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.
44 #include <sys/ioctl.h>
45 #include <sys/socket.h>
46 #include <sys/utsname.h>
47 #include <netinet/in.h>
48 #include <linux/ethtool.h>
49 #include <linux/sockios.h>
50 #include <linux/version.h>
53 /* DPDK headers don't like -pedantic. */
55 #pragma GCC diagnostic ignored "-Wpedantic"
57 #include <rte_atomic.h>
58 #include <rte_ethdev.h>
60 #include <rte_common.h>
61 #include <rte_interrupts.h>
62 #include <rte_alarm.h>
63 #include <rte_malloc.h>
65 #pragma GCC diagnostic error "-Wpedantic"
69 #include "mlx5_rxtx.h"
70 #include "mlx5_utils.h"
72 /* Add defines in case the running kernel is not the same as user headers. */
73 #ifndef ETHTOOL_GLINKSETTINGS
74 struct ethtool_link_settings {
83 uint8_t eth_tp_mdix_ctrl;
84 int8_t link_mode_masks_nwords;
86 uint32_t link_mode_masks[];
89 #define ETHTOOL_GLINKSETTINGS 0x0000004c
90 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
91 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
92 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
93 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
94 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
95 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
96 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
97 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
98 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
99 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
100 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
101 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
102 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
103 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
104 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
105 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
107 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
108 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
109 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
110 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
112 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
113 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
114 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
116 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
117 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
118 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
119 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
120 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
124 * Return private structure associated with an Ethernet device.
127 * Pointer to Ethernet device structure.
130 * Pointer to private structure.
133 mlx5_get_priv(struct rte_eth_dev *dev)
135 struct mlx5_secondary_data *sd;
137 if (!mlx5_is_secondary())
138 return dev->data->dev_private;
139 sd = &mlx5_secondary_data[dev->data->port_id];
140 return sd->data.dev_private;
144 * Check if running as a secondary process.
147 * Nonzero if running as a secondary process.
150 mlx5_is_secondary(void)
152 return rte_eal_process_type() != RTE_PROC_PRIMARY;
156 * Get interface name from private structure.
159 * Pointer to private structure.
161 * Interface name output buffer.
164 * 0 on success, -1 on failure and errno is set.
167 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
171 unsigned int dev_type = 0;
172 unsigned int dev_port_prev = ~0u;
173 char match[IF_NAMESIZE] = "";
176 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
182 while ((dent = readdir(dir)) != NULL) {
183 char *name = dent->d_name;
185 unsigned int dev_port;
188 if ((name[0] == '.') &&
189 ((name[1] == '\0') ||
190 ((name[1] == '.') && (name[2] == '\0'))))
193 MKSTR(path, "%s/device/net/%s/%s",
194 priv->ctx->device->ibdev_path, name,
195 (dev_type ? "dev_id" : "dev_port"));
197 file = fopen(path, "rb");
202 * Switch to dev_id when dev_port does not exist as
203 * is the case with Linux kernel versions < 3.15.
214 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
219 * Switch to dev_id when dev_port returns the same value for
220 * all ports. May happen when using a MOFED release older than
221 * 3.0 with a Linux kernel >= 3.15.
223 if (dev_port == dev_port_prev)
225 dev_port_prev = dev_port;
226 if (dev_port == (priv->port - 1u))
227 snprintf(match, sizeof(match), "%s", name);
230 if (match[0] == '\0')
232 strncpy(*ifname, match, sizeof(*ifname));
237 * Check if the counter is located on ib counters file.
243 * 1 if counter is located on ib counters file , 0 otherwise.
246 priv_is_ib_cntr(const char *cntr)
248 if (!strcmp(cntr, "out_of_buffer"))
254 * Read from sysfs entry.
257 * Pointer to private structure.
259 * Entry name relative to sysfs path.
261 * Data output buffer.
266 * 0 on success, -1 on failure and errno is set.
269 priv_sysfs_read(const struct priv *priv, const char *entry,
270 char *buf, size_t size)
272 char ifname[IF_NAMESIZE];
277 if (priv_get_ifname(priv, &ifname))
280 if (priv_is_ib_cntr(entry)) {
281 MKSTR(path, "%s/ports/1/hw_counters/%s",
282 priv->ctx->device->ibdev_path, entry);
283 file = fopen(path, "rb");
285 MKSTR(path, "%s/device/net/%s/%s",
286 priv->ctx->device->ibdev_path, ifname, entry);
287 file = fopen(path, "rb");
291 ret = fread(buf, 1, size, file);
293 if (((size_t)ret < size) && (ferror(file)))
303 * Write to sysfs entry.
306 * Pointer to private structure.
308 * Entry name relative to sysfs path.
315 * 0 on success, -1 on failure and errno is set.
318 priv_sysfs_write(const struct priv *priv, const char *entry,
319 char *buf, size_t size)
321 char ifname[IF_NAMESIZE];
326 if (priv_get_ifname(priv, &ifname))
329 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
332 file = fopen(path, "wb");
335 ret = fwrite(buf, 1, size, file);
337 if (((size_t)ret < size) || (ferror(file)))
347 * Get unsigned long sysfs property.
350 * Pointer to private structure.
352 * Entry name relative to sysfs path.
354 * Value output buffer.
357 * 0 on success, -1 on failure and errno is set.
360 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
363 unsigned long value_ret;
366 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
368 DEBUG("cannot read %s value from sysfs: %s",
369 name, strerror(errno));
372 value_str[ret] = '\0';
374 value_ret = strtoul(value_str, NULL, 0);
376 DEBUG("invalid %s value `%s': %s", name, value_str,
385 * Set unsigned long sysfs property.
388 * Pointer to private structure.
390 * Entry name relative to sysfs path.
395 * 0 on success, -1 on failure and errno is set.
398 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
401 MKSTR(value_str, "%lu", value);
403 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
405 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
406 name, value_str, value, strerror(errno));
413 * Perform ifreq ioctl() on associated Ethernet device.
416 * Pointer to private structure.
418 * Request number to pass to ioctl().
420 * Interface request structure output buffer.
423 * 0 on success, -1 on failure and errno is set.
426 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
428 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
433 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
434 ret = ioctl(sock, req, ifr);
440 * Return the number of active VFs for the current device.
443 * Pointer to private structure.
444 * @param[out] num_vfs
445 * Number of active VFs.
448 * 0 on success, -1 on failure and errno is set.
451 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
453 /* The sysfs entry name depends on the operating system. */
454 const char **name = (const char *[]){
455 "device/sriov_numvfs",
456 "device/mlx5_num_vfs",
462 unsigned long ulong_num_vfs;
464 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
466 *num_vfs = ulong_num_vfs;
467 } while (*(++name) && ret);
475 * Pointer to private structure.
477 * MTU value output buffer.
480 * 0 on success, -1 on failure and errno is set.
483 priv_get_mtu(struct priv *priv, uint16_t *mtu)
485 unsigned long ulong_mtu;
487 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
494 * Read device counter from sysfs.
497 * Pointer to private structure.
501 * Counter output buffer.
504 * 0 on success, -1 on failure and errno is set.
507 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr)
509 unsigned long ulong_ctr;
511 if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1)
521 * Pointer to private structure.
526 * 0 on success, -1 on failure and errno is set.
529 priv_set_mtu(struct priv *priv, uint16_t mtu)
533 if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
534 priv_get_mtu(priv, &new_mtu))
546 * Pointer to private structure.
548 * Bitmask for flags that must remain untouched.
550 * Bitmask for flags to modify.
553 * 0 on success, -1 on failure and errno is set.
556 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
560 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
563 tmp |= (flags & (~keep));
564 return priv_set_sysfs_ulong(priv, "flags", tmp);
568 * Ethernet device configuration.
570 * Prepare the driver for a given number of TX and RX queues.
573 * Pointer to Ethernet device structure.
576 * 0 on success, errno value on failure.
579 dev_configure(struct rte_eth_dev *dev)
581 struct priv *priv = dev->data->dev_private;
582 unsigned int rxqs_n = dev->data->nb_rx_queues;
583 unsigned int txqs_n = dev->data->nb_tx_queues;
586 unsigned int reta_idx_n;
588 priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
589 priv->rxqs = (void *)dev->data->rx_queues;
590 priv->txqs = (void *)dev->data->tx_queues;
591 if (txqs_n != priv->txqs_n) {
592 INFO("%p: TX queues number update: %u -> %u",
593 (void *)dev, priv->txqs_n, txqs_n);
594 priv->txqs_n = txqs_n;
596 if (rxqs_n > priv->ind_table_max_size) {
597 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
600 if (rxqs_n == priv->rxqs_n)
602 INFO("%p: RX queues number update: %u -> %u",
603 (void *)dev, priv->rxqs_n, rxqs_n);
604 priv->rxqs_n = rxqs_n;
605 /* If the requested number of RX queues is not a power of two, use the
606 * maximum indirection table size for better balancing.
607 * The result is always rounded to the next power of two. */
608 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
609 priv->ind_table_max_size :
611 if (priv_rss_reta_index_resize(priv, reta_idx_n))
613 /* When the number of RX queues is not a power of two, the remaining
614 * table entries are padded with reused WQs and hashes are not spread
616 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
617 (*priv->reta_idx)[i] = j;
625 * DPDK callback for Ethernet device configuration.
628 * Pointer to Ethernet device structure.
631 * 0 on success, negative errno value on failure.
634 mlx5_dev_configure(struct rte_eth_dev *dev)
636 struct priv *priv = dev->data->dev_private;
639 if (mlx5_is_secondary())
640 return -E_RTE_SECONDARY;
643 ret = dev_configure(dev);
650 * DPDK callback to get information about the device.
653 * Pointer to Ethernet device structure.
655 * Info structure output buffer.
658 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
660 struct priv *priv = mlx5_get_priv(dev);
662 char ifname[IF_NAMESIZE];
664 info->pci_dev = RTE_DEV_TO_PCI(dev->device);
667 /* FIXME: we should ask the device for these values. */
668 info->min_rx_bufsize = 32;
669 info->max_rx_pktlen = 65536;
671 * Since we need one CQ per QP, the limit is the minimum number
672 * between the two values.
674 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
675 priv->device_attr.max_qp : priv->device_attr.max_cq);
676 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
679 info->max_rx_queues = max;
680 info->max_tx_queues = max;
681 info->max_mac_addrs = RTE_DIM(priv->mac);
682 info->rx_offload_capa =
684 (DEV_RX_OFFLOAD_IPV4_CKSUM |
685 DEV_RX_OFFLOAD_UDP_CKSUM |
686 DEV_RX_OFFLOAD_TCP_CKSUM) :
688 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0);
690 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
692 info->tx_offload_capa |=
693 (DEV_TX_OFFLOAD_IPV4_CKSUM |
694 DEV_TX_OFFLOAD_UDP_CKSUM |
695 DEV_TX_OFFLOAD_TCP_CKSUM);
697 info->tx_offload_capa |= DEV_TX_OFFLOAD_TCP_TSO;
699 info->tx_offload_capa |= (DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
700 DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
701 DEV_TX_OFFLOAD_GRE_TNL_TSO);
702 if (priv_get_ifname(priv, &ifname) == 0)
703 info->if_index = if_nametoindex(ifname);
704 info->reta_size = priv->reta_idx_n ?
705 priv->reta_idx_n : priv->ind_table_max_size;
706 info->hash_key_size = ((*priv->rss_conf) ?
707 (*priv->rss_conf)[0]->rss_key_len :
709 info->speed_capa = priv->link_speed_capa;
714 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
716 static const uint32_t ptypes[] = {
717 /* refers to rxq_cq_to_pkt_type() */
718 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
719 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
720 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
721 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
726 if (dev->rx_pkt_burst == mlx5_rx_burst)
732 * DPDK callback to retrieve physical link information.
735 * Pointer to Ethernet device structure.
736 * @param wait_to_complete
737 * Wait for request completion (ignored).
740 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
742 struct priv *priv = mlx5_get_priv(dev);
743 struct ethtool_cmd edata = {
744 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
747 struct rte_eth_link dev_link;
750 /* priv_lock() is not taken to allow concurrent calls. */
752 (void)wait_to_complete;
753 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
754 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
757 memset(&dev_link, 0, sizeof(dev_link));
758 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
759 (ifr.ifr_flags & IFF_RUNNING));
760 ifr.ifr_data = (void *)&edata;
761 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
762 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
766 link_speed = ethtool_cmd_speed(&edata);
767 if (link_speed == -1)
768 dev_link.link_speed = 0;
770 dev_link.link_speed = link_speed;
771 priv->link_speed_capa = 0;
772 if (edata.supported & SUPPORTED_Autoneg)
773 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
774 if (edata.supported & (SUPPORTED_1000baseT_Full |
775 SUPPORTED_1000baseKX_Full))
776 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
777 if (edata.supported & SUPPORTED_10000baseKR_Full)
778 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
779 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
780 SUPPORTED_40000baseCR4_Full |
781 SUPPORTED_40000baseSR4_Full |
782 SUPPORTED_40000baseLR4_Full))
783 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
784 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
785 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
786 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
787 ETH_LINK_SPEED_FIXED);
788 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
789 /* Link status changed. */
790 dev->data->dev_link = dev_link;
793 /* Link status is still the same. */
798 * Retrieve physical link information (unlocked version using new ioctl).
801 * Pointer to Ethernet device structure.
802 * @param wait_to_complete
803 * Wait for request completion (ignored).
806 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
808 struct priv *priv = mlx5_get_priv(dev);
809 struct ethtool_link_settings edata = {
810 .cmd = ETHTOOL_GLINKSETTINGS,
813 struct rte_eth_link dev_link;
816 (void)wait_to_complete;
817 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
818 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
821 memset(&dev_link, 0, sizeof(dev_link));
822 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
823 (ifr.ifr_flags & IFF_RUNNING));
824 ifr.ifr_data = (void *)&edata;
825 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
826 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
830 dev_link.link_speed = edata.speed;
831 sc = edata.link_mode_masks[0] |
832 ((uint64_t)edata.link_mode_masks[1] << 32);
833 priv->link_speed_capa = 0;
834 if (sc & ETHTOOL_LINK_MODE_Autoneg_BIT)
835 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
836 if (sc & (ETHTOOL_LINK_MODE_1000baseT_Full_BIT |
837 ETHTOOL_LINK_MODE_1000baseKX_Full_BIT))
838 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
839 if (sc & (ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT |
840 ETHTOOL_LINK_MODE_10000baseKR_Full_BIT |
841 ETHTOOL_LINK_MODE_10000baseR_FEC_BIT))
842 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
843 if (sc & (ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT |
844 ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT))
845 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
846 if (sc & (ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT |
847 ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT |
848 ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT |
849 ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT))
850 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
851 if (sc & (ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT |
852 ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT |
853 ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT |
854 ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT))
855 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
856 if (sc & (ETHTOOL_LINK_MODE_25000baseCR_Full_BIT |
857 ETHTOOL_LINK_MODE_25000baseKR_Full_BIT |
858 ETHTOOL_LINK_MODE_25000baseSR_Full_BIT))
859 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
860 if (sc & (ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT |
861 ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT))
862 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
863 if (sc & (ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT |
864 ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT |
865 ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT |
866 ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT))
867 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
868 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
869 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
870 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
871 ETH_LINK_SPEED_FIXED);
872 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
873 /* Link status changed. */
874 dev->data->dev_link = dev_link;
877 /* Link status is still the same. */
882 * DPDK callback to retrieve physical link information.
885 * Pointer to Ethernet device structure.
886 * @param wait_to_complete
887 * Wait for request completion (ignored).
890 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
892 struct utsname utsname;
895 if (uname(&utsname) == -1 ||
896 sscanf(utsname.release, "%d.%d.%d",
897 &ver[0], &ver[1], &ver[2]) != 3 ||
898 KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
899 return mlx5_link_update_unlocked_gset(dev, wait_to_complete);
900 return mlx5_link_update_unlocked_gs(dev, wait_to_complete);
904 * DPDK callback to change the MTU.
906 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
907 * received). Use this as a hint to enable/disable scattered packets support
908 * and improve performance when not needed.
909 * Since failure is not an option, reconfiguring queues on the fly is not
913 * Pointer to Ethernet device structure.
918 * 0 on success, negative errno value on failure.
921 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
923 struct priv *priv = dev->data->dev_private;
926 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
928 unsigned int max_frame_len;
930 int restart = priv->started;
932 if (mlx5_is_secondary())
933 return -E_RTE_SECONDARY;
936 /* Set kernel interface MTU first. */
937 if (priv_set_mtu(priv, mtu)) {
939 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
943 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
944 /* Temporarily replace RX handler with a fake one, assuming it has not
945 * been copied elsewhere. */
946 dev->rx_pkt_burst = removed_rx_burst;
947 /* Make sure everyone has left mlx5_rx_burst() and uses
948 * removed_rx_burst() instead. */
951 /* MTU does not include header and CRC. */
952 max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN;
953 /* Check if at least one queue is going to need a SGE update. */
954 for (i = 0; i != priv->rxqs_n; ++i) {
955 struct rxq *rxq = (*priv->rxqs)[i];
957 unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len;
962 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
963 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
965 * Determine the number of SGEs needed for a full packet
966 * and round it to the next power of two.
968 sges_n = log2above((size / mb_len) + !!(size % mb_len));
969 if (sges_n != rxq->sges_n)
973 * If all queues have the right number of SGEs, a simple rehash
974 * of their buffers is enough, otherwise SGE information can only
975 * be updated in a queue by recreating it. All resources that depend
976 * on queues (flows, indirection tables) must be recreated as well in
979 rehash = (i == priv->rxqs_n);
981 /* Clean up everything as with mlx5_dev_stop(). */
982 priv_special_flow_disable_all(priv);
983 priv_mac_addrs_disable(priv);
984 priv_destroy_hash_rxqs(priv);
985 priv_fdir_disable(priv);
986 priv_dev_interrupt_handler_uninstall(priv, dev);
989 /* Reconfigure each RX queue. */
990 for (i = 0; (i != priv->rxqs_n); ++i) {
991 struct rxq *rxq = (*priv->rxqs)[i];
992 struct rxq_ctrl *rxq_ctrl =
993 container_of(rxq, struct rxq_ctrl, rxq);
999 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
1000 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
1001 /* Provide new values to rxq_setup(). */
1002 dev->data->dev_conf.rxmode.jumbo_frame =
1003 (max_frame_len > ETHER_MAX_LEN);
1004 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
1006 ret = rxq_rehash(dev, rxq_ctrl);
1008 ret = rxq_ctrl_setup(dev, rxq_ctrl, 1 << rxq->elts_n,
1009 rxq_ctrl->socket, NULL, rxq->mp);
1012 /* Attempt to roll back in case of error. */
1013 tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM;
1014 if (max_frame_len != tmp) {
1015 max_frame_len = tmp;
1018 /* Double fault, disable RX. */
1022 * Use a safe RX burst function in case of error, otherwise mimic
1026 ERROR("unable to reconfigure RX queues, RX disabled");
1027 rx_func = removed_rx_burst;
1028 } else if (restart &&
1030 !priv_create_hash_rxqs(priv) &&
1031 !priv_rehash_flows(priv)) {
1032 if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE)
1033 priv_fdir_enable(priv);
1034 priv_dev_interrupt_handler_install(priv, dev);
1037 /* Burst functions can now be called again. */
1039 dev->rx_pkt_burst = rx_func;
1047 * DPDK callback to get flow control status.
1050 * Pointer to Ethernet device structure.
1051 * @param[out] fc_conf
1052 * Flow control output buffer.
1055 * 0 on success, negative errno value on failure.
1058 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1060 struct priv *priv = dev->data->dev_private;
1062 struct ethtool_pauseparam ethpause = {
1063 .cmd = ETHTOOL_GPAUSEPARAM
1067 if (mlx5_is_secondary())
1068 return -E_RTE_SECONDARY;
1070 ifr.ifr_data = (void *)ðpause;
1072 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1074 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1080 fc_conf->autoneg = ethpause.autoneg;
1081 if (ethpause.rx_pause && ethpause.tx_pause)
1082 fc_conf->mode = RTE_FC_FULL;
1083 else if (ethpause.rx_pause)
1084 fc_conf->mode = RTE_FC_RX_PAUSE;
1085 else if (ethpause.tx_pause)
1086 fc_conf->mode = RTE_FC_TX_PAUSE;
1088 fc_conf->mode = RTE_FC_NONE;
1098 * DPDK callback to modify flow control parameters.
1101 * Pointer to Ethernet device structure.
1102 * @param[in] fc_conf
1103 * Flow control parameters.
1106 * 0 on success, negative errno value on failure.
1109 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1111 struct priv *priv = dev->data->dev_private;
1113 struct ethtool_pauseparam ethpause = {
1114 .cmd = ETHTOOL_SPAUSEPARAM
1118 if (mlx5_is_secondary())
1119 return -E_RTE_SECONDARY;
1121 ifr.ifr_data = (void *)ðpause;
1122 ethpause.autoneg = fc_conf->autoneg;
1123 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1124 (fc_conf->mode & RTE_FC_RX_PAUSE))
1125 ethpause.rx_pause = 1;
1127 ethpause.rx_pause = 0;
1129 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1130 (fc_conf->mode & RTE_FC_TX_PAUSE))
1131 ethpause.tx_pause = 1;
1133 ethpause.tx_pause = 0;
1136 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1138 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1152 * Get PCI information from struct ibv_device.
1155 * Pointer to Ethernet device structure.
1156 * @param[out] pci_addr
1157 * PCI bus address output buffer.
1160 * 0 on success, -1 on failure and errno is set.
1163 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1164 struct rte_pci_addr *pci_addr)
1168 MKSTR(path, "%s/device/uevent", device->ibdev_path);
1170 file = fopen(path, "rb");
1173 while (fgets(line, sizeof(line), file) == line) {
1174 size_t len = strlen(line);
1177 /* Truncate long lines. */
1178 if (len == (sizeof(line) - 1))
1179 while (line[(len - 1)] != '\n') {
1183 line[(len - 1)] = ret;
1185 /* Extract information. */
1188 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1192 &pci_addr->function) == 4) {
1202 * Link status handler.
1205 * Pointer to private structure.
1207 * Pointer to the rte_eth_dev structure.
1210 * Nonzero if the callback process can be called immediately.
1213 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
1215 struct ibv_async_event event;
1216 struct rte_eth_link *link = &dev->data->dev_link;
1219 /* Read all message and acknowledge them. */
1221 if (ibv_get_async_event(priv->ctx, &event))
1224 if (event.event_type != IBV_EVENT_PORT_ACTIVE &&
1225 event.event_type != IBV_EVENT_PORT_ERR)
1226 DEBUG("event type %d on port %d not handled",
1227 event.event_type, event.element.port_num);
1228 ibv_ack_async_event(&event);
1230 mlx5_link_update(dev, 0);
1231 if (((link->link_speed == 0) && link->link_status) ||
1232 ((link->link_speed != 0) && !link->link_status)) {
1233 if (!priv->pending_alarm) {
1234 /* Inconsistent status, check again later. */
1235 priv->pending_alarm = 1;
1236 rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1237 mlx5_dev_link_status_handler,
1247 * Handle delayed link status event.
1250 * Registered argument.
1253 mlx5_dev_link_status_handler(void *arg)
1255 struct rte_eth_dev *dev = arg;
1256 struct priv *priv = dev->data->dev_private;
1260 assert(priv->pending_alarm == 1);
1261 priv->pending_alarm = 0;
1262 ret = priv_dev_link_status_handler(priv, dev);
1265 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1269 * Handle interrupts from the NIC.
1271 * @param[in] intr_handle
1272 * Interrupt handler.
1274 * Callback argument.
1277 mlx5_dev_interrupt_handler(void *cb_arg)
1279 struct rte_eth_dev *dev = cb_arg;
1280 struct priv *priv = dev->data->dev_private;
1284 ret = priv_dev_link_status_handler(priv, dev);
1287 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1291 * Uninstall interrupt handler.
1294 * Pointer to private structure.
1296 * Pointer to the rte_eth_dev structure.
1299 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1301 if (!dev->data->dev_conf.intr_conf.lsc)
1303 rte_intr_callback_unregister(&priv->intr_handle,
1304 mlx5_dev_interrupt_handler,
1306 if (priv->pending_alarm)
1307 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1308 priv->pending_alarm = 0;
1309 priv->intr_handle.fd = 0;
1310 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1314 * Install interrupt handler.
1317 * Pointer to private structure.
1319 * Pointer to the rte_eth_dev structure.
1322 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1326 if (!dev->data->dev_conf.intr_conf.lsc)
1328 assert(priv->ctx->async_fd > 0);
1329 flags = fcntl(priv->ctx->async_fd, F_GETFL);
1330 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1332 INFO("failed to change file descriptor async event queue");
1333 dev->data->dev_conf.intr_conf.lsc = 0;
1335 priv->intr_handle.fd = priv->ctx->async_fd;
1336 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1337 rte_intr_callback_register(&priv->intr_handle,
1338 mlx5_dev_interrupt_handler,
1344 * Change the link state (UP / DOWN).
1347 * Pointer to Ethernet device structure.
1349 * Nonzero for link up, otherwise link down.
1352 * 0 on success, errno value on failure.
1355 priv_set_link(struct priv *priv, int up)
1357 struct rte_eth_dev *dev = priv->dev;
1361 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1364 priv_select_tx_function(priv);
1365 priv_select_rx_function(priv);
1367 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1370 dev->rx_pkt_burst = removed_rx_burst;
1371 dev->tx_pkt_burst = removed_tx_burst;
1377 * DPDK callback to bring the link DOWN.
1380 * Pointer to Ethernet device structure.
1383 * 0 on success, errno value on failure.
1386 mlx5_set_link_down(struct rte_eth_dev *dev)
1388 struct priv *priv = dev->data->dev_private;
1392 err = priv_set_link(priv, 0);
1398 * DPDK callback to bring the link UP.
1401 * Pointer to Ethernet device structure.
1404 * 0 on success, errno value on failure.
1407 mlx5_set_link_up(struct rte_eth_dev *dev)
1409 struct priv *priv = dev->data->dev_private;
1413 err = priv_set_link(priv, 1);
1419 * Configure secondary process queues from a private data pointer (primary
1420 * or secondary) and update burst callbacks. Can take place only once.
1422 * All queues must have been previously created by the primary process to
1423 * avoid undefined behavior.
1426 * Private data pointer from either primary or secondary process.
1429 * Private data pointer from secondary process, NULL in case of error.
1432 mlx5_secondary_data_setup(struct priv *priv)
1434 unsigned int port_id = 0;
1435 struct mlx5_secondary_data *sd;
1438 unsigned int nb_tx_queues;
1439 unsigned int nb_rx_queues;
1442 /* priv must be valid at this point. */
1443 assert(priv != NULL);
1444 /* priv->dev must also be valid but may point to local memory from
1445 * another process, possibly with the same address and must not
1446 * be dereferenced yet. */
1447 assert(priv->dev != NULL);
1448 /* Determine port ID by finding out where priv comes from. */
1450 sd = &mlx5_secondary_data[port_id];
1451 rte_spinlock_lock(&sd->lock);
1452 /* Primary process? */
1453 if (sd->primary_priv == priv)
1455 /* Secondary process? */
1456 if (sd->data.dev_private == priv)
1458 rte_spinlock_unlock(&sd->lock);
1459 if (++port_id == RTE_DIM(mlx5_secondary_data))
1462 /* Switch to secondary private structure. If private data has already
1463 * been updated by another thread, there is nothing else to do. */
1464 priv = sd->data.dev_private;
1465 if (priv->dev->data == &sd->data)
1467 /* Sanity checks. Secondary private structure is supposed to point
1468 * to local eth_dev, itself still pointing to the shared device data
1469 * structure allocated by the primary process. */
1470 assert(sd->shared_dev_data != &sd->data);
1471 assert(sd->data.nb_tx_queues == 0);
1472 assert(sd->data.tx_queues == NULL);
1473 assert(sd->data.nb_rx_queues == 0);
1474 assert(sd->data.rx_queues == NULL);
1475 assert(priv != sd->primary_priv);
1476 assert(priv->dev->data == sd->shared_dev_data);
1477 assert(priv->txqs_n == 0);
1478 assert(priv->txqs == NULL);
1479 assert(priv->rxqs_n == 0);
1480 assert(priv->rxqs == NULL);
1481 nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1482 nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1483 /* Allocate local storage for queues. */
1484 tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1485 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1486 RTE_CACHE_LINE_SIZE);
1487 rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1488 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1489 RTE_CACHE_LINE_SIZE);
1490 if (tx_queues == NULL || rx_queues == NULL)
1492 /* Lock to prevent control operations during setup. */
1495 for (i = 0; i != nb_tx_queues; ++i) {
1496 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1497 struct txq_ctrl *primary_txq_ctrl;
1498 struct txq_ctrl *txq_ctrl;
1500 if (primary_txq == NULL)
1502 primary_txq_ctrl = container_of(primary_txq,
1503 struct txq_ctrl, txq);
1504 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) +
1505 (1 << primary_txq->elts_n) *
1506 sizeof(struct rte_mbuf *), 0,
1507 primary_txq_ctrl->socket);
1508 if (txq_ctrl != NULL) {
1509 if (txq_ctrl_setup(priv->dev,
1511 1 << primary_txq->elts_n,
1512 primary_txq_ctrl->socket,
1514 txq_ctrl->txq.stats.idx =
1515 primary_txq->stats.idx;
1516 tx_queues[i] = &txq_ctrl->txq;
1522 txq_ctrl = tx_queues[--i];
1523 txq_cleanup(txq_ctrl);
1529 for (i = 0; i != nb_rx_queues; ++i) {
1530 struct rxq_ctrl *primary_rxq =
1531 container_of((*sd->primary_priv->rxqs)[i],
1532 struct rxq_ctrl, rxq);
1534 if (primary_rxq == NULL)
1536 /* Not supported yet. */
1537 rx_queues[i] = NULL;
1539 /* Update everything. */
1540 priv->txqs = (void *)tx_queues;
1541 priv->txqs_n = nb_tx_queues;
1542 priv->rxqs = (void *)rx_queues;
1543 priv->rxqs_n = nb_rx_queues;
1544 sd->data.rx_queues = rx_queues;
1545 sd->data.tx_queues = tx_queues;
1546 sd->data.nb_rx_queues = nb_rx_queues;
1547 sd->data.nb_tx_queues = nb_tx_queues;
1548 sd->data.dev_link = sd->shared_dev_data->dev_link;
1549 sd->data.mtu = sd->shared_dev_data->mtu;
1550 memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1551 sizeof(sd->data.rx_queue_state));
1552 memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1553 sizeof(sd->data.tx_queue_state));
1554 sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1555 /* Use local data from now on. */
1557 priv->dev->data = &sd->data;
1559 priv_select_tx_function(priv);
1560 priv_select_rx_function(priv);
1563 /* More sanity checks. */
1564 assert(priv->dev->data == &sd->data);
1565 rte_spinlock_unlock(&sd->lock);
1569 rte_free(tx_queues);
1570 rte_free(rx_queues);
1571 rte_spinlock_unlock(&sd->lock);
1576 * Configure the TX function to use.
1579 * Pointer to private structure.
1582 priv_select_tx_function(struct priv *priv)
1584 priv->dev->tx_pkt_burst = mlx5_tx_burst;
1585 /* Select appropriate TX function. */
1586 if (priv->mps == MLX5_MPW_ENHANCED) {
1587 priv->dev->tx_pkt_burst =
1589 DEBUG("selected Enhanced MPW TX function");
1590 } else if (priv->mps && priv->txq_inline) {
1591 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1592 DEBUG("selected MPW inline TX function");
1593 } else if (priv->mps) {
1594 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1595 DEBUG("selected MPW TX function");
1600 * Configure the RX function to use.
1603 * Pointer to private structure.
1606 priv_select_rx_function(struct priv *priv)
1608 priv->dev->rx_pkt_burst = mlx5_rx_burst;