4 * Copyright 2015 6WIND S.A.
5 * Copyright 2015 Mellanox.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of 6WIND S.A. nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46 #include <sys/ioctl.h>
47 #include <sys/socket.h>
48 #include <sys/utsname.h>
49 #include <netinet/in.h>
50 #include <linux/ethtool.h>
51 #include <linux/sockios.h>
52 #include <linux/version.h>
57 #include <rte_atomic.h>
58 #include <rte_ethdev_driver.h>
59 #include <rte_bus_pci.h>
61 #include <rte_common.h>
62 #include <rte_interrupts.h>
63 #include <rte_alarm.h>
64 #include <rte_malloc.h>
67 #include "mlx5_rxtx.h"
68 #include "mlx5_utils.h"
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 interface name from private structure.
125 * Pointer to private structure.
127 * Interface name output buffer.
130 * 0 on success, -1 on failure and errno is set.
133 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
137 unsigned int dev_type = 0;
138 unsigned int dev_port_prev = ~0u;
139 char match[IF_NAMESIZE] = "";
142 MKSTR(path, "%s/device/net", priv->ibdev_path);
148 while ((dent = readdir(dir)) != NULL) {
149 char *name = dent->d_name;
151 unsigned int dev_port;
154 if ((name[0] == '.') &&
155 ((name[1] == '\0') ||
156 ((name[1] == '.') && (name[2] == '\0'))))
159 MKSTR(path, "%s/device/net/%s/%s",
160 priv->ibdev_path, name,
161 (dev_type ? "dev_id" : "dev_port"));
163 file = fopen(path, "rb");
168 * Switch to dev_id when dev_port does not exist as
169 * is the case with Linux kernel versions < 3.15.
180 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
185 * Switch to dev_id when dev_port returns the same value for
186 * all ports. May happen when using a MOFED release older than
187 * 3.0 with a Linux kernel >= 3.15.
189 if (dev_port == dev_port_prev)
191 dev_port_prev = dev_port;
192 if (dev_port == (priv->port - 1u))
193 snprintf(match, sizeof(match), "%s", name);
196 if (match[0] == '\0')
198 strncpy(*ifname, match, sizeof(*ifname));
203 * Check if the counter is located on ib counters file.
209 * 1 if counter is located on ib counters file , 0 otherwise.
212 priv_is_ib_cntr(const char *cntr)
214 if (!strcmp(cntr, "out_of_buffer"))
220 * Read from sysfs entry.
223 * Pointer to private structure.
225 * Entry name relative to sysfs path.
227 * Data output buffer.
232 * 0 on success, -1 on failure and errno is set.
235 priv_sysfs_read(const struct priv *priv, const char *entry,
236 char *buf, size_t size)
238 char ifname[IF_NAMESIZE];
243 if (priv_get_ifname(priv, &ifname))
246 if (priv_is_ib_cntr(entry)) {
247 MKSTR(path, "%s/ports/1/hw_counters/%s",
248 priv->ibdev_path, entry);
249 file = fopen(path, "rb");
251 MKSTR(path, "%s/device/net/%s/%s",
252 priv->ibdev_path, ifname, entry);
253 file = fopen(path, "rb");
257 ret = fread(buf, 1, size, file);
259 if (((size_t)ret < size) && (ferror(file)))
269 * Write to sysfs entry.
272 * Pointer to private structure.
274 * Entry name relative to sysfs path.
281 * 0 on success, -1 on failure and errno is set.
284 priv_sysfs_write(const struct priv *priv, const char *entry,
285 char *buf, size_t size)
287 char ifname[IF_NAMESIZE];
292 if (priv_get_ifname(priv, &ifname))
295 MKSTR(path, "%s/device/net/%s/%s", priv->ibdev_path, ifname, entry);
297 file = fopen(path, "wb");
300 ret = fwrite(buf, 1, size, file);
302 if (((size_t)ret < size) || (ferror(file)))
312 * Get unsigned long sysfs property.
315 * Pointer to private structure.
317 * Entry name relative to sysfs path.
319 * Value output buffer.
322 * 0 on success, -1 on failure and errno is set.
325 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
328 unsigned long value_ret;
331 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
333 DEBUG("cannot read %s value from sysfs: %s",
334 name, strerror(errno));
337 value_str[ret] = '\0';
339 value_ret = strtoul(value_str, NULL, 0);
341 DEBUG("invalid %s value `%s': %s", name, value_str,
350 * Set unsigned long sysfs property.
353 * Pointer to private structure.
355 * Entry name relative to sysfs path.
360 * 0 on success, -1 on failure and errno is set.
363 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
366 MKSTR(value_str, "%lu", value);
368 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
370 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
371 name, value_str, value, strerror(errno));
378 * Perform ifreq ioctl() on associated Ethernet device.
381 * Pointer to private structure.
383 * Request number to pass to ioctl().
385 * Interface request structure output buffer.
388 * 0 on success, -1 on failure and errno is set.
391 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
393 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
398 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
399 ret = ioctl(sock, req, ifr);
405 * Return the number of active VFs for the current device.
408 * Pointer to private structure.
409 * @param[out] num_vfs
410 * Number of active VFs.
413 * 0 on success, -1 on failure and errno is set.
416 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
418 /* The sysfs entry name depends on the operating system. */
419 const char **name = (const char *[]){
420 "device/sriov_numvfs",
421 "device/mlx5_num_vfs",
427 unsigned long ulong_num_vfs;
429 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
431 *num_vfs = ulong_num_vfs;
432 } while (*(++name) && ret);
440 * Pointer to private structure.
442 * MTU value output buffer.
445 * 0 on success, -1 on failure and errno is set.
448 priv_get_mtu(struct priv *priv, uint16_t *mtu)
450 unsigned long ulong_mtu;
452 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
459 * Read device counter from sysfs.
462 * Pointer to private structure.
466 * Counter output buffer.
469 * 0 on success, -1 on failure and errno is set.
472 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr)
474 unsigned long ulong_ctr;
476 if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1)
486 * Pointer to private structure.
491 * 0 on success, -1 on failure and errno is set.
494 priv_set_mtu(struct priv *priv, uint16_t mtu)
498 if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
499 priv_get_mtu(priv, &new_mtu))
511 * Pointer to private structure.
513 * Bitmask for flags that must remain untouched.
515 * Bitmask for flags to modify.
518 * 0 on success, -1 on failure and errno is set.
521 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
525 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
528 tmp |= (flags & (~keep));
529 return priv_set_sysfs_ulong(priv, "flags", tmp);
533 * Ethernet device configuration.
535 * Prepare the driver for a given number of TX and RX queues.
538 * Pointer to Ethernet device structure.
541 * 0 on success, errno value on failure.
544 dev_configure(struct rte_eth_dev *dev)
546 struct priv *priv = dev->data->dev_private;
547 unsigned int rxqs_n = dev->data->nb_rx_queues;
548 unsigned int txqs_n = dev->data->nb_tx_queues;
551 unsigned int reta_idx_n;
552 const uint8_t use_app_rss_key =
553 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
554 uint64_t supp_tx_offloads = mlx5_priv_get_tx_port_offloads(priv);
555 uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
556 uint64_t supp_rx_offloads =
557 (mlx5_priv_get_rx_port_offloads(priv) |
558 mlx5_priv_get_rx_queue_offloads(priv));
559 uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
561 if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
562 ERROR("Some Tx offloads are not supported "
563 "requested 0x%" PRIx64 " supported 0x%" PRIx64,
564 tx_offloads, supp_tx_offloads);
567 if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
568 ERROR("Some Rx offloads are not supported "
569 "requested 0x%" PRIx64 " supported 0x%" PRIx64,
570 rx_offloads, supp_rx_offloads);
573 if (use_app_rss_key &&
574 (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
575 rss_hash_default_key_len)) {
576 /* MLX5 RSS only support 40bytes key. */
579 priv->rss_conf.rss_key =
580 rte_realloc(priv->rss_conf.rss_key,
581 rss_hash_default_key_len, 0);
582 if (!priv->rss_conf.rss_key) {
583 ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
586 memcpy(priv->rss_conf.rss_key,
588 dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
589 rss_hash_default_key,
590 rss_hash_default_key_len);
591 priv->rss_conf.rss_key_len = rss_hash_default_key_len;
592 priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
593 priv->rxqs = (void *)dev->data->rx_queues;
594 priv->txqs = (void *)dev->data->tx_queues;
595 if (txqs_n != priv->txqs_n) {
596 INFO("%p: TX queues number update: %u -> %u",
597 (void *)dev, priv->txqs_n, txqs_n);
598 priv->txqs_n = txqs_n;
600 if (rxqs_n > priv->config.ind_table_max_size) {
601 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
604 if (rxqs_n == priv->rxqs_n)
606 INFO("%p: RX queues number update: %u -> %u",
607 (void *)dev, priv->rxqs_n, rxqs_n);
608 priv->rxqs_n = rxqs_n;
609 /* If the requested number of RX queues is not a power of two, use the
610 * maximum indirection table size for better balancing.
611 * The result is always rounded to the next power of two. */
612 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
613 priv->config.ind_table_max_size :
615 if (priv_rss_reta_index_resize(priv, reta_idx_n))
617 /* When the number of RX queues is not a power of two, the remaining
618 * table entries are padded with reused WQs and hashes are not spread
620 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
621 (*priv->reta_idx)[i] = j;
629 * DPDK callback for Ethernet device configuration.
632 * Pointer to Ethernet device structure.
635 * 0 on success, negative errno value on failure.
638 mlx5_dev_configure(struct rte_eth_dev *dev)
640 struct priv *priv = dev->data->dev_private;
644 ret = dev_configure(dev);
651 * DPDK callback to get information about the device.
654 * Pointer to Ethernet device structure.
656 * Info structure output buffer.
659 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
661 struct priv *priv = dev->data->dev_private;
662 struct mlx5_dev_config *config = &priv->config;
664 char ifname[IF_NAMESIZE];
666 info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
669 /* FIXME: we should ask the device for these values. */
670 info->min_rx_bufsize = 32;
671 info->max_rx_pktlen = 65536;
673 * Since we need one CQ per QP, the limit is the minimum number
674 * between the two values.
676 max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
677 priv->device_attr.orig_attr.max_qp);
678 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
681 info->max_rx_queues = max;
682 info->max_tx_queues = max;
683 info->max_mac_addrs = RTE_DIM(priv->mac);
684 info->rx_queue_offload_capa =
685 mlx5_priv_get_rx_queue_offloads(priv);
686 info->rx_offload_capa = (mlx5_priv_get_rx_port_offloads(priv) |
687 info->rx_queue_offload_capa);
688 info->tx_offload_capa = mlx5_priv_get_tx_port_offloads(priv);
689 if (priv_get_ifname(priv, &ifname) == 0)
690 info->if_index = if_nametoindex(ifname);
691 info->reta_size = priv->reta_idx_n ?
692 priv->reta_idx_n : config->ind_table_max_size;
693 info->hash_key_size = priv->rss_conf.rss_key_len;
694 info->speed_capa = priv->link_speed_capa;
695 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
700 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
702 static const uint32_t ptypes[] = {
703 /* refers to rxq_cq_to_pkt_type() */
705 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
706 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
707 RTE_PTYPE_L4_NONFRAG,
711 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
712 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
713 RTE_PTYPE_INNER_L4_NONFRAG,
714 RTE_PTYPE_INNER_L4_FRAG,
715 RTE_PTYPE_INNER_L4_TCP,
716 RTE_PTYPE_INNER_L4_UDP,
720 if (dev->rx_pkt_burst == mlx5_rx_burst ||
721 dev->rx_pkt_burst == mlx5_rx_burst_vec)
727 * DPDK callback to retrieve physical link information.
730 * Pointer to Ethernet device structure.
731 * @param wait_to_complete
732 * Wait for request completion (ignored).
735 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
737 struct priv *priv = dev->data->dev_private;
738 struct ethtool_cmd edata = {
739 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
742 struct rte_eth_link dev_link;
745 /* priv_lock() is not taken to allow concurrent calls. */
747 (void)wait_to_complete;
748 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
749 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
752 memset(&dev_link, 0, sizeof(dev_link));
753 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
754 (ifr.ifr_flags & IFF_RUNNING));
755 ifr.ifr_data = (void *)&edata;
756 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
757 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
761 link_speed = ethtool_cmd_speed(&edata);
762 if (link_speed == -1)
763 dev_link.link_speed = 0;
765 dev_link.link_speed = link_speed;
766 priv->link_speed_capa = 0;
767 if (edata.supported & SUPPORTED_Autoneg)
768 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
769 if (edata.supported & (SUPPORTED_1000baseT_Full |
770 SUPPORTED_1000baseKX_Full))
771 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
772 if (edata.supported & SUPPORTED_10000baseKR_Full)
773 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
774 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
775 SUPPORTED_40000baseCR4_Full |
776 SUPPORTED_40000baseSR4_Full |
777 SUPPORTED_40000baseLR4_Full))
778 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
779 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
780 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
781 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
782 ETH_LINK_SPEED_FIXED);
783 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
784 /* Link status changed. */
785 dev->data->dev_link = dev_link;
788 /* Link status is still the same. */
793 * Retrieve physical link information (unlocked version using new ioctl).
796 * Pointer to Ethernet device structure.
797 * @param wait_to_complete
798 * Wait for request completion (ignored).
801 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
803 struct priv *priv = dev->data->dev_private;
804 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
806 struct rte_eth_link dev_link;
809 (void)wait_to_complete;
810 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
811 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
814 memset(&dev_link, 0, sizeof(dev_link));
815 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
816 (ifr.ifr_flags & IFF_RUNNING));
817 ifr.ifr_data = (void *)&gcmd;
818 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
819 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
823 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
825 alignas(struct ethtool_link_settings)
826 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
827 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
828 struct ethtool_link_settings *ecmd = (void *)data;
831 ifr.ifr_data = (void *)ecmd;
832 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
833 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
837 dev_link.link_speed = ecmd->speed;
838 sc = ecmd->link_mode_masks[0] |
839 ((uint64_t)ecmd->link_mode_masks[1] << 32);
840 priv->link_speed_capa = 0;
841 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
842 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
843 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
844 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
845 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
846 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
847 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
848 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
849 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
850 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
851 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
852 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
853 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
854 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
855 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
856 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
857 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
858 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
859 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
860 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
861 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
862 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
863 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
864 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
865 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
866 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
867 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
868 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
869 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
870 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
871 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
872 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
873 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
874 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
875 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
876 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
877 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
878 ETH_LINK_SPEED_FIXED);
879 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
880 /* Link status changed. */
881 dev->data->dev_link = dev_link;
884 /* Link status is still the same. */
889 * Enable receiving and transmitting traffic.
892 * Pointer to private structure.
895 priv_link_start(struct priv *priv)
897 struct rte_eth_dev *dev = priv->dev;
900 dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
901 dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
902 err = priv_dev_traffic_enable(priv, dev);
904 ERROR("%p: error occurred while configuring control flows: %s",
905 (void *)priv, strerror(err));
906 err = priv_flow_start(priv, &priv->flows);
908 ERROR("%p: error occurred while configuring flows: %s",
909 (void *)priv, strerror(err));
913 * Disable receiving and transmitting traffic.
916 * Pointer to private structure.
919 priv_link_stop(struct priv *priv)
921 struct rte_eth_dev *dev = priv->dev;
923 priv_flow_stop(priv, &priv->flows);
924 priv_dev_traffic_disable(priv, dev);
925 dev->rx_pkt_burst = removed_rx_burst;
926 dev->tx_pkt_burst = removed_tx_burst;
930 * Retrieve physical link information and update rx/tx_pkt_burst callbacks
934 * Pointer to private structure.
935 * @param wait_to_complete
936 * Wait for request completion (ignored).
939 priv_link_update(struct priv *priv, int wait_to_complete)
941 struct rte_eth_dev *dev = priv->dev;
942 struct utsname utsname;
945 struct rte_eth_link dev_link = dev->data->dev_link;
947 if (uname(&utsname) == -1 ||
948 sscanf(utsname.release, "%d.%d.%d",
949 &ver[0], &ver[1], &ver[2]) != 3 ||
950 KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
951 ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
953 ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
954 /* If lsc interrupt is disabled, should always be ready for traffic. */
955 if (!dev->data->dev_conf.intr_conf.lsc) {
956 priv_link_start(priv);
959 /* Re-select burst callbacks only if link status has been changed. */
960 if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
961 if (dev->data->dev_link.link_status == ETH_LINK_UP)
962 priv_link_start(priv);
964 priv_link_stop(priv);
970 * Querying the link status till it changes to the desired state.
971 * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
974 * Pointer to private structure.
976 * Link desired status.
979 * 0 on success, negative errno value on failure.
982 priv_force_link_status_change(struct priv *priv, int status)
986 while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
987 priv_link_update(priv, 0);
988 if (priv->dev->data->dev_link.link_status == status)
997 * DPDK callback to retrieve physical link information.
1000 * Pointer to Ethernet device structure.
1001 * @param wait_to_complete
1002 * Wait for request completion (ignored).
1005 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1007 struct priv *priv = dev->data->dev_private;
1011 ret = priv_link_update(priv, wait_to_complete);
1017 * DPDK callback to change the MTU.
1020 * Pointer to Ethernet device structure.
1025 * 0 on success, negative errno value on failure.
1028 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1030 struct priv *priv = dev->data->dev_private;
1035 ret = priv_get_mtu(priv, &kern_mtu);
1038 /* Set kernel interface MTU first. */
1039 ret = priv_set_mtu(priv, mtu);
1042 ret = priv_get_mtu(priv, &kern_mtu);
1045 if (kern_mtu == mtu) {
1047 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
1053 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
1061 * DPDK callback to get flow control status.
1064 * Pointer to Ethernet device structure.
1065 * @param[out] fc_conf
1066 * Flow control output buffer.
1069 * 0 on success, negative errno value on failure.
1072 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1074 struct priv *priv = dev->data->dev_private;
1076 struct ethtool_pauseparam ethpause = {
1077 .cmd = ETHTOOL_GPAUSEPARAM
1081 ifr.ifr_data = (void *)ðpause;
1083 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1085 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1091 fc_conf->autoneg = ethpause.autoneg;
1092 if (ethpause.rx_pause && ethpause.tx_pause)
1093 fc_conf->mode = RTE_FC_FULL;
1094 else if (ethpause.rx_pause)
1095 fc_conf->mode = RTE_FC_RX_PAUSE;
1096 else if (ethpause.tx_pause)
1097 fc_conf->mode = RTE_FC_TX_PAUSE;
1099 fc_conf->mode = RTE_FC_NONE;
1109 * DPDK callback to modify flow control parameters.
1112 * Pointer to Ethernet device structure.
1113 * @param[in] fc_conf
1114 * Flow control parameters.
1117 * 0 on success, negative errno value on failure.
1120 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1122 struct priv *priv = dev->data->dev_private;
1124 struct ethtool_pauseparam ethpause = {
1125 .cmd = ETHTOOL_SPAUSEPARAM
1129 ifr.ifr_data = (void *)ðpause;
1130 ethpause.autoneg = fc_conf->autoneg;
1131 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1132 (fc_conf->mode & RTE_FC_RX_PAUSE))
1133 ethpause.rx_pause = 1;
1135 ethpause.rx_pause = 0;
1137 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1138 (fc_conf->mode & RTE_FC_TX_PAUSE))
1139 ethpause.tx_pause = 1;
1141 ethpause.tx_pause = 0;
1144 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1146 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1160 * Get PCI information from struct ibv_device.
1163 * Pointer to Ethernet device structure.
1164 * @param[out] pci_addr
1165 * PCI bus address output buffer.
1168 * 0 on success, -1 on failure and errno is set.
1171 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1172 struct rte_pci_addr *pci_addr)
1176 MKSTR(path, "%s/device/uevent", device->ibdev_path);
1178 file = fopen(path, "rb");
1181 while (fgets(line, sizeof(line), file) == line) {
1182 size_t len = strlen(line);
1185 /* Truncate long lines. */
1186 if (len == (sizeof(line) - 1))
1187 while (line[(len - 1)] != '\n') {
1191 line[(len - 1)] = ret;
1193 /* Extract information. */
1196 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1200 &pci_addr->function) == 4) {
1210 * Update the link status.
1213 * Pointer to private structure.
1216 * Zero if the callback process can be called immediately.
1219 priv_link_status_update(struct priv *priv)
1221 struct rte_eth_link *link = &priv->dev->data->dev_link;
1223 priv_link_update(priv, 0);
1224 if (((link->link_speed == 0) && link->link_status) ||
1225 ((link->link_speed != 0) && !link->link_status)) {
1227 * Inconsistent status. Event likely occurred before the
1228 * kernel netdevice exposes the new status.
1230 if (!priv->pending_alarm) {
1231 priv->pending_alarm = 1;
1232 rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1233 mlx5_dev_link_status_handler,
1237 } else if (unlikely(priv->pending_alarm)) {
1238 /* Link interrupt occurred while alarm is already scheduled. */
1239 priv->pending_alarm = 0;
1240 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
1246 * Device status handler.
1249 * Pointer to private structure.
1251 * Pointer to event flags holder.
1254 * Events bitmap of callback process which can be called immediately.
1257 priv_dev_status_handler(struct priv *priv)
1259 struct ibv_async_event event;
1262 /* Read all message and acknowledge them. */
1264 if (ibv_get_async_event(priv->ctx, &event))
1266 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1267 event.event_type == IBV_EVENT_PORT_ERR) &&
1268 (priv->dev->data->dev_conf.intr_conf.lsc == 1))
1269 ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1270 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1271 priv->dev->data->dev_conf.intr_conf.rmv == 1)
1272 ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1274 DEBUG("event type %d on port %d not handled",
1275 event.event_type, event.element.port_num);
1276 ibv_ack_async_event(&event);
1278 if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1279 if (priv_link_status_update(priv))
1280 ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1285 * Handle delayed link status event.
1288 * Registered argument.
1291 mlx5_dev_link_status_handler(void *arg)
1293 struct rte_eth_dev *dev = arg;
1294 struct priv *priv = dev->data->dev_private;
1297 while (!priv_trylock(priv)) {
1298 /* Alarm is being canceled. */
1299 if (priv->pending_alarm == 0)
1303 priv->pending_alarm = 0;
1304 ret = priv_link_status_update(priv);
1307 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1311 * Handle interrupts from the NIC.
1313 * @param[in] intr_handle
1314 * Interrupt handler.
1316 * Callback argument.
1319 mlx5_dev_interrupt_handler(void *cb_arg)
1321 struct rte_eth_dev *dev = cb_arg;
1322 struct priv *priv = dev->data->dev_private;
1326 events = priv_dev_status_handler(priv);
1328 if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1329 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1330 if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1331 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1335 * Handle interrupts from the socket.
1338 * Callback argument.
1341 mlx5_dev_handler_socket(void *cb_arg)
1343 struct rte_eth_dev *dev = cb_arg;
1344 struct priv *priv = dev->data->dev_private;
1347 priv_socket_handle(priv);
1352 * Uninstall interrupt handler.
1355 * Pointer to private structure.
1357 * Pointer to the rte_eth_dev structure.
1360 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1362 if (dev->data->dev_conf.intr_conf.lsc ||
1363 dev->data->dev_conf.intr_conf.rmv)
1364 rte_intr_callback_unregister(&priv->intr_handle,
1365 mlx5_dev_interrupt_handler, dev);
1366 if (priv->primary_socket)
1367 rte_intr_callback_unregister(&priv->intr_handle_socket,
1368 mlx5_dev_handler_socket, dev);
1369 if (priv->pending_alarm) {
1370 priv->pending_alarm = 0;
1371 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1373 priv->intr_handle.fd = 0;
1374 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1375 priv->intr_handle_socket.fd = 0;
1376 priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1380 * Install interrupt handler.
1383 * Pointer to private structure.
1385 * Pointer to the rte_eth_dev structure.
1388 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1392 assert(priv->ctx->async_fd > 0);
1393 flags = fcntl(priv->ctx->async_fd, F_GETFL);
1394 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1396 INFO("failed to change file descriptor async event queue");
1397 dev->data->dev_conf.intr_conf.lsc = 0;
1398 dev->data->dev_conf.intr_conf.rmv = 0;
1400 if (dev->data->dev_conf.intr_conf.lsc ||
1401 dev->data->dev_conf.intr_conf.rmv) {
1402 priv->intr_handle.fd = priv->ctx->async_fd;
1403 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1404 rte_intr_callback_register(&priv->intr_handle,
1405 mlx5_dev_interrupt_handler, dev);
1408 rc = priv_socket_init(priv);
1409 if (!rc && priv->primary_socket) {
1410 priv->intr_handle_socket.fd = priv->primary_socket;
1411 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1412 rte_intr_callback_register(&priv->intr_handle_socket,
1413 mlx5_dev_handler_socket, dev);
1418 * Change the link state (UP / DOWN).
1421 * Pointer to private data structure.
1423 * Nonzero for link up, otherwise link down.
1426 * 0 on success, errno value on failure.
1429 priv_dev_set_link(struct priv *priv, int up)
1431 return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1435 * DPDK callback to bring the link DOWN.
1438 * Pointer to Ethernet device structure.
1441 * 0 on success, errno value on failure.
1444 mlx5_set_link_down(struct rte_eth_dev *dev)
1446 struct priv *priv = dev->data->dev_private;
1450 err = priv_dev_set_link(priv, 0);
1456 * DPDK callback to bring the link UP.
1459 * Pointer to Ethernet device structure.
1462 * 0 on success, errno value on failure.
1465 mlx5_set_link_up(struct rte_eth_dev *dev)
1467 struct priv *priv = dev->data->dev_private;
1471 err = priv_dev_set_link(priv, 1);
1477 * Configure the TX function to use.
1480 * Pointer to private data structure.
1482 * Pointer to rte_eth_dev structure.
1485 * Pointer to selected Tx burst function.
1488 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1490 eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1491 struct mlx5_dev_config *config = &priv->config;
1492 uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1493 int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1494 DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1495 DEV_TX_OFFLOAD_GRE_TNL_TSO));
1496 int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1498 assert(priv != NULL);
1499 /* Select appropriate TX function. */
1500 if (vlan_insert || tso)
1501 return tx_pkt_burst;
1502 if (config->mps == MLX5_MPW_ENHANCED) {
1503 if (priv_check_vec_tx_support(priv, dev) > 0) {
1504 if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1505 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1507 tx_pkt_burst = mlx5_tx_burst_vec;
1508 DEBUG("selected Enhanced MPW TX vectorized function");
1510 tx_pkt_burst = mlx5_tx_burst_empw;
1511 DEBUG("selected Enhanced MPW TX function");
1513 } else if (config->mps && (config->txq_inline > 0)) {
1514 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1515 DEBUG("selected MPW inline TX function");
1516 } else if (config->mps) {
1517 tx_pkt_burst = mlx5_tx_burst_mpw;
1518 DEBUG("selected MPW TX function");
1520 return tx_pkt_burst;
1524 * Configure the RX function to use.
1527 * Pointer to private data structure.
1529 * Pointer to rte_eth_dev structure.
1532 * Pointer to selected Rx burst function.
1535 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1537 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1539 assert(priv != NULL);
1540 if (priv_check_vec_rx_support(priv) > 0) {
1541 rx_pkt_burst = mlx5_rx_burst_vec;
1542 DEBUG("selected RX vectorized function");
1544 return rx_pkt_burst;
1548 * Check if mlx5 device was removed.
1551 * Pointer to Ethernet device structure.
1554 * 1 when device is removed, otherwise 0.
1557 mlx5_is_removed(struct rte_eth_dev *dev)
1559 struct ibv_device_attr device_attr;
1560 struct priv *priv = dev->data->dev_private;
1562 if (ibv_query_device(priv->ctx, &device_attr) == EIO)