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;
386 const uint8_t use_app_rss_key =
387 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
390 if (use_app_rss_key &&
391 (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
392 MLX5_RSS_HASH_KEY_LEN)) {
393 DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
394 dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
398 priv->rss_conf.rss_key =
399 rte_realloc(priv->rss_conf.rss_key,
400 MLX5_RSS_HASH_KEY_LEN, 0);
401 if (!priv->rss_conf.rss_key) {
402 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
403 dev->data->port_id, rxqs_n);
408 if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
409 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
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 priv->skip_default_rss_reta = 0;
437 ret = mlx5_proc_priv_init(dev);
444 * Configure default RSS reta.
447 * Pointer to Ethernet device structure.
450 * 0 on success, a negative errno value otherwise and rte_errno is set.
453 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
455 struct mlx5_priv *priv = dev->data->dev_private;
456 unsigned int rxqs_n = dev->data->nb_rx_queues;
459 unsigned int reta_idx_n;
461 unsigned int *rss_queue_arr = NULL;
462 unsigned int rss_queue_n = 0;
464 if (priv->skip_default_rss_reta)
466 rss_queue_arr = rte_malloc("", rxqs_n * sizeof(unsigned int), 0);
467 if (!rss_queue_arr) {
468 DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
469 dev->data->port_id, rxqs_n);
473 for (i = 0, j = 0; i < rxqs_n; i++) {
474 struct mlx5_rxq_data *rxq_data;
475 struct mlx5_rxq_ctrl *rxq_ctrl;
477 rxq_data = (*priv->rxqs)[i];
478 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
479 if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
480 rss_queue_arr[j++] = i;
483 if (rss_queue_n > priv->config.ind_table_max_size) {
484 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
485 dev->data->port_id, rss_queue_n);
487 rte_free(rss_queue_arr);
490 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
491 dev->data->port_id, priv->rxqs_n, rxqs_n);
492 priv->rxqs_n = rxqs_n;
494 * If the requested number of RX queues is not a power of two,
495 * use the maximum indirection table size for better balancing.
496 * The result is always rounded to the next power of two.
498 reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
499 priv->config.ind_table_max_size :
501 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
503 rte_free(rss_queue_arr);
507 * When the number of RX queues is not a power of two,
508 * the remaining table entries are padded with reused WQs
509 * and hashes are not spread uniformly.
511 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
512 (*priv->reta_idx)[i] = rss_queue_arr[j];
513 if (++j == rss_queue_n)
516 rte_free(rss_queue_arr);
521 * Sets default tuning parameters.
524 * Pointer to Ethernet device.
526 * Info structure output buffer.
529 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
531 struct mlx5_priv *priv = dev->data->dev_private;
533 /* Minimum CPU utilization. */
534 info->default_rxportconf.ring_size = 256;
535 info->default_txportconf.ring_size = 256;
536 info->default_rxportconf.burst_size = MLX5_RX_DEFAULT_BURST;
537 info->default_txportconf.burst_size = MLX5_TX_DEFAULT_BURST;
538 if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
539 info->default_rxportconf.nb_queues = 16;
540 info->default_txportconf.nb_queues = 16;
541 if (dev->data->nb_rx_queues > 2 ||
542 dev->data->nb_tx_queues > 2) {
543 /* Max Throughput. */
544 info->default_rxportconf.ring_size = 2048;
545 info->default_txportconf.ring_size = 2048;
548 info->default_rxportconf.nb_queues = 8;
549 info->default_txportconf.nb_queues = 8;
550 if (dev->data->nb_rx_queues > 2 ||
551 dev->data->nb_tx_queues > 2) {
552 /* Max Throughput. */
553 info->default_rxportconf.ring_size = 4096;
554 info->default_txportconf.ring_size = 4096;
560 * Sets tx mbuf limiting parameters.
563 * Pointer to Ethernet device.
565 * Info structure output buffer.
568 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
570 struct mlx5_priv *priv = dev->data->dev_private;
571 struct mlx5_dev_config *config = &priv->config;
575 inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
576 MLX5_SEND_DEF_INLINE_LEN :
577 (unsigned int)config->txq_inline_max;
578 assert(config->txq_inline_min >= 0);
579 inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
580 inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
581 MLX5_ESEG_MIN_INLINE_SIZE -
584 MLX5_WQE_DSEG_SIZE * 2);
585 nb_max = (MLX5_WQE_SIZE_MAX +
586 MLX5_ESEG_MIN_INLINE_SIZE -
590 inlen) / MLX5_WSEG_SIZE;
591 info->tx_desc_lim.nb_seg_max = nb_max;
592 info->tx_desc_lim.nb_mtu_seg_max = nb_max;
596 * DPDK callback to get information about the device.
599 * Pointer to Ethernet device structure.
601 * Info structure output buffer.
604 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
606 struct mlx5_priv *priv = dev->data->dev_private;
607 struct mlx5_dev_config *config = &priv->config;
610 /* FIXME: we should ask the device for these values. */
611 info->min_rx_bufsize = 32;
612 info->max_rx_pktlen = 65536;
613 info->max_lro_pkt_size = MLX5_MAX_LRO_SIZE;
615 * Since we need one CQ per QP, the limit is the minimum number
616 * between the two values.
618 max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
619 priv->sh->device_attr.orig_attr.max_qp);
620 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
623 info->max_rx_queues = max;
624 info->max_tx_queues = max;
625 info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
626 info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
627 info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
628 info->rx_queue_offload_capa);
629 info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
630 info->if_index = mlx5_ifindex(dev);
631 info->reta_size = priv->reta_idx_n ?
632 priv->reta_idx_n : config->ind_table_max_size;
633 info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
634 info->speed_capa = priv->link_speed_capa;
635 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
636 mlx5_set_default_params(dev, info);
637 mlx5_set_txlimit_params(dev, info);
638 info->switch_info.name = dev->data->name;
639 info->switch_info.domain_id = priv->domain_id;
640 info->switch_info.port_id = priv->representor_id;
641 if (priv->representor) {
644 if (priv->pf_bond >= 0) {
646 * Switch port ID is opaque value with driver defined
647 * format. Push the PF index in bonding configurations
648 * in upper four bits of port ID. If we get too many
649 * representors (more than 4K) or PFs (more than 15)
650 * this approach must be reconsidered.
652 if ((info->switch_info.port_id >>
653 MLX5_PORT_ID_BONDING_PF_SHIFT) ||
654 priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
655 DRV_LOG(ERR, "can't update switch port ID"
656 " for bonding device");
660 info->switch_info.port_id |=
661 priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
663 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
664 struct mlx5_priv *opriv =
665 rte_eth_devices[port_id].data->dev_private;
668 opriv->representor ||
669 opriv->sh != priv->sh ||
670 opriv->domain_id != priv->domain_id)
673 * Override switch name with that of the master
676 info->switch_info.name = opriv->dev_data->name;
684 * Get device current raw clock counter
687 * Pointer to Ethernet device structure.
689 * Current raw clock counter of the device.
692 * 0 if the clock has correctly been read
693 * The value of errno in case of error
696 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
698 struct mlx5_priv *priv = dev->data->dev_private;
699 struct ibv_context *ctx = priv->sh->ctx;
700 struct ibv_values_ex values;
703 values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
704 err = mlx5_glue->query_rt_values_ex(ctx, &values);
706 DRV_LOG(WARNING, "Could not query the clock !");
709 *clock = values.raw_clock.tv_nsec;
714 * Get firmware version of a device.
717 * Ethernet device port.
719 * String output allocated by caller.
721 * Size of the output string, including terminating null byte.
724 * 0 on success, or the size of the non truncated string if too big.
726 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
728 struct mlx5_priv *priv = dev->data->dev_private;
729 struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
730 size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
735 strlcpy(fw_ver, attr->fw_ver, fw_size);
740 * Get supported packet types.
743 * Pointer to Ethernet device structure.
746 * A pointer to the supported Packet types array.
749 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
751 static const uint32_t ptypes[] = {
752 /* refers to rxq_cq_to_pkt_type() */
754 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
755 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
756 RTE_PTYPE_L4_NONFRAG,
760 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
761 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
762 RTE_PTYPE_INNER_L4_NONFRAG,
763 RTE_PTYPE_INNER_L4_FRAG,
764 RTE_PTYPE_INNER_L4_TCP,
765 RTE_PTYPE_INNER_L4_UDP,
769 if (dev->rx_pkt_burst == mlx5_rx_burst ||
770 dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
771 dev->rx_pkt_burst == mlx5_rx_burst_vec)
777 * Retrieve the master device for representor in the same switch domain.
780 * Pointer to representor Ethernet device structure.
783 * Master device structure on success, NULL otherwise.
786 static struct rte_eth_dev *
787 mlx5_find_master_dev(struct rte_eth_dev *dev)
789 struct mlx5_priv *priv;
793 priv = dev->data->dev_private;
794 domain_id = priv->domain_id;
795 assert(priv->representor);
796 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
797 struct mlx5_priv *opriv =
798 rte_eth_devices[port_id].data->dev_private;
801 opriv->domain_id == domain_id &&
802 opriv->sh == priv->sh)
803 return &rte_eth_devices[port_id];
809 * DPDK callback to retrieve physical link information.
812 * Pointer to Ethernet device structure.
814 * Storage for current link status.
817 * 0 on success, a negative errno value otherwise and rte_errno is set.
820 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
821 struct rte_eth_link *link)
823 struct mlx5_priv *priv = dev->data->dev_private;
824 struct ethtool_cmd edata = {
825 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
828 struct rte_eth_link dev_link;
832 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
834 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
835 dev->data->port_id, strerror(rte_errno));
838 dev_link = (struct rte_eth_link) {
839 .link_status = ((ifr.ifr_flags & IFF_UP) &&
840 (ifr.ifr_flags & IFF_RUNNING)),
842 ifr = (struct ifreq) {
843 .ifr_data = (void *)&edata,
845 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
847 if (ret == -ENOTSUP && priv->representor) {
848 struct rte_eth_dev *master;
851 * For representors we can try to inherit link
852 * settings from the master device. Actually
853 * link settings do not make a lot of sense
854 * for representors due to missing physical
855 * link. The old kernel drivers supported
856 * emulated settings query for representors,
857 * the new ones do not, so we have to add
858 * this code for compatibility issues.
860 master = mlx5_find_master_dev(dev);
862 ifr = (struct ifreq) {
863 .ifr_data = (void *)&edata,
865 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
870 "port %u ioctl(SIOCETHTOOL,"
871 " ETHTOOL_GSET) failed: %s",
872 dev->data->port_id, strerror(rte_errno));
876 link_speed = ethtool_cmd_speed(&edata);
877 if (link_speed == -1)
878 dev_link.link_speed = ETH_SPEED_NUM_NONE;
880 dev_link.link_speed = link_speed;
881 priv->link_speed_capa = 0;
882 if (edata.supported & SUPPORTED_Autoneg)
883 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
884 if (edata.supported & (SUPPORTED_1000baseT_Full |
885 SUPPORTED_1000baseKX_Full))
886 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
887 if (edata.supported & SUPPORTED_10000baseKR_Full)
888 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
889 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
890 SUPPORTED_40000baseCR4_Full |
891 SUPPORTED_40000baseSR4_Full |
892 SUPPORTED_40000baseLR4_Full))
893 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
894 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
895 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
896 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
897 ETH_LINK_SPEED_FIXED);
898 if (((dev_link.link_speed && !dev_link.link_status) ||
899 (!dev_link.link_speed && dev_link.link_status))) {
908 * Retrieve physical link information (unlocked version using new ioctl).
911 * Pointer to Ethernet device structure.
913 * Storage for current link status.
916 * 0 on success, a negative errno value otherwise and rte_errno is set.
919 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
920 struct rte_eth_link *link)
923 struct mlx5_priv *priv = dev->data->dev_private;
924 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
926 struct rte_eth_link dev_link;
927 struct rte_eth_dev *master = NULL;
931 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
933 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
934 dev->data->port_id, strerror(rte_errno));
937 dev_link = (struct rte_eth_link) {
938 .link_status = ((ifr.ifr_flags & IFF_UP) &&
939 (ifr.ifr_flags & IFF_RUNNING)),
941 ifr = (struct ifreq) {
942 .ifr_data = (void *)&gcmd,
944 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
946 if (ret == -ENOTSUP && priv->representor) {
948 * For representors we can try to inherit link
949 * settings from the master device. Actually
950 * link settings do not make a lot of sense
951 * for representors due to missing physical
952 * link. The old kernel drivers supported
953 * emulated settings query for representors,
954 * the new ones do not, so we have to add
955 * this code for compatibility issues.
957 master = mlx5_find_master_dev(dev);
959 ifr = (struct ifreq) {
960 .ifr_data = (void *)&gcmd,
962 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
967 "port %u ioctl(SIOCETHTOOL,"
968 " ETHTOOL_GLINKSETTINGS) failed: %s",
969 dev->data->port_id, strerror(rte_errno));
974 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
976 alignas(struct ethtool_link_settings)
977 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
978 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
979 struct ethtool_link_settings *ecmd = (void *)data;
982 ifr.ifr_data = (void *)ecmd;
983 ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
986 "port %u ioctl(SIOCETHTOOL,"
987 "ETHTOOL_GLINKSETTINGS) failed: %s",
988 dev->data->port_id, strerror(rte_errno));
991 dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
993 sc = ecmd->link_mode_masks[0] |
994 ((uint64_t)ecmd->link_mode_masks[1] << 32);
995 priv->link_speed_capa = 0;
996 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
997 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
998 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
999 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
1000 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
1001 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
1002 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
1003 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
1004 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
1005 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
1006 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
1007 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
1008 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
1009 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
1010 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
1011 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
1012 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
1013 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
1014 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
1015 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
1016 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
1017 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
1018 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
1019 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
1020 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
1021 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
1022 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
1023 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
1024 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
1025 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
1026 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
1027 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
1028 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
1029 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
1030 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
1031 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
1032 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
1033 ETH_LINK_SPEED_FIXED);
1034 if (((dev_link.link_speed && !dev_link.link_status) ||
1035 (!dev_link.link_speed && dev_link.link_status))) {
1044 * DPDK callback to retrieve physical link information.
1047 * Pointer to Ethernet device structure.
1048 * @param wait_to_complete
1049 * Wait for request completion.
1052 * 0 if link status was not updated, positive if it was, a negative errno
1053 * value otherwise and rte_errno is set.
1056 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1059 struct rte_eth_link dev_link;
1060 time_t start_time = time(NULL);
1061 int retry = MLX5_GET_LINK_STATUS_RETRY_COUNT;
1064 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1065 if (ret == -ENOTSUP)
1066 ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1069 /* Handle wait to complete situation. */
1070 if ((wait_to_complete || retry) && ret == -EAGAIN) {
1071 if (abs((int)difftime(time(NULL), start_time)) <
1072 MLX5_LINK_STATUS_TIMEOUT) {
1079 } else if (ret < 0) {
1082 } while (wait_to_complete || retry-- > 0);
1083 ret = !!memcmp(&dev->data->dev_link, &dev_link,
1084 sizeof(struct rte_eth_link));
1085 dev->data->dev_link = dev_link;
1090 * DPDK callback to change the MTU.
1093 * Pointer to Ethernet device structure.
1098 * 0 on success, a negative errno value otherwise and rte_errno is set.
1101 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1103 struct mlx5_priv *priv = dev->data->dev_private;
1104 uint16_t kern_mtu = 0;
1107 ret = mlx5_get_mtu(dev, &kern_mtu);
1110 /* Set kernel interface MTU first. */
1111 ret = mlx5_set_mtu(dev, mtu);
1114 ret = mlx5_get_mtu(dev, &kern_mtu);
1117 if (kern_mtu == mtu) {
1119 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1120 dev->data->port_id, mtu);
1128 * DPDK callback to get flow control status.
1131 * Pointer to Ethernet device structure.
1132 * @param[out] fc_conf
1133 * Flow control output buffer.
1136 * 0 on success, a negative errno value otherwise and rte_errno is set.
1139 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1142 struct ethtool_pauseparam ethpause = {
1143 .cmd = ETHTOOL_GPAUSEPARAM
1147 ifr.ifr_data = (void *)ðpause;
1148 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1151 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1153 dev->data->port_id, strerror(rte_errno));
1156 fc_conf->autoneg = ethpause.autoneg;
1157 if (ethpause.rx_pause && ethpause.tx_pause)
1158 fc_conf->mode = RTE_FC_FULL;
1159 else if (ethpause.rx_pause)
1160 fc_conf->mode = RTE_FC_RX_PAUSE;
1161 else if (ethpause.tx_pause)
1162 fc_conf->mode = RTE_FC_TX_PAUSE;
1164 fc_conf->mode = RTE_FC_NONE;
1169 * DPDK callback to modify flow control parameters.
1172 * Pointer to Ethernet device structure.
1173 * @param[in] fc_conf
1174 * Flow control parameters.
1177 * 0 on success, a negative errno value otherwise and rte_errno is set.
1180 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1183 struct ethtool_pauseparam ethpause = {
1184 .cmd = ETHTOOL_SPAUSEPARAM
1188 ifr.ifr_data = (void *)ðpause;
1189 ethpause.autoneg = fc_conf->autoneg;
1190 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1191 (fc_conf->mode & RTE_FC_RX_PAUSE))
1192 ethpause.rx_pause = 1;
1194 ethpause.rx_pause = 0;
1196 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1197 (fc_conf->mode & RTE_FC_TX_PAUSE))
1198 ethpause.tx_pause = 1;
1200 ethpause.tx_pause = 0;
1201 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1204 "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1206 dev->data->port_id, strerror(rte_errno));
1213 * Get PCI information by sysfs device path.
1216 * Pointer to device sysfs folder name.
1217 * @param[out] pci_addr
1218 * PCI bus address output buffer.
1221 * 0 on success, a negative errno value otherwise and rte_errno is set.
1224 mlx5_dev_to_pci_addr(const char *dev_path,
1225 struct rte_pci_addr *pci_addr)
1229 MKSTR(path, "%s/device/uevent", dev_path);
1231 file = fopen(path, "rb");
1236 while (fgets(line, sizeof(line), file) == line) {
1237 size_t len = strlen(line);
1240 /* Truncate long lines. */
1241 if (len == (sizeof(line) - 1))
1242 while (line[(len - 1)] != '\n') {
1246 line[(len - 1)] = ret;
1248 /* Extract information. */
1251 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1255 &pci_addr->function) == 4) {
1265 * Handle asynchronous removal event for entire multiport device.
1268 * Infiniband device shared context.
1271 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1275 for (i = 0; i < sh->max_port; ++i) {
1276 struct rte_eth_dev *dev;
1278 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1280 * Or not existing port either no
1281 * handler installed for this port.
1285 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1287 if (dev->data->dev_conf.intr_conf.rmv)
1288 _rte_eth_dev_callback_process
1289 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1294 * Handle shared asynchronous events the NIC (removal event
1295 * and link status change). Supports multiport IB device.
1298 * Callback argument.
1301 mlx5_dev_interrupt_handler(void *cb_arg)
1303 struct mlx5_ibv_shared *sh = cb_arg;
1304 struct ibv_async_event event;
1306 /* Read all message from the IB device and acknowledge them. */
1308 struct rte_eth_dev *dev;
1311 if (mlx5_glue->get_async_event(sh->ctx, &event))
1313 /* Retrieve and check IB port index. */
1314 tmp = (uint32_t)event.element.port_num;
1315 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1317 * The DEVICE_FATAL event is called once for
1318 * entire device without port specifying.
1319 * We should notify all existing ports.
1321 mlx5_glue->ack_async_event(&event);
1322 mlx5_dev_interrupt_device_fatal(sh);
1325 assert(tmp && (tmp <= sh->max_port));
1327 /* Unsupported devive level event. */
1328 mlx5_glue->ack_async_event(&event);
1330 "unsupported common event (type %d)",
1334 if (tmp > sh->max_port) {
1335 /* Invalid IB port index. */
1336 mlx5_glue->ack_async_event(&event);
1338 "cannot handle an event (type %d)"
1339 "due to invalid IB port index (%u)",
1340 event.event_type, tmp);
1343 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1344 /* No handler installed. */
1345 mlx5_glue->ack_async_event(&event);
1347 "cannot handle an event (type %d)"
1348 "due to no handler installed for port %u",
1349 event.event_type, tmp);
1352 /* Retrieve ethernet device descriptor. */
1353 tmp = sh->port[tmp - 1].ih_port_id;
1354 dev = &rte_eth_devices[tmp];
1356 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1357 event.event_type == IBV_EVENT_PORT_ERR) &&
1358 dev->data->dev_conf.intr_conf.lsc) {
1359 mlx5_glue->ack_async_event(&event);
1360 if (mlx5_link_update(dev, 0) == -EAGAIN) {
1364 _rte_eth_dev_callback_process
1365 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1369 "port %u cannot handle an unknown event (type %d)",
1370 dev->data->port_id, event.event_type);
1371 mlx5_glue->ack_async_event(&event);
1376 * Unregister callback handler safely. The handler may be active
1377 * while we are trying to unregister it, in this case code -EAGAIN
1378 * is returned by rte_intr_callback_unregister(). This routine checks
1379 * the return code and tries to unregister handler again.
1384 * pointer to callback routine
1386 * opaque callback parameter
1389 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1390 rte_intr_callback_fn cb_fn, void *cb_arg)
1393 * Try to reduce timeout management overhead by not calling
1394 * the timer related routines on the first iteration. If the
1395 * unregistering succeeds on first call there will be no
1396 * timer calls at all.
1404 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1407 if (ret != -EAGAIN) {
1408 DRV_LOG(INFO, "failed to unregister interrupt"
1409 " handler (error: %d)", ret);
1414 struct timespec onems;
1416 /* Wait one millisecond and try again. */
1418 onems.tv_nsec = NS_PER_S / MS_PER_S;
1419 nanosleep(&onems, 0);
1420 /* Check whether one second elapsed. */
1421 if ((rte_get_timer_cycles() - start) <= twait)
1425 * We get the amount of timer ticks for one second.
1426 * If this amount elapsed it means we spent one
1427 * second in waiting. This branch is executed once
1428 * on first iteration.
1430 twait = rte_get_timer_hz();
1434 * Timeout elapsed, show message (once a second) and retry.
1435 * We have no other acceptable option here, if we ignore
1436 * the unregistering return code the handler will not
1437 * be unregistered, fd will be closed and we may get the
1438 * crush. Hanging and messaging in the loop seems not to be
1441 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1442 start = rte_get_timer_cycles();
1447 * Handle DEVX interrupts from the NIC.
1448 * This function is probably called from the DPDK host thread.
1451 * Callback argument.
1454 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1456 #ifndef HAVE_IBV_DEVX_ASYNC
1460 struct mlx5_ibv_shared *sh = cb_arg;
1462 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1463 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1464 MLX5_ST_SZ_BYTES(traffic_counter) +
1465 sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1467 uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1469 while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1472 mlx5_flow_async_pool_query_handle
1473 (sh, (uint64_t)out.cmd_resp.wr_id,
1474 mlx5_devx_get_out_command_status(buf));
1475 #endif /* HAVE_IBV_DEVX_ASYNC */
1479 * Uninstall shared asynchronous device events handler.
1480 * This function is implemented to support event sharing
1481 * between multiple ports of single IB device.
1484 * Pointer to Ethernet device.
1487 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1489 struct mlx5_priv *priv = dev->data->dev_private;
1490 struct mlx5_ibv_shared *sh = priv->sh;
1492 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1494 pthread_mutex_lock(&sh->intr_mutex);
1495 assert(priv->ibv_port);
1496 assert(priv->ibv_port <= sh->max_port);
1497 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1498 if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1500 assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1501 (uint32_t)dev->data->port_id);
1502 assert(sh->intr_cnt);
1503 sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1504 if (!sh->intr_cnt || --sh->intr_cnt)
1506 mlx5_intr_callback_unregister(&sh->intr_handle,
1507 mlx5_dev_interrupt_handler, sh);
1508 sh->intr_handle.fd = 0;
1509 sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1511 pthread_mutex_unlock(&sh->intr_mutex);
1515 * Uninstall devx shared asynchronous device events handler.
1516 * This function is implemeted to support event sharing
1517 * between multiple ports of single IB device.
1520 * Pointer to Ethernet device.
1523 mlx5_dev_shared_handler_devx_uninstall(struct rte_eth_dev *dev)
1525 struct mlx5_priv *priv = dev->data->dev_private;
1526 struct mlx5_ibv_shared *sh = priv->sh;
1528 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1530 pthread_mutex_lock(&sh->intr_mutex);
1531 assert(priv->ibv_port);
1532 assert(priv->ibv_port <= sh->max_port);
1533 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1534 if (sh->port[priv->ibv_port - 1].devx_ih_port_id >= RTE_MAX_ETHPORTS)
1536 assert(sh->port[priv->ibv_port - 1].devx_ih_port_id ==
1537 (uint32_t)dev->data->port_id);
1538 sh->port[priv->ibv_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1539 if (!sh->devx_intr_cnt || --sh->devx_intr_cnt)
1541 if (sh->intr_handle_devx.fd) {
1542 rte_intr_callback_unregister(&sh->intr_handle_devx,
1543 mlx5_dev_interrupt_handler_devx,
1545 sh->intr_handle_devx.fd = 0;
1546 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1548 if (sh->devx_comp) {
1549 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1550 sh->devx_comp = NULL;
1553 pthread_mutex_unlock(&sh->intr_mutex);
1557 * Install shared asynchronous device events handler.
1558 * This function is implemented to support event sharing
1559 * between multiple ports of single IB device.
1562 * Pointer to Ethernet device.
1565 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1567 struct mlx5_priv *priv = dev->data->dev_private;
1568 struct mlx5_ibv_shared *sh = priv->sh;
1572 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1574 pthread_mutex_lock(&sh->intr_mutex);
1575 assert(priv->ibv_port);
1576 assert(priv->ibv_port <= sh->max_port);
1577 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1578 if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1579 /* The handler is already installed for this port. */
1580 assert(sh->intr_cnt);
1584 sh->port[priv->ibv_port - 1].ih_port_id =
1585 (uint32_t)dev->data->port_id;
1589 /* No shared handler installed. */
1590 assert(sh->ctx->async_fd > 0);
1591 flags = fcntl(sh->ctx->async_fd, F_GETFL);
1592 ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1594 DRV_LOG(INFO, "failed to change file descriptor async event"
1596 /* Indicate there will be no interrupts. */
1597 dev->data->dev_conf.intr_conf.lsc = 0;
1598 dev->data->dev_conf.intr_conf.rmv = 0;
1600 sh->intr_handle.fd = sh->ctx->async_fd;
1601 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1602 rte_intr_callback_register(&sh->intr_handle,
1603 mlx5_dev_interrupt_handler, sh);
1605 sh->port[priv->ibv_port - 1].ih_port_id =
1606 (uint32_t)dev->data->port_id;
1609 pthread_mutex_unlock(&sh->intr_mutex);
1613 * Install devx shared asyncronous device events handler.
1614 * This function is implemeted to support event sharing
1615 * between multiple ports of single IB device.
1618 * Pointer to Ethernet device.
1621 mlx5_dev_shared_handler_devx_install(struct rte_eth_dev *dev)
1623 struct mlx5_priv *priv = dev->data->dev_private;
1624 struct mlx5_ibv_shared *sh = priv->sh;
1626 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1628 pthread_mutex_lock(&sh->intr_mutex);
1629 assert(priv->ibv_port);
1630 assert(priv->ibv_port <= sh->max_port);
1631 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1632 if (sh->port[priv->ibv_port - 1].devx_ih_port_id < RTE_MAX_ETHPORTS) {
1633 /* The handler is already installed for this port. */
1634 assert(sh->devx_intr_cnt);
1637 if (sh->devx_intr_cnt) {
1638 sh->devx_intr_cnt++;
1639 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1640 (uint32_t)dev->data->port_id;
1643 if (priv->config.devx) {
1644 #ifndef HAVE_IBV_DEVX_ASYNC
1647 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1648 if (sh->devx_comp) {
1649 int flags = fcntl(sh->devx_comp->fd, F_GETFL);
1650 int ret = fcntl(sh->devx_comp->fd, F_SETFL,
1651 flags | O_NONBLOCK);
1654 DRV_LOG(INFO, "failed to change file descriptor"
1655 " devx async event queue");
1657 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1658 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1659 rte_intr_callback_register
1660 (&sh->intr_handle_devx,
1661 mlx5_dev_interrupt_handler_devx, sh);
1662 sh->devx_intr_cnt++;
1663 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1664 (uint32_t)dev->data->port_id;
1667 #endif /* HAVE_IBV_DEVX_ASYNC */
1670 pthread_mutex_unlock(&sh->intr_mutex);
1674 * Uninstall interrupt handler.
1677 * Pointer to Ethernet device.
1680 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1682 mlx5_dev_shared_handler_uninstall(dev);
1686 * Install interrupt handler.
1689 * Pointer to Ethernet device.
1692 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1694 mlx5_dev_shared_handler_install(dev);
1698 * Devx uninstall interrupt handler.
1701 * Pointer to Ethernet device.
1704 mlx5_dev_interrupt_handler_devx_uninstall(struct rte_eth_dev *dev)
1706 mlx5_dev_shared_handler_devx_uninstall(dev);
1710 * Devx install interrupt handler.
1713 * Pointer to Ethernet device.
1716 mlx5_dev_interrupt_handler_devx_install(struct rte_eth_dev *dev)
1718 mlx5_dev_shared_handler_devx_install(dev);
1722 * DPDK callback to bring the link DOWN.
1725 * Pointer to Ethernet device structure.
1728 * 0 on success, a negative errno value otherwise and rte_errno is set.
1731 mlx5_set_link_down(struct rte_eth_dev *dev)
1733 return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1737 * DPDK callback to bring the link UP.
1740 * Pointer to Ethernet device structure.
1743 * 0 on success, a negative errno value otherwise and rte_errno is set.
1746 mlx5_set_link_up(struct rte_eth_dev *dev)
1748 return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1752 * Configure the RX function to use.
1755 * Pointer to private data structure.
1758 * Pointer to selected Rx burst function.
1761 mlx5_select_rx_function(struct rte_eth_dev *dev)
1763 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1765 assert(dev != NULL);
1766 if (mlx5_check_vec_rx_support(dev) > 0) {
1767 rx_pkt_burst = mlx5_rx_burst_vec;
1768 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1769 dev->data->port_id);
1770 } else if (mlx5_mprq_enabled(dev)) {
1771 rx_pkt_burst = mlx5_rx_burst_mprq;
1773 return rx_pkt_burst;
1777 * Check if mlx5 device was removed.
1780 * Pointer to Ethernet device structure.
1783 * 1 when device is removed, otherwise 0.
1786 mlx5_is_removed(struct rte_eth_dev *dev)
1788 struct ibv_device_attr device_attr;
1789 struct mlx5_priv *priv = dev->data->dev_private;
1791 if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1797 * Get the E-Switch parameters by port id.
1802 * Device port id is valid, skip check. This flag is useful
1803 * when trials are performed from probing and device is not
1804 * flagged as valid yet (in attaching process).
1805 * @param[out] es_domain_id
1806 * E-Switch domain id.
1807 * @param[out] es_port_id
1808 * The port id of the port in the E-Switch.
1811 * pointer to device private data structure containing data needed
1812 * on success, NULL otherwise and rte_errno is set.
1815 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
1817 struct rte_eth_dev *dev;
1818 struct mlx5_priv *priv;
1820 if (port >= RTE_MAX_ETHPORTS) {
1824 if (!valid && !rte_eth_dev_is_valid_port(port)) {
1828 dev = &rte_eth_devices[port];
1829 priv = dev->data->dev_private;
1830 if (!(priv->representor || priv->master)) {
1838 * Get the E-Switch parameters by device instance.
1842 * @param[out] es_domain_id
1843 * E-Switch domain id.
1844 * @param[out] es_port_id
1845 * The port id of the port in the E-Switch.
1848 * pointer to device private data structure containing data needed
1849 * on success, NULL otherwise and rte_errno is set.
1852 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1854 struct mlx5_priv *priv;
1856 priv = dev->data->dev_private;
1857 if (!(priv->representor || priv->master)) {
1865 * Get switch information associated with network interface.
1868 * Network interface index.
1870 * Switch information object, populated in case of success.
1873 * 0 on success, a negative errno value otherwise and rte_errno is set.
1876 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1878 char ifname[IF_NAMESIZE];
1879 char port_name[IF_NAMESIZE];
1881 struct mlx5_switch_info data = {
1884 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1889 bool port_switch_id_set = false;
1890 bool device_dir = false;
1894 if (!if_indextoname(ifindex, ifname)) {
1899 MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1901 MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1903 MKSTR(pci_device, "/sys/class/net/%s/device",
1906 file = fopen(phys_port_name, "rb");
1908 ret = fscanf(file, "%s", port_name);
1911 mlx5_translate_port_name(port_name, &data);
1913 file = fopen(phys_switch_id, "rb");
1918 port_switch_id_set =
1919 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1922 dir = opendir(pci_device);
1927 if (port_switch_id_set) {
1928 /* We have some E-Switch configuration. */
1929 mlx5_sysfs_check_switch_info(device_dir, &data);
1932 assert(!(data.master && data.representor));
1933 if (data.master && data.representor) {
1934 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1935 " and as representor", ifindex);
1943 * Analyze gathered port parameters via Netlink to recognize master
1944 * and representor devices for E-Switch configuration.
1946 * @param[in] num_vf_set
1947 * flag of presence of number of VFs port attribute.
1948 * @param[inout] switch_info
1949 * Port information, including port name as a number and port name
1950 * type if recognized
1953 * master and representor flags are set in switch_info according to
1954 * recognized parameters (if any).
1957 mlx5_nl_check_switch_info(bool num_vf_set,
1958 struct mlx5_switch_info *switch_info)
1960 switch (switch_info->name_type) {
1961 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1963 * Name is not recognized, assume the master,
1964 * check the number of VFs key presence.
1966 switch_info->master = num_vf_set;
1968 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1970 * Name is not set, this assumes the legacy naming
1971 * schema for master, just check if there is a
1972 * number of VFs key.
1974 switch_info->master = num_vf_set;
1976 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1977 /* New uplink naming schema recognized. */
1978 switch_info->master = 1;
1980 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1981 /* Legacy representors naming schema. */
1982 switch_info->representor = !num_vf_set;
1984 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1985 /* New representors naming schema. */
1986 switch_info->representor = 1;
1992 * Analyze gathered port parameters via sysfs to recognize master
1993 * and representor devices for E-Switch configuration.
1995 * @param[in] device_dir
1996 * flag of presence of "device" directory under port device key.
1997 * @param[inout] switch_info
1998 * Port information, including port name as a number and port name
1999 * type if recognized
2002 * master and representor flags are set in switch_info according to
2003 * recognized parameters (if any).
2006 mlx5_sysfs_check_switch_info(bool device_dir,
2007 struct mlx5_switch_info *switch_info)
2009 switch (switch_info->name_type) {
2010 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
2012 * Name is not recognized, assume the master,
2013 * check the device directory presence.
2015 switch_info->master = device_dir;
2017 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
2019 * Name is not set, this assumes the legacy naming
2020 * schema for master, just check if there is
2021 * a device directory.
2023 switch_info->master = device_dir;
2025 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2026 /* New uplink naming schema recognized. */
2027 switch_info->master = 1;
2029 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
2030 /* Legacy representors naming schema. */
2031 switch_info->representor = !device_dir;
2033 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2034 /* New representors naming schema. */
2035 switch_info->representor = 1;
2041 * Extract port name, as a number, from sysfs or netlink information.
2043 * @param[in] port_name_in
2044 * String representing the port name.
2045 * @param[out] port_info_out
2046 * Port information, including port name as a number and port name
2047 * type if recognized
2050 * port_name field set according to recognized name format.
2053 mlx5_translate_port_name(const char *port_name_in,
2054 struct mlx5_switch_info *port_info_out)
2056 char pf_c1, pf_c2, vf_c1, vf_c2;
2061 * Check for port-name as a string of the form pf0vf0
2062 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
2064 sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
2065 &pf_c1, &pf_c2, &port_info_out->pf_num,
2066 &vf_c1, &vf_c2, &port_info_out->port_name);
2067 if (sc_items == 6 &&
2068 pf_c1 == 'p' && pf_c2 == 'f' &&
2069 vf_c1 == 'v' && vf_c2 == 'f') {
2070 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
2074 * Check for port-name as a string of the form p0
2075 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
2077 sc_items = sscanf(port_name_in, "%c%d",
2078 &pf_c1, &port_info_out->port_name);
2079 if (sc_items == 2 && pf_c1 == 'p') {
2080 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
2083 /* Check for port-name as a number (support kernel ver < 5.0 */
2085 port_info_out->port_name = strtol(port_name_in, &end, 0);
2087 (size_t)(end - port_name_in) == strlen(port_name_in)) {
2088 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
2091 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
2096 * DPDK callback to retrieve plug-in module EEPROM information (type and size).
2099 * Pointer to Ethernet device structure.
2100 * @param[out] modinfo
2101 * Storage for plug-in module EEPROM information.
2104 * 0 on success, a negative errno value otherwise and rte_errno is set.
2107 mlx5_get_module_info(struct rte_eth_dev *dev,
2108 struct rte_eth_dev_module_info *modinfo)
2110 struct ethtool_modinfo info = {
2111 .cmd = ETHTOOL_GMODULEINFO,
2113 struct ifreq ifr = (struct ifreq) {
2114 .ifr_data = (void *)&info,
2118 if (!dev || !modinfo) {
2119 DRV_LOG(WARNING, "missing argument, cannot get module info");
2123 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2125 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2126 dev->data->port_id, strerror(rte_errno));
2129 modinfo->type = info.type;
2130 modinfo->eeprom_len = info.eeprom_len;
2135 * DPDK callback to retrieve plug-in module EEPROM data.
2138 * Pointer to Ethernet device structure.
2140 * Storage for plug-in module EEPROM data.
2143 * 0 on success, a negative errno value otherwise and rte_errno is set.
2145 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
2146 struct rte_dev_eeprom_info *info)
2148 struct ethtool_eeprom *eeprom;
2152 if (!dev || !info) {
2153 DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
2157 eeprom = rte_calloc(__func__, 1,
2158 (sizeof(struct ethtool_eeprom) + info->length), 0);
2160 DRV_LOG(WARNING, "port %u cannot allocate memory for "
2161 "eeprom data", dev->data->port_id);
2165 eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2166 eeprom->offset = info->offset;
2167 eeprom->len = info->length;
2168 ifr = (struct ifreq) {
2169 .ifr_data = (void *)eeprom,
2171 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2173 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2174 dev->data->port_id, strerror(rte_errno));
2176 rte_memcpy(info->data, eeprom->data, info->length);
2182 * DPDK callback to retrieve hairpin capabilities.
2185 * Pointer to Ethernet device structure.
2187 * Storage for hairpin capability data.
2190 * 0 on success, a negative errno value otherwise and rte_errno is set.
2192 int mlx5_hairpin_cap_get(struct rte_eth_dev *dev,
2193 struct rte_eth_hairpin_cap *cap)
2195 struct mlx5_priv *priv = dev->data->dev_private;
2197 if (priv->sh->devx == 0) {
2198 rte_errno = ENOTSUP;
2201 cap->max_nb_queues = UINT16_MAX;
2202 cap->max_rx_2_tx = 1;
2203 cap->max_tx_2_rx = 1;
2204 cap->max_nb_desc = 8192;