addr *pci_addr)
1150 MKSTR(path, "%s/device/uevent", device->ibdev_path);
1152 file = fopen(path, "rb");
1157 while (fgets(line, sizeof(line), file) == line) {
1158 size_t len = strlen(line);
1161 /* Truncate long lines. */
1162 if (len == (sizeof(line) - 1))
1163 while (line[(len - 1)] != '\n') {
1167 line[(len - 1)] = ret;
1169 /* Extract information. */
1172 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1176 &pci_addr->function) == 4) {
1186 * Handle asynchronous removal event for entire multiport device.
1189 * Infiniband device shared context.
1192 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1196 for (i = 0; i < sh->max_port; ++i) {
1197 struct rte_eth_dev *dev;
1199 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1201 * Or not existing port either no
1202 * handler installed for this port.
1206 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1208 if (dev->data->dev_conf.intr_conf.rmv)
1209 _rte_eth_dev_callback_process
1210 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1215 * Handle shared asynchronous events the NIC (removal event
1216 * and link status change). Supports multiport IB device.
1219 * Callback argument.
1222 mlx5_dev_interrupt_handler(void *cb_arg)
1224 struct mlx5_ibv_shared *sh = cb_arg;
1225 struct ibv_async_event event;
1227 /* Read all message from the IB device and acknowledge them. */
1229 struct rte_eth_dev *dev;
1232 if (mlx5_glue->get_async_event(sh->ctx, &event))
1234 /* Retrieve and check IB port index. */
1235 tmp = (uint32_t)event.element.port_num;
1236 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1238 * The DEVICE_FATAL event is called once for
1239 * entire device without port specifying.
1240 * We should notify all existing ports.
1242 mlx5_glue->ack_async_event(&event);
1243 mlx5_dev_interrupt_device_fatal(sh);
1246 assert(tmp && (tmp <= sh->max_port));
1248 /* Unsupported devive level event. */
1249 mlx5_glue->ack_async_event(&event);
1251 "unsupported common event (type %d)",
1255 if (tmp > sh->max_port) {
1256 /* Invalid IB port index. */
1257 mlx5_glue->ack_async_event(&event);
1259 "cannot handle an event (type %d)"
1260 "due to invalid IB port index (%u)",
1261 event.event_type, tmp);
1264 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1265 /* No handler installed. */
1266 mlx5_glue->ack_async_event(&event);
1268 "cannot handle an event (type %d)"
1269 "due to no handler installed for port %u",
1270 event.event_type, tmp);
1273 /* Retrieve ethernet device descriptor. */
1274 tmp = sh->port[tmp - 1].ih_port_id;
1275 dev = &rte_eth_devices[tmp];
1277 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1278 event.event_type == IBV_EVENT_PORT_ERR) &&
1279 dev->data->dev_conf.intr_conf.lsc) {
1280 mlx5_glue->ack_async_event(&event);
1281 if (mlx5_link_update(dev, 0) == -EAGAIN) {
1285 _rte_eth_dev_callback_process
1286 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1290 "port %u cannot handle an unknown event (type %d)",
1291 dev->data->port_id, event.event_type);
1292 mlx5_glue->ack_async_event(&event);
1297 * Unregister callback handler safely. The handler may be active
1298 * while we are trying to unregister it, in this case code -EAGAIN
1299 * is returned by rte_intr_callback_unregister(). This routine checks
1300 * the return code and tries to unregister handler again.
1305 * pointer to callback routine
1307 * opaque callback parameter
1310 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1311 rte_intr_callback_fn cb_fn, void *cb_arg)
1314 * Try to reduce timeout management overhead by not calling
1315 * the timer related routines on the first iteration. If the
1316 * unregistering succeeds on first call there will be no
1317 * timer calls at all.
1325 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1328 if (ret != -EAGAIN) {
1329 DRV_LOG(INFO, "failed to unregister interrupt"
1330 " handler (error: %d)", ret);
1335 struct timespec onems;
1337 /* Wait one millisecond and try again. */
1339 onems.tv_nsec = NS_PER_S / MS_PER_S;
1340 nanosleep(&onems, 0);
1341 /* Check whether one second elapsed. */
1342 if ((rte_get_timer_cycles() - start) <= twait)
1346 * We get the amount of timer ticks for one second.
1347 * If this amount elapsed it means we spent one
1348 * second in waiting. This branch is executed once
1349 * on first iteration.
1351 twait = rte_get_timer_hz();
1355 * Timeout elapsed, show message (once a second) and retry.
1356 * We have no other acceptable option here, if we ignore
1357 * the unregistering return code the handler will not
1358 * be unregistered, fd will be closed and we may get the
1359 * crush. Hanging and messaging in the loop seems not to be
1362 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1363 start = rte_get_timer_cycles();
1368 * Handle DEVX interrupts from the NIC.
1369 * This function is probably called from the DPDK host thread.
1372 * Callback argument.
1375 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1377 #ifndef HAVE_IBV_DEVX_ASYNC
1381 struct mlx5_ibv_shared *sh = cb_arg;
1383 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1384 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1385 MLX5_ST_SZ_BYTES(traffic_counter) +
1386 sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1388 uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1390 while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1393 mlx5_flow_async_pool_query_handle
1394 (sh, (uint64_t)out.cmd_resp.wr_id,
1395 mlx5_devx_get_out_command_status(buf));
1396 #endif /* HAVE_IBV_DEVX_ASYNC */
1400 * Uninstall shared asynchronous device events handler.
1401 * This function is implemented to support event sharing
1402 * between multiple ports of single IB device.
1405 * Pointer to Ethernet device.
1408 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1410 struct mlx5_priv *priv = dev->data->dev_private;
1411 struct mlx5_ibv_shared *sh = priv->sh;
1413 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1415 pthread_mutex_lock(&sh->intr_mutex);
1416 assert(priv->ibv_port);
1417 assert(priv->ibv_port <= sh->max_port);
1418 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1419 if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1421 assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1422 (uint32_t)dev->data->port_id);
1423 assert(sh->intr_cnt);
1424 sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1425 if (!sh->intr_cnt || --sh->intr_cnt)
1427 mlx5_intr_callback_unregister(&sh->intr_handle,
1428 mlx5_dev_interrupt_handler, sh);
1429 sh->intr_handle.fd = 0;
1430 sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1431 if (sh->intr_handle_devx.fd) {
1432 rte_intr_callback_unregister(&sh->intr_handle_devx,
1433 mlx5_dev_interrupt_handler_devx,
1435 sh->intr_handle_devx.fd = 0;
1436 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1438 if (sh->devx_comp) {
1439 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1440 sh->devx_comp = NULL;
1443 pthread_mutex_unlock(&sh->intr_mutex);
1447 * Install shared asynchronous device events handler.
1448 * This function is implemented to support event sharing
1449 * between multiple ports of single IB device.
1452 * Pointer to Ethernet device.
1455 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1457 struct mlx5_priv *priv = dev->data->dev_private;
1458 struct mlx5_ibv_shared *sh = priv->sh;
1462 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1464 pthread_mutex_lock(&sh->intr_mutex);
1465 assert(priv->ibv_port);
1466 assert(priv->ibv_port <= sh->max_port);
1467 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1468 if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1469 /* The handler is already installed for this port. */
1470 assert(sh->intr_cnt);
1473 sh->port[priv->ibv_port - 1].ih_port_id = (uint32_t)dev->data->port_id;
1478 /* No shared handler installed. */
1479 assert(sh->ctx->async_fd > 0);
1480 flags = fcntl(sh->ctx->async_fd, F_GETFL);
1481 ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1483 DRV_LOG(INFO, "failed to change file descriptor"
1484 " async event queue");
1487 sh->intr_handle.fd = sh->ctx->async_fd;
1488 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1489 rte_intr_callback_register(&sh->intr_handle,
1490 mlx5_dev_interrupt_handler, sh);
1491 if (priv->config.devx) {
1492 #ifndef HAVE_IBV_DEVX_ASYNC
1493 goto error_unregister;
1495 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1496 if (sh->devx_comp) {
1497 flags = fcntl(sh->devx_comp->fd, F_GETFL);
1498 ret = fcntl(sh->devx_comp->fd, F_SETFL,
1499 flags | O_NONBLOCK);
1501 DRV_LOG(INFO, "failed to change file descriptor"
1502 " devx async event queue");
1503 goto error_unregister;
1505 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1506 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1507 rte_intr_callback_register
1508 (&sh->intr_handle_devx,
1509 mlx5_dev_interrupt_handler_devx, sh);
1511 DRV_LOG(INFO, "failed to create devx async command "
1513 goto error_unregister;
1515 #endif /* HAVE_IBV_DEVX_ASYNC */
1520 rte_intr_callback_unregister(&sh->intr_handle,
1521 mlx5_dev_interrupt_handler, sh);
1523 /* Indicate there will be no interrupts. */
1524 dev->data->dev_conf.intr_conf.lsc = 0;
1525 dev->data->dev_conf.intr_conf.rmv = 0;
1526 sh->intr_handle.fd = 0;
1527 sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1528 sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1530 pthread_mutex_unlock(&sh->intr_mutex);
1534 * Uninstall interrupt handler.
1537 * Pointer to Ethernet device.
1540 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1542 mlx5_dev_shared_handler_uninstall(dev);
1546 * Install interrupt handler.
1549 * Pointer to Ethernet device.
1552 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1554 mlx5_dev_shared_handler_install(dev);
1558 * DPDK callback to bring the link DOWN.
1561 * Pointer to Ethernet device structure.
1564 * 0 on success, a negative errno value otherwise and rte_errno is set.
1567 mlx5_set_link_down(struct rte_eth_dev *dev)
1569 return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1573 * DPDK callback to bring the link UP.
1576 * Pointer to Ethernet device structure.
1579 * 0 on success, a negative errno value otherwise and rte_errno is set.
1582 mlx5_set_link_up(struct rte_eth_dev *dev)
1584 return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1588 * Configure the RX function to use.
1591 * Pointer to private data structure.
1594 * Pointer to selected Rx burst function.
1597 mlx5_select_rx_function(struct rte_eth_dev *dev)
1599 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1601 assert(dev != NULL);
1602 if (mlx5_check_vec_rx_support(dev) > 0) {
1603 rx_pkt_burst = mlx5_rx_burst_vec;
1604 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1605 dev->data->port_id);
1606 } else if (mlx5_mprq_enabled(dev)) {
1607 rx_pkt_burst = mlx5_rx_burst_mprq;
1609 return rx_pkt_burst;
1613 * Check if mlx5 device was removed.
1616 * Pointer to Ethernet device structure.
1619 * 1 when device is removed, otherwise 0.
1622 mlx5_is_removed(struct rte_eth_dev *dev)
1624 struct ibv_device_attr device_attr;
1625 struct mlx5_priv *priv = dev->data->dev_private;
1627 if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1633 * Get port ID list of mlx5 instances sharing a common device.
1636 * Device to look for.
1637 * @param[out] port_list
1638 * Result buffer for collected port IDs.
1639 * @param port_list_n
1640 * Maximum number of entries in result buffer. If 0, @p port_list can be
1644 * Number of matching instances regardless of the @p port_list_n
1645 * parameter, 0 if none were found.
1648 mlx5_dev_to_port_id(const struct rte_device *dev, uint16_t *port_list,
1649 unsigned int port_list_n)
1654 RTE_ETH_FOREACH_DEV_OF(id, dev) {
1655 if (n < port_list_n)
1663 * Get the E-Switch domain id this port belongs to.
1667 * @param[out] es_domain_id
1668 * E-Switch domain id.
1669 * @param[out] es_port_id
1670 * The port id of the port in the E-Switch.
1673 * 0 on success, a negative errno value otherwise and rte_errno is set.
1676 mlx5_port_to_eswitch_info(uint16_t port,
1677 uint16_t *es_domain_id, uint16_t *es_port_id)
1679 struct rte_eth_dev *dev;
1680 struct mlx5_priv *priv;
1682 if (port >= RTE_MAX_ETHPORTS) {
1686 if (!rte_eth_dev_is_valid_port(port)) {
1690 dev = &rte_eth_devices[port];
1691 priv = dev->data->dev_private;
1692 if (!(priv->representor || priv->master)) {
1697 *es_domain_id = priv->domain_id;
1699 *es_port_id = priv->vport_id;
1704 * Get switch information associated with network interface.
1707 * Network interface index.
1709 * Switch information object, populated in case of success.
1712 * 0 on success, a negative errno value otherwise and rte_errno is set.
1715 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1717 char ifname[IF_NAMESIZE];
1718 char port_name[IF_NAMESIZE];
1720 struct mlx5_switch_info data = {
1723 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1728 bool port_switch_id_set = false;
1729 bool device_dir = false;
1733 if (!if_indextoname(ifindex, ifname)) {
1738 MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1740 MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1742 MKSTR(pci_device, "/sys/class/net/%s/device",
1745 file = fopen(phys_port_name, "rb");
1747 ret = fscanf(file, "%s", port_name);
1750 mlx5_translate_port_name(port_name, &data);
1752 file = fopen(phys_switch_id, "rb");
1757 port_switch_id_set =
1758 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1761 dir = opendir(pci_device);
1766 if (port_switch_id_set) {
1767 /* We have some E-Switch configuration. */
1768 mlx5_sysfs_check_switch_info(device_dir, &data);
1771 assert(!(data.master && data.representor));
1772 if (data.master && data.representor) {
1773 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1774 " and as representor", ifindex);
1782 * Analyze gathered port parameters via Netlink to recognize master
1783 * and representor devices for E-Switch configuration.
1785 * @param[in] num_vf_set
1786 * flag of presence of number of VFs port attribute.
1787 * @param[inout] switch_info
1788 * Port information, including port name as a number and port name
1789 * type if recognized
1792 * master and representor flags are set in switch_info according to
1793 * recognized parameters (if any).
1796 mlx5_nl_check_switch_info(bool num_vf_set,
1797 struct mlx5_switch_info *switch_info)
1799 switch (switch_info->name_type) {
1800 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1802 * Name is not recognized, assume the master,
1803 * check the number of VFs key presence.
1805 switch_info->master = num_vf_set;
1807 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1809 * Name is not set, this assumes the legacy naming
1810 * schema for master, just check if there is a
1811 * number of VFs key.
1813 switch_info->master = num_vf_set;
1815 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1816 /* New uplink naming schema recognized. */
1817 switch_info->master = 1;
1819 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1820 /* Legacy representors naming schema. */
1821 switch_info->representor = !num_vf_set;
1823 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1824 /* New representors naming schema. */
1825 switch_info->representor = 1;
1831 * Analyze gathered port parameters via sysfs to recognize master
1832 * and representor devices for E-Switch configuration.
1834 * @param[in] device_dir
1835 * flag of presence of "device" directory under port device key.
1836 * @param[inout] switch_info
1837 * Port information, including port name as a number and port name
1838 * type if recognized
1841 * master and representor flags are set in switch_info according to
1842 * recognized parameters (if any).
1845 mlx5_sysfs_check_switch_info(bool device_dir,
1846 struct mlx5_switch_info *switch_info)
1848 switch (switch_info->name_type) {
1849 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1851 * Name is not recognized, assume the master,
1852 * check the device directory presence.
1854 switch_info->master = device_dir;
1856 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1858 * Name is not set, this assumes the legacy naming
1859 * schema for master, just check if there is
1860 * a device directory.
1862 switch_info->master = device_dir;
1864 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1865 /* New uplink naming schema recognized. */
1866 switch_info->master = 1;
1868 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1869 /* Legacy representors naming schema. */
1870 switch_info->representor = !device_dir;
1872 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1873 /* New representors naming schema. */
1874 switch_info->representor = 1;
1880 * Extract port name, as a number, from sysfs or netlink information.
1882 * @param[in] port_name_in
1883 * String representing the port name.
1884 * @param[out] port_info_out
1885 * Port information, including port name as a number and port name
1886 * type if recognized
1889 * port_name field set according to recognized name format.
1892 mlx5_translate_port_name(const char *port_name_in,
1893 struct mlx5_switch_info *port_info_out)
1895 char pf_c1, pf_c2, vf_c1, vf_c2;
1900 * Check for port-name as a string of the form pf0vf0
1901 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
1903 sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
1904 &pf_c1, &pf_c2, &port_info_out->pf_num,
1905 &vf_c1, &vf_c2, &port_info_out->port_name);
1906 if (sc_items == 6 &&
1907 pf_c1 == 'p' && pf_c2 == 'f' &&
1908 vf_c1 == 'v' && vf_c2 == 'f') {
1909 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
1913 * Check for port-name as a string of the form p0
1914 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
1916 sc_items = sscanf(port_name_in, "%c%d",
1917 &pf_c1, &port_info_out->port_name);
1918 if (sc_items == 2 && pf_c1 == 'p') {
1919 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1922 /* Check for port-name as a number (support kernel ver < 5.0 */
1924 port_info_out->port_name = strtol(port_name_in, &end, 0);
1926 (size_t)(end - port_name_in) == strlen(port_name_in)) {
1927 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
1930 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
1935 * DPDK callback to retrieve plug-in module EEPROM information (type and size).
1938 * Pointer to Ethernet device structure.
1939 * @param[out] modinfo
1940 * Storage for plug-in module EEPROM information.
1943 * 0 on success, a negative errno value otherwise and rte_errno is set.
1946 mlx5_get_module_info(struct rte_eth_dev *dev,
1947 struct rte_eth_dev_module_info *modinfo)
1949 struct ethtool_modinfo info = {
1950 .cmd = ETHTOOL_GMODULEINFO,
1952 struct ifreq ifr = (struct ifreq) {
1953 .ifr_data = (void *)&info,
1957 if (!dev || !modinfo) {
1958 DRV_LOG(WARNING, "missing argument, cannot get module info");
1962 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1964 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
1965 dev->data->port_id, strerror(rte_errno));
1968 modinfo->type = info.type;
1969 modinfo->eeprom_len = info.eeprom_len;
1974 * DPDK callback to retrieve plug-in module EEPROM data.
1977 * Pointer to Ethernet device structure.
1979 * Storage for plug-in module EEPROM data.
1982 * 0 on success, a negative errno value otherwise and rte_errno is set.
1984 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
1985 struct rte_dev_eeprom_info *info)
1987 struct ethtool_eeprom *eeprom;
1991 if (!dev || !info) {
1992 DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
1996 eeprom = rte_calloc(__func__, 1,
1997 (sizeof(struct ethtool_eeprom) + info->length), 0);
1999 DRV_LOG(WARNING, "port %u cannot allocate memory for "
2000 "eeprom data", dev->data->port_id);
2004 eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2005 eeprom->offset = info->offset;
2006 eeprom->len = info->length;
2007 ifr = (struct ifreq) {
2008 .ifr_data = (void *)eeprom,
2010 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2012 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2013 dev->data->port_id, strerror(rte_errno));
2015 rte_memcpy(info->data, eeprom->data, info->length);