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);
407 memcpy(priv->rss_conf.rss_key,
409 dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
410 rss_hash_default_key,
411 MLX5_RSS_HASH_KEY_LEN);
412 priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
413 priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
414 priv->rxqs = (void *)dev->data->rx_queues;
415 priv->txqs = (void *)dev->data->tx_queues;
416 if (txqs_n != priv->txqs_n) {
417 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
418 dev->data->port_id, priv->txqs_n, txqs_n);
419 priv->txqs_n = txqs_n;
421 if (rxqs_n > priv->config.ind_table_max_size) {
422 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
423 dev->data->port_id, rxqs_n);
427 if (rxqs_n != priv->rxqs_n) {
428 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
429 dev->data->port_id, priv->rxqs_n, rxqs_n);
430 priv->rxqs_n = rxqs_n;
432 priv->skip_default_rss_reta = 0;
433 ret = mlx5_proc_priv_init(dev);
440 * Configure default RSS reta.
443 * Pointer to Ethernet device structure.
446 * 0 on success, a negative errno value otherwise and rte_errno is set.
449 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
451 struct mlx5_priv *priv = dev->data->dev_private;
452 unsigned int rxqs_n = dev->data->nb_rx_queues;
455 unsigned int reta_idx_n;
457 unsigned int *rss_queue_arr = NULL;
458 unsigned int rss_queue_n = 0;
460 if (priv->skip_default_rss_reta)
462 rss_queue_arr = rte_malloc("", rxqs_n * sizeof(unsigned int), 0);
463 if (!rss_queue_arr) {
464 DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
465 dev->data->port_id, rxqs_n);
469 for (i = 0, j = 0; i < rxqs_n; i++) {
470 struct mlx5_rxq_data *rxq_data;
471 struct mlx5_rxq_ctrl *rxq_ctrl;
473 rxq_data = (*priv->rxqs)[i];
474 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
475 if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
476 rss_queue_arr[j++] = i;
479 if (rss_queue_n > priv->config.ind_table_max_size) {
480 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
481 dev->data->port_id, rss_queue_n);
483 rte_free(rss_queue_arr);
486 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
487 dev->data->port_id, priv->rxqs_n, rxqs_n);
488 priv->rxqs_n = rxqs_n;
490 * If the requested number of RX queues is not a power of two,
491 * use the maximum indirection table size for better balancing.
492 * The result is always rounded to the next power of two.
494 reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
495 priv->config.ind_table_max_size :
497 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
499 rte_free(rss_queue_arr);
503 * When the number of RX queues is not a power of two,
504 * the remaining table entries are padded with reused WQs
505 * and hashes are not spread uniformly.
507 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
508 (*priv->reta_idx)[i] = rss_queue_arr[j];
509 if (++j == rss_queue_n)
512 rte_free(rss_queue_arr);
517 * Sets default tuning parameters.
520 * Pointer to Ethernet device.
522 * Info structure output buffer.
525 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
527 struct mlx5_priv *priv = dev->data->dev_private;
529 /* Minimum CPU utilization. */
530 info->default_rxportconf.ring_size = 256;
531 info->default_txportconf.ring_size = 256;
532 info->default_rxportconf.burst_size = 64;
533 info->default_txportconf.burst_size = 64;
534 if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
535 info->default_rxportconf.nb_queues = 16;
536 info->default_txportconf.nb_queues = 16;
537 if (dev->data->nb_rx_queues > 2 ||
538 dev->data->nb_tx_queues > 2) {
539 /* Max Throughput. */
540 info->default_rxportconf.ring_size = 2048;
541 info->default_txportconf.ring_size = 2048;
544 info->default_rxportconf.nb_queues = 8;
545 info->default_txportconf.nb_queues = 8;
546 if (dev->data->nb_rx_queues > 2 ||
547 dev->data->nb_tx_queues > 2) {
548 /* Max Throughput. */
549 info->default_rxportconf.ring_size = 4096;
550 info->default_txportconf.ring_size = 4096;
556 * Sets tx mbuf limiting parameters.
559 * Pointer to Ethernet device.
561 * Info structure output buffer.
564 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
566 struct mlx5_priv *priv = dev->data->dev_private;
567 struct mlx5_dev_config *config = &priv->config;
571 inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
572 MLX5_SEND_DEF_INLINE_LEN :
573 (unsigned int)config->txq_inline_max;
574 assert(config->txq_inline_min >= 0);
575 inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
576 inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
577 MLX5_ESEG_MIN_INLINE_SIZE -
580 MLX5_WQE_DSEG_SIZE * 2);
581 nb_max = (MLX5_WQE_SIZE_MAX +
582 MLX5_ESEG_MIN_INLINE_SIZE -
586 inlen) / MLX5_WSEG_SIZE;
587 info->tx_desc_lim.nb_seg_max = nb_max;
588 info->tx_desc_lim.nb_mtu_seg_max = nb_max;
592 * DPDK callback to get information about the device.
595 * Pointer to Ethernet device structure.
597 * Info structure output buffer.
600 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
602 struct mlx5_priv *priv = dev->data->dev_private;
603 struct mlx5_dev_config *config = &priv->config;
606 /* FIXME: we should ask the device for these values. */
607 info->min_rx_bufsize = 32;
608 info->max_rx_pktlen = 65536;
610 * Since we need one CQ per QP, the limit is the minimum number
611 * between the two values.
613 max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
614 priv->sh->device_attr.orig_attr.max_qp);
615 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
618 info->max_rx_queues = max;
619 info->max_tx_queues = max;
620 info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
621 info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
622 info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
623 info->rx_queue_offload_capa);
624 info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
625 info->if_index = mlx5_ifindex(dev);
626 info->reta_size = priv->reta_idx_n ?
627 priv->reta_idx_n : config->ind_table_max_size;
628 info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
629 info->speed_capa = priv->link_speed_capa;
630 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
631 mlx5_set_default_params(dev, info);
632 mlx5_set_txlimit_params(dev, info);
633 info->switch_info.name = dev->data->name;
634 info->switch_info.domain_id = priv->domain_id;
635 info->switch_info.port_id = priv->representor_id;
636 if (priv->representor) {
639 if (priv->pf_bond >= 0) {
641 * Switch port ID is opaque value with driver defined
642 * format. Push the PF index in bonding configurations
643 * in upper four bits of port ID. If we get too many
644 * representors (more than 4K) or PFs (more than 15)
645 * this approach must be reconsidered.
647 if ((info->switch_info.port_id >>
648 MLX5_PORT_ID_BONDING_PF_SHIFT) ||
649 priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
650 DRV_LOG(ERR, "can't update switch port ID"
651 " for bonding device");
655 info->switch_info.port_id |=
656 priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
658 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
659 struct mlx5_priv *opriv =
660 rte_eth_devices[port_id].data->dev_private;
663 opriv->representor ||
664 opriv->sh != priv->sh ||
665 opriv->domain_id != priv->domain_id)
668 * Override switch name with that of the master
671 info->switch_info.name = opriv->dev_data->name;
679 * Get device current raw clock counter
682 * Pointer to Ethernet device structure.
684 * Current raw clock counter of the device.
687 * 0 if the clock has correctly been read
688 * The value of errno in case of error
691 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
693 struct mlx5_priv *priv = dev->data->dev_private;
694 struct ibv_context *ctx = priv->sh->ctx;
695 struct ibv_values_ex values;
698 values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
699 err = mlx5_glue->query_rt_values_ex(ctx, &values);
701 DRV_LOG(WARNING, "Could not query the clock !");
704 *clock = values.raw_clock.tv_nsec;
709 * Get firmware version of a device.
712 * Ethernet device port.
714 * String output allocated by caller.
716 * Size of the output string, including terminating null byte.
719 * 0 on success, or the size of the non truncated string if too big.
721 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
723 struct mlx5_priv *priv = dev->data->dev_private;
724 struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
725 size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
730 strlcpy(fw_ver, attr->fw_ver, fw_size);
735 * Get supported packet types.
738 * Pointer to Ethernet device structure.
741 * A pointer to the supported Packet types array.
744 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
746 static const uint32_t ptypes[] = {
747 /* refers to rxq_cq_to_pkt_type() */
749 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
750 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
751 RTE_PTYPE_L4_NONFRAG,
755 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
756 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
757 RTE_PTYPE_INNER_L4_NONFRAG,
758 RTE_PTYPE_INNER_L4_FRAG,
759 RTE_PTYPE_INNER_L4_TCP,
760 RTE_PTYPE_INNER_L4_UDP,
764 if (dev->rx_pkt_burst == mlx5_rx_burst ||
765 dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
766 dev->rx_pkt_burst == mlx5_rx_burst_vec)
772 * Retrieve the master device for representor in the same switch domain.
775 * Pointer to representor Ethernet device structure.
778 * Master device structure on success, NULL otherwise.
781 static struct rte_eth_dev *
782 mlx5_find_master_dev(struct rte_eth_dev *dev)
784 struct mlx5_priv *priv;
788 priv = dev->data->dev_private;
789 domain_id = priv->domain_id;
790 assert(priv->representor);
791 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
792 struct mlx5_priv *opriv =
793 rte_eth_devices[port_id].data->dev_private;
796 opriv->domain_id == domain_id &&
797 opriv->sh == priv->sh)
798 return &rte_eth_devices[port_id];
804 * DPDK callback to retrieve physical link information.
807 * Pointer to Ethernet device structure.
809 * Storage for current link status.
812 * 0 on success, a negative errno value otherwise and rte_errno is set.
815 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
816 struct rte_eth_link *link)
818 struct mlx5_priv *priv = dev->data->dev_private;
819 struct ethtool_cmd edata = {
820 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
823 struct rte_eth_link dev_link;
827 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
829 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
830 dev->data->port_id, strerror(rte_errno));
833 dev_link = (struct rte_eth_link) {
834 .link_status = ((ifr.ifr_flags & IFF_UP) &&
835 (ifr.ifr_flags & IFF_RUNNING)),
837 ifr = (struct ifreq) {
838 .ifr_data = (void *)&edata,
840 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
842 if (ret == -ENOTSUP && priv->representor) {
843 struct rte_eth_dev *master;
846 * For representors we can try to inherit link
847 * settings from the master device. Actually
848 * link settings do not make a lot of sense
849 * for representors due to missing physical
850 * link. The old kernel drivers supported
851 * emulated settings query for representors,
852 * the new ones do not, so we have to add
853 * this code for compatibility issues.
855 master = mlx5_find_master_dev(dev);
857 ifr = (struct ifreq) {
858 .ifr_data = (void *)&edata,
860 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
865 "port %u ioctl(SIOCETHTOOL,"
866 " ETHTOOL_GSET) failed: %s",
867 dev->data->port_id, strerror(rte_errno));
871 link_speed = ethtool_cmd_speed(&edata);
872 if (link_speed == -1)
873 dev_link.link_speed = ETH_SPEED_NUM_NONE;
875 dev_link.link_speed = link_speed;
876 priv->link_speed_capa = 0;
877 if (edata.supported & SUPPORTED_Autoneg)
878 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
879 if (edata.supported & (SUPPORTED_1000baseT_Full |
880 SUPPORTED_1000baseKX_Full))
881 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
882 if (edata.supported & SUPPORTED_10000baseKR_Full)
883 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
884 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
885 SUPPORTED_40000baseCR4_Full |
886 SUPPORTED_40000baseSR4_Full |
887 SUPPORTED_40000baseLR4_Full))
888 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
889 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
890 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
891 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
892 ETH_LINK_SPEED_FIXED);
893 if (((dev_link.link_speed && !dev_link.link_status) ||
894 (!dev_link.link_speed && dev_link.link_status))) {
903 * Retrieve physical link information (unlocked version using new ioctl).
906 * Pointer to Ethernet device structure.
908 * Storage for current link status.
911 * 0 on success, a negative errno value otherwise and rte_errno is set.
914 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
915 struct rte_eth_link *link)
918 struct mlx5_priv *priv = dev->data->dev_private;
919 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
921 struct rte_eth_link dev_link;
922 struct rte_eth_dev *master = NULL;
926 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
928 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
929 dev->data->port_id, strerror(rte_errno));
932 dev_link = (struct rte_eth_link) {
933 .link_status = ((ifr.ifr_flags & IFF_UP) &&
934 (ifr.ifr_flags & IFF_RUNNING)),
936 ifr = (struct ifreq) {
937 .ifr_data = (void *)&gcmd,
939 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
941 if (ret == -ENOTSUP && priv->representor) {
943 * For representors we can try to inherit link
944 * settings from the master device. Actually
945 * link settings do not make a lot of sense
946 * for representors due to missing physical
947 * link. The old kernel drivers supported
948 * emulated settings query for representors,
949 * the new ones do not, so we have to add
950 * this code for compatibility issues.
952 master = mlx5_find_master_dev(dev);
954 ifr = (struct ifreq) {
955 .ifr_data = (void *)&gcmd,
957 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
962 "port %u ioctl(SIOCETHTOOL,"
963 " ETHTOOL_GLINKSETTINGS) failed: %s",
964 dev->data->port_id, strerror(rte_errno));
969 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
971 alignas(struct ethtool_link_settings)
972 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
973 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
974 struct ethtool_link_settings *ecmd = (void *)data;
977 ifr.ifr_data = (void *)ecmd;
978 ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
981 "port %u ioctl(SIOCETHTOOL,"
982 "ETHTOOL_GLINKSETTINGS) failed: %s",
983 dev->data->port_id, strerror(rte_errno));
986 dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
988 sc = ecmd->link_mode_masks[0] |
989 ((uint64_t)ecmd->link_mode_masks[1] << 32);
990 priv->link_speed_capa = 0;
991 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
992 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
993 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
994 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
995 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
996 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
997 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
998 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
999 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
1000 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
1001 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
1002 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
1003 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
1004 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
1005 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
1006 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
1007 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
1008 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
1009 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
1010 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
1011 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
1012 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
1013 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
1014 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
1015 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
1016 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
1017 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
1018 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
1019 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
1020 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
1021 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
1022 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
1023 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
1024 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
1025 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
1026 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
1027 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
1028 ETH_LINK_SPEED_FIXED);
1029 if (((dev_link.link_speed && !dev_link.link_status) ||
1030 (!dev_link.link_speed && dev_link.link_status))) {
1039 * DPDK callback to retrieve physical link information.
1042 * Pointer to Ethernet device structure.
1043 * @param wait_to_complete
1044 * Wait for request completion.
1047 * 0 if link status was not updated, positive if it was, a negative errno
1048 * value otherwise and rte_errno is set.
1051 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1054 struct rte_eth_link dev_link;
1055 time_t start_time = time(NULL);
1056 int retry = MLX5_GET_LINK_STATUS_RETRY_COUNT;
1059 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1060 if (ret == -ENOTSUP)
1061 ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1064 /* Handle wait to complete situation. */
1065 if ((wait_to_complete || retry) && ret == -EAGAIN) {
1066 if (abs((int)difftime(time(NULL), start_time)) <
1067 MLX5_LINK_STATUS_TIMEOUT) {
1074 } else if (ret < 0) {
1077 } while (wait_to_complete || retry-- > 0);
1078 ret = !!memcmp(&dev->data->dev_link, &dev_link,
1079 sizeof(struct rte_eth_link));
1080 dev->data->dev_link = dev_link;
1085 * DPDK callback to change the MTU.
1088 * Pointer to Ethernet device structure.
1093 * 0 on success, a negative errno value otherwise and rte_errno is set.
1096 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1098 struct mlx5_priv *priv = dev->data->dev_private;
1099 uint16_t kern_mtu = 0;
1102 ret = mlx5_get_mtu(dev, &kern_mtu);
1105 /* Set kernel interface MTU first. */
1106 ret = mlx5_set_mtu(dev, mtu);
1109 ret = mlx5_get_mtu(dev, &kern_mtu);
1112 if (kern_mtu == mtu) {
1114 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1115 dev->data->port_id, mtu);
1123 * DPDK callback to get flow control status.
1126 * Pointer to Ethernet device structure.
1127 * @param[out] fc_conf
1128 * Flow control output buffer.
1131 * 0 on success, a negative errno value otherwise and rte_errno is set.
1134 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1137 struct ethtool_pauseparam ethpause = {
1138 .cmd = ETHTOOL_GPAUSEPARAM
1142 ifr.ifr_data = (void *)ðpause;
1143 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1146 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1148 dev->data->port_id, strerror(rte_errno));
1151 fc_conf->autoneg = ethpause.autoneg;
1152 if (ethpause.rx_pause && ethpause.tx_pause)
1153 fc_conf->mode = RTE_FC_FULL;
1154 else if (ethpause.rx_pause)
1155 fc_conf->mode = RTE_FC_RX_PAUSE;
1156 else if (ethpause.tx_pause)
1157 fc_conf->mode = RTE_FC_TX_PAUSE;
1159 fc_conf->mode = RTE_FC_NONE;
1164 * DPDK callback to modify flow control parameters.
1167 * Pointer to Ethernet device structure.
1168 * @param[in] fc_conf
1169 * Flow control parameters.
1172 * 0 on success, a negative errno value otherwise and rte_errno is set.
1175 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1178 struct ethtool_pauseparam ethpause = {
1179 .cmd = ETHTOOL_SPAUSEPARAM
1183 ifr.ifr_data = (void *)ðpause;
1184 ethpause.autoneg = fc_conf->autoneg;
1185 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1186 (fc_conf->mode & RTE_FC_RX_PAUSE))
1187 ethpause.rx_pause = 1;
1189 ethpause.rx_pause = 0;
1191 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1192 (fc_conf->mode & RTE_FC_TX_PAUSE))
1193 ethpause.tx_pause = 1;
1195 ethpause.tx_pause = 0;
1196 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1199 "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1201 dev->data->port_id, strerror(rte_errno));
1208 * Get PCI information by sysfs device path.
1211 * Pointer to device sysfs folder name.
1212 * @param[out] pci_addr
1213 * PCI bus address output buffer.
1216 * 0 on success, a negative errno value otherwise and rte_errno is set.
1219 mlx5_dev_to_pci_addr(const char *dev_path,
1220 struct rte_pci_addr *pci_addr)
1224 MKSTR(path, "%s/device/uevent", dev_path);
1226 file = fopen(path, "rb");
1231 while (fgets(line, sizeof(line), file) == line) {
1232 size_t len = strlen(line);
1235 /* Truncate long lines. */
1236 if (len == (sizeof(line) - 1))
1237 while (line[(len - 1)] != '\n') {
1241 line[(len - 1)] = ret;
1243 /* Extract information. */
1246 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1250 &pci_addr->function) == 4) {
1260 * Handle asynchronous removal event for entire multiport device.
1263 * Infiniband device shared context.
1266 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1270 for (i = 0; i < sh->max_port; ++i) {
1271 struct rte_eth_dev *dev;
1273 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1275 * Or not existing port either no
1276 * handler installed for this port.
1280 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1282 if (dev->data->dev_conf.intr_conf.rmv)
1283 _rte_eth_dev_callback_process
1284 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1289 * Handle shared asynchronous events the NIC (removal event
1290 * and link status change). Supports multiport IB device.
1293 * Callback argument.
1296 mlx5_dev_interrupt_handler(void *cb_arg)
1298 struct mlx5_ibv_shared *sh = cb_arg;
1299 struct ibv_async_event event;
1301 /* Read all message from the IB device and acknowledge them. */
1303 struct rte_eth_dev *dev;
1306 if (mlx5_glue->get_async_event(sh->ctx, &event))
1308 /* Retrieve and check IB port index. */
1309 tmp = (uint32_t)event.element.port_num;
1310 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1312 * The DEVICE_FATAL event is called once for
1313 * entire device without port specifying.
1314 * We should notify all existing ports.
1316 mlx5_glue->ack_async_event(&event);
1317 mlx5_dev_interrupt_device_fatal(sh);
1320 assert(tmp && (tmp <= sh->max_port));
1322 /* Unsupported devive level event. */
1323 mlx5_glue->ack_async_event(&event);
1325 "unsupported common event (type %d)",
1329 if (tmp > sh->max_port) {
1330 /* Invalid IB port index. */
1331 mlx5_glue->ack_async_event(&event);
1333 "cannot handle an event (type %d)"
1334 "due to invalid IB port index (%u)",
1335 event.event_type, tmp);
1338 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1339 /* No handler installed. */
1340 mlx5_glue->ack_async_event(&event);
1342 "cannot handle an event (type %d)"
1343 "due to no handler installed for port %u",
1344 event.event_type, tmp);
1347 /* Retrieve ethernet device descriptor. */
1348 tmp = sh->port[tmp - 1].ih_port_id;
1349 dev = &rte_eth_devices[tmp];
1351 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1352 event.event_type == IBV_EVENT_PORT_ERR) &&
1353 dev->data->dev_conf.intr_conf.lsc) {
1354 mlx5_glue->ack_async_event(&event);
1355 if (mlx5_link_update(dev, 0) == -EAGAIN) {
1359 _rte_eth_dev_callback_process
1360 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1364 "port %u cannot handle an unknown event (type %d)",
1365 dev->data->port_id, event.event_type);
1366 mlx5_glue->ack_async_event(&event);
1371 * Unregister callback handler safely. The handler may be active
1372 * while we are trying to unregister it, in this case code -EAGAIN
1373 * is returned by rte_intr_callback_unregister(). This routine checks
1374 * the return code and tries to unregister handler again.
1379 * pointer to callback routine
1381 * opaque callback parameter
1384 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1385 rte_intr_callback_fn cb_fn, void *cb_arg)
1388 * Try to reduce timeout management overhead by not calling
1389 * the timer related routines on the first iteration. If the
1390 * unregistering succeeds on first call there will be no
1391 * timer calls at all.
1399 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1402 if (ret != -EAGAIN) {
1403 DRV_LOG(INFO, "failed to unregister interrupt"
1404 " handler (error: %d)", ret);
1409 struct timespec onems;
1411 /* Wait one millisecond and try again. */
1413 onems.tv_nsec = NS_PER_S / MS_PER_S;
1414 nanosleep(&onems, 0);
1415 /* Check whether one second elapsed. */
1416 if ((rte_get_timer_cycles() - start) <= twait)
1420 * We get the amount of timer ticks for one second.
1421 * If this amount elapsed it means we spent one
1422 * second in waiting. This branch is executed once
1423 * on first iteration.
1425 twait = rte_get_timer_hz();
1429 * Timeout elapsed, show message (once a second) and retry.
1430 * We have no other acceptable option here, if we ignore
1431 * the unregistering return code the handler will not
1432 * be unregistered, fd will be closed and we may get the
1433 * crush. Hanging and messaging in the loop seems not to be
1436 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1437 start = rte_get_timer_cycles();
1442 * Handle DEVX interrupts from the NIC.
1443 * This function is probably called from the DPDK host thread.
1446 * Callback argument.
1449 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1451 #ifndef HAVE_IBV_DEVX_ASYNC
1455 struct mlx5_ibv_shared *sh = cb_arg;
1457 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1458 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1459 MLX5_ST_SZ_BYTES(traffic_counter) +
1460 sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1462 uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1464 while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1467 mlx5_flow_async_pool_query_handle
1468 (sh, (uint64_t)out.cmd_resp.wr_id,
1469 mlx5_devx_get_out_command_status(buf));
1470 #endif /* HAVE_IBV_DEVX_ASYNC */
1474 * Uninstall shared asynchronous device events handler.
1475 * This function is implemented to support event sharing
1476 * between multiple ports of single IB device.
1479 * Pointer to Ethernet device.
1482 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1484 struct mlx5_priv *priv = dev->data->dev_private;
1485 struct mlx5_ibv_shared *sh = priv->sh;
1487 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1489 pthread_mutex_lock(&sh->intr_mutex);
1490 assert(priv->ibv_port);
1491 assert(priv->ibv_port <= sh->max_port);
1492 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1493 if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1495 assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1496 (uint32_t)dev->data->port_id);
1497 assert(sh->intr_cnt);
1498 sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1499 if (!sh->intr_cnt || --sh->intr_cnt)
1501 mlx5_intr_callback_unregister(&sh->intr_handle,
1502 mlx5_dev_interrupt_handler, sh);
1503 sh->intr_handle.fd = 0;
1504 sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1506 pthread_mutex_unlock(&sh->intr_mutex);
1510 * Uninstall devx shared asynchronous device events handler.
1511 * This function is implemeted to support event sharing
1512 * between multiple ports of single IB device.
1515 * Pointer to Ethernet device.
1518 mlx5_dev_shared_handler_devx_uninstall(struct rte_eth_dev *dev)
1520 struct mlx5_priv *priv = dev->data->dev_private;
1521 struct mlx5_ibv_shared *sh = priv->sh;
1523 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1525 pthread_mutex_lock(&sh->intr_mutex);
1526 assert(priv->ibv_port);
1527 assert(priv->ibv_port <= sh->max_port);
1528 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1529 if (sh->port[priv->ibv_port - 1].devx_ih_port_id >= RTE_MAX_ETHPORTS)
1531 assert(sh->port[priv->ibv_port - 1].devx_ih_port_id ==
1532 (uint32_t)dev->data->port_id);
1533 sh->port[priv->ibv_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1534 if (!sh->devx_intr_cnt || --sh->devx_intr_cnt)
1536 if (sh->intr_handle_devx.fd) {
1537 rte_intr_callback_unregister(&sh->intr_handle_devx,
1538 mlx5_dev_interrupt_handler_devx,
1540 sh->intr_handle_devx.fd = 0;
1541 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1543 if (sh->devx_comp) {
1544 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1545 sh->devx_comp = NULL;
1548 pthread_mutex_unlock(&sh->intr_mutex);
1552 * Install shared asynchronous device events handler.
1553 * This function is implemented to support event sharing
1554 * between multiple ports of single IB device.
1557 * Pointer to Ethernet device.
1560 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1562 struct mlx5_priv *priv = dev->data->dev_private;
1563 struct mlx5_ibv_shared *sh = priv->sh;
1567 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1569 pthread_mutex_lock(&sh->intr_mutex);
1570 assert(priv->ibv_port);
1571 assert(priv->ibv_port <= sh->max_port);
1572 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1573 if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1574 /* The handler is already installed for this port. */
1575 assert(sh->intr_cnt);
1579 sh->port[priv->ibv_port - 1].ih_port_id =
1580 (uint32_t)dev->data->port_id;
1584 /* No shared handler installed. */
1585 assert(sh->ctx->async_fd > 0);
1586 flags = fcntl(sh->ctx->async_fd, F_GETFL);
1587 ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1589 DRV_LOG(INFO, "failed to change file descriptor async event"
1591 /* Indicate there will be no interrupts. */
1592 dev->data->dev_conf.intr_conf.lsc = 0;
1593 dev->data->dev_conf.intr_conf.rmv = 0;
1595 sh->intr_handle.fd = sh->ctx->async_fd;
1596 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1597 rte_intr_callback_register(&sh->intr_handle,
1598 mlx5_dev_interrupt_handler, sh);
1600 sh->port[priv->ibv_port - 1].ih_port_id =
1601 (uint32_t)dev->data->port_id;
1604 pthread_mutex_unlock(&sh->intr_mutex);
1608 * Install devx shared asyncronous device events handler.
1609 * This function is implemeted to support event sharing
1610 * between multiple ports of single IB device.
1613 * Pointer to Ethernet device.
1616 mlx5_dev_shared_handler_devx_install(struct rte_eth_dev *dev)
1618 struct mlx5_priv *priv = dev->data->dev_private;
1619 struct mlx5_ibv_shared *sh = priv->sh;
1621 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1623 pthread_mutex_lock(&sh->intr_mutex);
1624 assert(priv->ibv_port);
1625 assert(priv->ibv_port <= sh->max_port);
1626 assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1627 if (sh->port[priv->ibv_port - 1].devx_ih_port_id < RTE_MAX_ETHPORTS) {
1628 /* The handler is already installed for this port. */
1629 assert(sh->devx_intr_cnt);
1632 if (sh->devx_intr_cnt) {
1633 sh->devx_intr_cnt++;
1634 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1635 (uint32_t)dev->data->port_id;
1638 if (priv->config.devx) {
1639 #ifndef HAVE_IBV_DEVX_ASYNC
1642 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1643 if (sh->devx_comp) {
1644 int flags = fcntl(sh->devx_comp->fd, F_GETFL);
1645 int ret = fcntl(sh->devx_comp->fd, F_SETFL,
1646 flags | O_NONBLOCK);
1649 DRV_LOG(INFO, "failed to change file descriptor"
1650 " devx async event queue");
1652 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1653 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1654 rte_intr_callback_register
1655 (&sh->intr_handle_devx,
1656 mlx5_dev_interrupt_handler_devx, sh);
1657 sh->devx_intr_cnt++;
1658 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1659 (uint32_t)dev->data->port_id;
1662 #endif /* HAVE_IBV_DEVX_ASYNC */
1665 pthread_mutex_unlock(&sh->intr_mutex);
1669 * Uninstall interrupt handler.
1672 * Pointer to Ethernet device.
1675 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1677 mlx5_dev_shared_handler_uninstall(dev);
1681 * Install interrupt handler.
1684 * Pointer to Ethernet device.
1687 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1689 mlx5_dev_shared_handler_install(dev);
1693 * Devx uninstall interrupt handler.
1696 * Pointer to Ethernet device.
1699 mlx5_dev_interrupt_handler_devx_uninstall(struct rte_eth_dev *dev)
1701 mlx5_dev_shared_handler_devx_uninstall(dev);
1705 * Devx install interrupt handler.
1708 * Pointer to Ethernet device.
1711 mlx5_dev_interrupt_handler_devx_install(struct rte_eth_dev *dev)
1713 mlx5_dev_shared_handler_devx_install(dev);
1717 * DPDK callback to bring the link DOWN.
1720 * Pointer to Ethernet device structure.
1723 * 0 on success, a negative errno value otherwise and rte_errno is set.
1726 mlx5_set_link_down(struct rte_eth_dev *dev)
1728 return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1732 * DPDK callback to bring the link UP.
1735 * Pointer to Ethernet device structure.
1738 * 0 on success, a negative errno value otherwise and rte_errno is set.
1741 mlx5_set_link_up(struct rte_eth_dev *dev)
1743 return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1747 * Configure the RX function to use.
1750 * Pointer to private data structure.
1753 * Pointer to selected Rx burst function.
1756 mlx5_select_rx_function(struct rte_eth_dev *dev)
1758 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1760 assert(dev != NULL);
1761 if (mlx5_check_vec_rx_support(dev) > 0) {
1762 rx_pkt_burst = mlx5_rx_burst_vec;
1763 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1764 dev->data->port_id);
1765 } else if (mlx5_mprq_enabled(dev)) {
1766 rx_pkt_burst = mlx5_rx_burst_mprq;
1768 return rx_pkt_burst;
1772 * Check if mlx5 device was removed.
1775 * Pointer to Ethernet device structure.
1778 * 1 when device is removed, otherwise 0.
1781 mlx5_is_removed(struct rte_eth_dev *dev)
1783 struct ibv_device_attr device_attr;
1784 struct mlx5_priv *priv = dev->data->dev_private;
1786 if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1792 * Get the E-Switch parameters by port id.
1797 * Device port id is valid, skip check. This flag is useful
1798 * when trials are performed from probing and device is not
1799 * flagged as valid yet (in attaching process).
1800 * @param[out] es_domain_id
1801 * E-Switch domain id.
1802 * @param[out] es_port_id
1803 * The port id of the port in the E-Switch.
1806 * pointer to device private data structure containing data needed
1807 * on success, NULL otherwise and rte_errno is set.
1810 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
1812 struct rte_eth_dev *dev;
1813 struct mlx5_priv *priv;
1815 if (port >= RTE_MAX_ETHPORTS) {
1819 if (!valid && !rte_eth_dev_is_valid_port(port)) {
1823 dev = &rte_eth_devices[port];
1824 priv = dev->data->dev_private;
1825 if (!(priv->representor || priv->master)) {
1833 * Get the E-Switch parameters by device instance.
1837 * @param[out] es_domain_id
1838 * E-Switch domain id.
1839 * @param[out] es_port_id
1840 * The port id of the port in the E-Switch.
1843 * pointer to device private data structure containing data needed
1844 * on success, NULL otherwise and rte_errno is set.
1847 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1849 struct mlx5_priv *priv;
1851 priv = dev->data->dev_private;
1852 if (!(priv->representor || priv->master)) {
1860 * Get switch information associated with network interface.
1863 * Network interface index.
1865 * Switch information object, populated in case of success.
1868 * 0 on success, a negative errno value otherwise and rte_errno is set.
1871 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1873 char ifname[IF_NAMESIZE];
1874 char port_name[IF_NAMESIZE];
1876 struct mlx5_switch_info data = {
1879 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1884 bool port_switch_id_set = false;
1885 bool device_dir = false;
1889 if (!if_indextoname(ifindex, ifname)) {
1894 MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1896 MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1898 MKSTR(pci_device, "/sys/class/net/%s/device",
1901 file = fopen(phys_port_name, "rb");
1903 ret = fscanf(file, "%s", port_name);
1906 mlx5_translate_port_name(port_name, &data);
1908 file = fopen(phys_switch_id, "rb");
1913 port_switch_id_set =
1914 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1917 dir = opendir(pci_device);
1922 if (port_switch_id_set) {
1923 /* We have some E-Switch configuration. */
1924 mlx5_sysfs_check_switch_info(device_dir, &data);
1927 assert(!(data.master && data.representor));
1928 if (data.master && data.representor) {
1929 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1930 " and as representor", ifindex);
1938 * Analyze gathered port parameters via Netlink to recognize master
1939 * and representor devices for E-Switch configuration.
1941 * @param[in] num_vf_set
1942 * flag of presence of number of VFs port attribute.
1943 * @param[inout] switch_info
1944 * Port information, including port name as a number and port name
1945 * type if recognized
1948 * master and representor flags are set in switch_info according to
1949 * recognized parameters (if any).
1952 mlx5_nl_check_switch_info(bool num_vf_set,
1953 struct mlx5_switch_info *switch_info)
1955 switch (switch_info->name_type) {
1956 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1958 * Name is not recognized, assume the master,
1959 * check the number of VFs key presence.
1961 switch_info->master = num_vf_set;
1963 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1965 * Name is not set, this assumes the legacy naming
1966 * schema for master, just check if there is a
1967 * number of VFs key.
1969 switch_info->master = num_vf_set;
1971 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1972 /* New uplink naming schema recognized. */
1973 switch_info->master = 1;
1975 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1976 /* Legacy representors naming schema. */
1977 switch_info->representor = !num_vf_set;
1979 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1980 /* New representors naming schema. */
1981 switch_info->representor = 1;
1987 * Analyze gathered port parameters via sysfs to recognize master
1988 * and representor devices for E-Switch configuration.
1990 * @param[in] device_dir
1991 * flag of presence of "device" directory under port device key.
1992 * @param[inout] switch_info
1993 * Port information, including port name as a number and port name
1994 * type if recognized
1997 * master and representor flags are set in switch_info according to
1998 * recognized parameters (if any).
2001 mlx5_sysfs_check_switch_info(bool device_dir,
2002 struct mlx5_switch_info *switch_info)
2004 switch (switch_info->name_type) {
2005 case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
2007 * Name is not recognized, assume the master,
2008 * check the device directory presence.
2010 switch_info->master = device_dir;
2012 case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
2014 * Name is not set, this assumes the legacy naming
2015 * schema for master, just check if there is
2016 * a device directory.
2018 switch_info->master = device_dir;
2020 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2021 /* New uplink naming schema recognized. */
2022 switch_info->master = 1;
2024 case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
2025 /* Legacy representors naming schema. */
2026 switch_info->representor = !device_dir;
2028 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2029 /* New representors naming schema. */
2030 switch_info->representor = 1;
2036 * Extract port name, as a number, from sysfs or netlink information.
2038 * @param[in] port_name_in
2039 * String representing the port name.
2040 * @param[out] port_info_out
2041 * Port information, including port name as a number and port name
2042 * type if recognized
2045 * port_name field set according to recognized name format.
2048 mlx5_translate_port_name(const char *port_name_in,
2049 struct mlx5_switch_info *port_info_out)
2051 char pf_c1, pf_c2, vf_c1, vf_c2;
2056 * Check for port-name as a string of the form pf0vf0
2057 * (support kernel ver >= 5.0 or OFED ver >= 4.6).
2059 sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
2060 &pf_c1, &pf_c2, &port_info_out->pf_num,
2061 &vf_c1, &vf_c2, &port_info_out->port_name);
2062 if (sc_items == 6 &&
2063 pf_c1 == 'p' && pf_c2 == 'f' &&
2064 vf_c1 == 'v' && vf_c2 == 'f') {
2065 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
2069 * Check for port-name as a string of the form p0
2070 * (support kernel ver >= 5.0, or OFED ver >= 4.6).
2072 sc_items = sscanf(port_name_in, "%c%d",
2073 &pf_c1, &port_info_out->port_name);
2074 if (sc_items == 2 && pf_c1 == 'p') {
2075 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
2078 /* Check for port-name as a number (support kernel ver < 5.0 */
2080 port_info_out->port_name = strtol(port_name_in, &end, 0);
2082 (size_t)(end - port_name_in) == strlen(port_name_in)) {
2083 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
2086 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
2091 * DPDK callback to retrieve plug-in module EEPROM information (type and size).
2094 * Pointer to Ethernet device structure.
2095 * @param[out] modinfo
2096 * Storage for plug-in module EEPROM information.
2099 * 0 on success, a negative errno value otherwise and rte_errno is set.
2102 mlx5_get_module_info(struct rte_eth_dev *dev,
2103 struct rte_eth_dev_module_info *modinfo)
2105 struct ethtool_modinfo info = {
2106 .cmd = ETHTOOL_GMODULEINFO,
2108 struct ifreq ifr = (struct ifreq) {
2109 .ifr_data = (void *)&info,
2113 if (!dev || !modinfo) {
2114 DRV_LOG(WARNING, "missing argument, cannot get module info");
2118 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2120 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2121 dev->data->port_id, strerror(rte_errno));
2124 modinfo->type = info.type;
2125 modinfo->eeprom_len = info.eeprom_len;
2130 * DPDK callback to retrieve plug-in module EEPROM data.
2133 * Pointer to Ethernet device structure.
2135 * Storage for plug-in module EEPROM data.
2138 * 0 on success, a negative errno value otherwise and rte_errno is set.
2140 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
2141 struct rte_dev_eeprom_info *info)
2143 struct ethtool_eeprom *eeprom;
2147 if (!dev || !info) {
2148 DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
2152 eeprom = rte_calloc(__func__, 1,
2153 (sizeof(struct ethtool_eeprom) + info->length), 0);
2155 DRV_LOG(WARNING, "port %u cannot allocate memory for "
2156 "eeprom data", dev->data->port_id);
2160 eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2161 eeprom->offset = info->offset;
2162 eeprom->len = info->length;
2163 ifr = (struct ifreq) {
2164 .ifr_data = (void *)eeprom,
2166 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2168 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2169 dev->data->port_id, strerror(rte_errno));
2171 rte_memcpy(info->data, eeprom->data, info->length);
2177 * DPDK callback to retrieve hairpin capabilities.
2180 * Pointer to Ethernet device structure.
2182 * Storage for hairpin capability data.
2185 * 0 on success, a negative errno value otherwise and rte_errno is set.
2187 int mlx5_hairpin_cap_get(struct rte_eth_dev *dev,
2188 struct rte_eth_hairpin_cap *cap)
2190 struct mlx5_priv *priv = dev->data->dev_private;
2192 if (priv->sh->devx == 0) {
2193 rte_errno = ENOTSUP;
2196 cap->max_nb_queues = UINT16_MAX;
2197 cap->max_rx_2_tx = 1;
2198 cap->max_tx_2_rx = 1;
2199 cap->max_nb_desc = 8192;