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 <netinet/in.h>
47 #include <linux/ethtool.h>
48 #include <linux/sockios.h>
51 /* DPDK headers don't like -pedantic. */
53 #pragma GCC diagnostic ignored "-pedantic"
55 #include <rte_atomic.h>
56 #include <rte_ethdev.h>
58 #include <rte_common.h>
59 #include <rte_interrupts.h>
60 #include <rte_alarm.h>
61 #include <rte_malloc.h>
63 #pragma GCC diagnostic error "-pedantic"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_utils.h"
71 * Return private structure associated with an Ethernet device.
74 * Pointer to Ethernet device structure.
77 * Pointer to private structure.
80 mlx5_get_priv(struct rte_eth_dev *dev)
82 struct mlx5_secondary_data *sd;
84 if (!mlx5_is_secondary())
85 return dev->data->dev_private;
86 sd = &mlx5_secondary_data[dev->data->port_id];
87 return sd->data.dev_private;
91 * Check if running as a secondary process.
94 * Nonzero if running as a secondary process.
97 mlx5_is_secondary(void)
99 return rte_eal_process_type() != RTE_PROC_PRIMARY;
103 * Get interface name from private structure.
106 * Pointer to private structure.
108 * Interface name output buffer.
111 * 0 on success, -1 on failure and errno is set.
114 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
118 unsigned int dev_type = 0;
119 unsigned int dev_port_prev = ~0u;
120 char match[IF_NAMESIZE] = "";
123 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
129 while ((dent = readdir(dir)) != NULL) {
130 char *name = dent->d_name;
132 unsigned int dev_port;
135 if ((name[0] == '.') &&
136 ((name[1] == '\0') ||
137 ((name[1] == '.') && (name[2] == '\0'))))
140 MKSTR(path, "%s/device/net/%s/%s",
141 priv->ctx->device->ibdev_path, name,
142 (dev_type ? "dev_id" : "dev_port"));
144 file = fopen(path, "rb");
149 * Switch to dev_id when dev_port does not exist as
150 * is the case with Linux kernel versions < 3.15.
161 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
166 * Switch to dev_id when dev_port returns the same value for
167 * all ports. May happen when using a MOFED release older than
168 * 3.0 with a Linux kernel >= 3.15.
170 if (dev_port == dev_port_prev)
172 dev_port_prev = dev_port;
173 if (dev_port == (priv->port - 1u))
174 snprintf(match, sizeof(match), "%s", name);
177 if (match[0] == '\0')
179 strncpy(*ifname, match, sizeof(*ifname));
184 * Read from sysfs entry.
187 * Pointer to private structure.
189 * Entry name relative to sysfs path.
191 * Data output buffer.
196 * 0 on success, -1 on failure and errno is set.
199 priv_sysfs_read(const struct priv *priv, const char *entry,
200 char *buf, size_t size)
202 char ifname[IF_NAMESIZE];
207 if (priv_get_ifname(priv, &ifname))
210 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
213 file = fopen(path, "rb");
216 ret = fread(buf, 1, size, file);
218 if (((size_t)ret < size) && (ferror(file)))
228 * Write to sysfs entry.
231 * Pointer to private structure.
233 * Entry name relative to sysfs path.
240 * 0 on success, -1 on failure and errno is set.
243 priv_sysfs_write(const struct priv *priv, const char *entry,
244 char *buf, size_t size)
246 char ifname[IF_NAMESIZE];
251 if (priv_get_ifname(priv, &ifname))
254 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
257 file = fopen(path, "wb");
260 ret = fwrite(buf, 1, size, file);
262 if (((size_t)ret < size) || (ferror(file)))
272 * Get unsigned long sysfs property.
275 * Pointer to private structure.
277 * Entry name relative to sysfs path.
279 * Value output buffer.
282 * 0 on success, -1 on failure and errno is set.
285 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
288 unsigned long value_ret;
291 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
293 DEBUG("cannot read %s value from sysfs: %s",
294 name, strerror(errno));
297 value_str[ret] = '\0';
299 value_ret = strtoul(value_str, NULL, 0);
301 DEBUG("invalid %s value `%s': %s", name, value_str,
310 * Set unsigned long sysfs property.
313 * Pointer to private structure.
315 * Entry name relative to sysfs path.
320 * 0 on success, -1 on failure and errno is set.
323 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
326 MKSTR(value_str, "%lu", value);
328 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
330 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
331 name, value_str, value, strerror(errno));
338 * Perform ifreq ioctl() on associated Ethernet device.
341 * Pointer to private structure.
343 * Request number to pass to ioctl().
345 * Interface request structure output buffer.
348 * 0 on success, -1 on failure and errno is set.
351 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
353 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
358 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
359 ret = ioctl(sock, req, ifr);
368 * Pointer to private structure.
370 * MTU value output buffer.
373 * 0 on success, -1 on failure and errno is set.
376 priv_get_mtu(struct priv *priv, uint16_t *mtu)
378 unsigned long ulong_mtu;
380 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
390 * Pointer to private structure.
395 * 0 on success, -1 on failure and errno is set.
398 priv_set_mtu(struct priv *priv, uint16_t mtu)
402 if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
403 priv_get_mtu(priv, &new_mtu))
415 * Pointer to private structure.
417 * Bitmask for flags that must remain untouched.
419 * Bitmask for flags to modify.
422 * 0 on success, -1 on failure and errno is set.
425 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
429 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
433 return priv_set_sysfs_ulong(priv, "flags", tmp);
437 * Ethernet device configuration.
439 * Prepare the driver for a given number of TX and RX queues.
442 * Pointer to Ethernet device structure.
445 * 0 on success, errno value on failure.
448 dev_configure(struct rte_eth_dev *dev)
450 struct priv *priv = dev->data->dev_private;
451 unsigned int rxqs_n = dev->data->nb_rx_queues;
452 unsigned int txqs_n = dev->data->nb_tx_queues;
455 unsigned int reta_idx_n;
457 priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
458 priv->rxqs = (void *)dev->data->rx_queues;
459 priv->txqs = (void *)dev->data->tx_queues;
460 if (txqs_n != priv->txqs_n) {
461 INFO("%p: TX queues number update: %u -> %u",
462 (void *)dev, priv->txqs_n, txqs_n);
463 priv->txqs_n = txqs_n;
465 if (rxqs_n > priv->ind_table_max_size) {
466 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
469 if (rxqs_n == priv->rxqs_n)
471 INFO("%p: RX queues number update: %u -> %u",
472 (void *)dev, priv->rxqs_n, rxqs_n);
473 priv->rxqs_n = rxqs_n;
474 /* If the requested number of RX queues is not a power of two, use the
475 * maximum indirection table size for better balancing.
476 * The result is always rounded to the next power of two. */
477 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
478 priv->ind_table_max_size :
480 if (priv_rss_reta_index_resize(priv, reta_idx_n))
482 /* When the number of RX queues is not a power of two, the remaining
483 * table entries are padded with reused WQs and hashes are not spread
485 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
486 (*priv->reta_idx)[i] = j;
494 * DPDK callback for Ethernet device configuration.
497 * Pointer to Ethernet device structure.
500 * 0 on success, negative errno value on failure.
503 mlx5_dev_configure(struct rte_eth_dev *dev)
505 struct priv *priv = dev->data->dev_private;
508 if (mlx5_is_secondary())
509 return -E_RTE_SECONDARY;
512 ret = dev_configure(dev);
519 * DPDK callback to get information about the device.
522 * Pointer to Ethernet device structure.
524 * Info structure output buffer.
527 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
529 struct priv *priv = mlx5_get_priv(dev);
531 char ifname[IF_NAMESIZE];
534 /* FIXME: we should ask the device for these values. */
535 info->min_rx_bufsize = 32;
536 info->max_rx_pktlen = 65536;
538 * Since we need one CQ per QP, the limit is the minimum number
539 * between the two values.
541 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
542 priv->device_attr.max_qp : priv->device_attr.max_cq);
543 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
546 info->max_rx_queues = max;
547 info->max_tx_queues = max;
548 info->max_mac_addrs = RTE_DIM(priv->mac);
549 info->rx_offload_capa =
551 (DEV_RX_OFFLOAD_IPV4_CKSUM |
552 DEV_RX_OFFLOAD_UDP_CKSUM |
553 DEV_RX_OFFLOAD_TCP_CKSUM) :
555 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
557 info->tx_offload_capa |=
558 (DEV_TX_OFFLOAD_IPV4_CKSUM |
559 DEV_TX_OFFLOAD_UDP_CKSUM |
560 DEV_TX_OFFLOAD_TCP_CKSUM);
561 if (priv_get_ifname(priv, &ifname) == 0)
562 info->if_index = if_nametoindex(ifname);
563 /* FIXME: RETA update/query API expects the callee to know the size of
564 * the indirection table, for this PMD the size varies depending on
565 * the number of RX queues, it becomes impossible to find the correct
566 * size if it is not fixed.
567 * The API should be updated to solve this problem. */
568 info->reta_size = priv->ind_table_max_size;
582 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
584 static const uint32_t ptypes[] = {
585 /* refers to rxq_cq_to_pkt_type() */
588 RTE_PTYPE_INNER_L3_IPV4,
589 RTE_PTYPE_INNER_L3_IPV6,
594 if (dev->rx_pkt_burst == mlx5_rx_burst ||
595 dev->rx_pkt_burst == mlx5_rx_burst_sp)
601 * DPDK callback to retrieve physical link information (unlocked version).
604 * Pointer to Ethernet device structure.
605 * @param wait_to_complete
606 * Wait for request completion (ignored).
609 mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
611 struct priv *priv = mlx5_get_priv(dev);
612 struct ethtool_cmd edata = {
616 struct rte_eth_link dev_link;
619 (void)wait_to_complete;
620 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
621 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
624 memset(&dev_link, 0, sizeof(dev_link));
625 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
626 (ifr.ifr_flags & IFF_RUNNING));
627 ifr.ifr_data = (void *)&edata;
628 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
629 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
633 link_speed = ethtool_cmd_speed(&edata);
634 if (link_speed == -1)
635 dev_link.link_speed = 0;
637 dev_link.link_speed = link_speed;
638 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
639 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
640 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
641 ETH_LINK_SPEED_FIXED);
642 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
643 /* Link status changed. */
644 dev->data->dev_link = dev_link;
647 /* Link status is still the same. */
652 * DPDK callback to retrieve physical link information.
655 * Pointer to Ethernet device structure.
656 * @param wait_to_complete
657 * Wait for request completion (ignored).
660 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
662 struct priv *priv = mlx5_get_priv(dev);
666 ret = mlx5_link_update_unlocked(dev, wait_to_complete);
672 * DPDK callback to change the MTU.
674 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
675 * received). Use this as a hint to enable/disable scattered packets support
676 * and improve performance when not needed.
677 * Since failure is not an option, reconfiguring queues on the fly is not
681 * Pointer to Ethernet device structure.
686 * 0 on success, negative errno value on failure.
689 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
691 struct priv *priv = dev->data->dev_private;
694 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
697 if (mlx5_is_secondary())
698 return -E_RTE_SECONDARY;
701 /* Set kernel interface MTU first. */
702 if (priv_set_mtu(priv, mtu)) {
704 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
708 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
710 /* Temporarily replace RX handler with a fake one, assuming it has not
711 * been copied elsewhere. */
712 dev->rx_pkt_burst = removed_rx_burst;
713 /* Make sure everyone has left mlx5_rx_burst() and uses
714 * removed_rx_burst() instead. */
717 /* Reconfigure each RX queue. */
718 for (i = 0; (i != priv->rxqs_n); ++i) {
719 struct rxq *rxq = (*priv->rxqs)[i];
721 unsigned int max_frame_len;
726 /* Calculate new maximum frame length according to MTU and
727 * toggle scattered support (sp) if necessary. */
728 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
729 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
730 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
731 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
732 sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM));
733 /* Provide new values to rxq_setup(). */
734 dev->data->dev_conf.rxmode.jumbo_frame = sp;
735 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
736 ret = rxq_rehash(dev, rxq);
738 /* Force SP RX if that queue requires it and abort. */
740 rx_func = mlx5_rx_burst_sp;
743 /* Scattered burst function takes priority. */
745 rx_func = mlx5_rx_burst_sp;
747 /* Burst functions can now be called again. */
749 dev->rx_pkt_burst = rx_func;
757 * DPDK callback to get flow control status.
760 * Pointer to Ethernet device structure.
761 * @param[out] fc_conf
762 * Flow control output buffer.
765 * 0 on success, negative errno value on failure.
768 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
770 struct priv *priv = dev->data->dev_private;
772 struct ethtool_pauseparam ethpause = {
773 .cmd = ETHTOOL_GPAUSEPARAM
777 if (mlx5_is_secondary())
778 return -E_RTE_SECONDARY;
780 ifr.ifr_data = (void *)ðpause;
782 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
784 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
790 fc_conf->autoneg = ethpause.autoneg;
791 if (ethpause.rx_pause && ethpause.tx_pause)
792 fc_conf->mode = RTE_FC_FULL;
793 else if (ethpause.rx_pause)
794 fc_conf->mode = RTE_FC_RX_PAUSE;
795 else if (ethpause.tx_pause)
796 fc_conf->mode = RTE_FC_TX_PAUSE;
798 fc_conf->mode = RTE_FC_NONE;
808 * DPDK callback to modify flow control parameters.
811 * Pointer to Ethernet device structure.
813 * Flow control parameters.
816 * 0 on success, negative errno value on failure.
819 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
821 struct priv *priv = dev->data->dev_private;
823 struct ethtool_pauseparam ethpause = {
824 .cmd = ETHTOOL_SPAUSEPARAM
828 if (mlx5_is_secondary())
829 return -E_RTE_SECONDARY;
831 ifr.ifr_data = (void *)ðpause;
832 ethpause.autoneg = fc_conf->autoneg;
833 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
834 (fc_conf->mode & RTE_FC_RX_PAUSE))
835 ethpause.rx_pause = 1;
837 ethpause.rx_pause = 0;
839 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
840 (fc_conf->mode & RTE_FC_TX_PAUSE))
841 ethpause.tx_pause = 1;
843 ethpause.tx_pause = 0;
846 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
848 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
862 * Get PCI information from struct ibv_device.
865 * Pointer to Ethernet device structure.
866 * @param[out] pci_addr
867 * PCI bus address output buffer.
870 * 0 on success, -1 on failure and errno is set.
873 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
874 struct rte_pci_addr *pci_addr)
878 MKSTR(path, "%s/device/uevent", device->ibdev_path);
880 file = fopen(path, "rb");
883 while (fgets(line, sizeof(line), file) == line) {
884 size_t len = strlen(line);
887 /* Truncate long lines. */
888 if (len == (sizeof(line) - 1))
889 while (line[(len - 1)] != '\n') {
893 line[(len - 1)] = ret;
895 /* Extract information. */
898 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
902 &pci_addr->function) == 4) {
912 * Link status handler.
915 * Pointer to private structure.
917 * Pointer to the rte_eth_dev structure.
920 * Nonzero if the callback process can be called immediately.
923 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
925 struct ibv_async_event event;
929 /* Read all message and acknowledge them. */
931 if (ibv_get_async_event(priv->ctx, &event))
934 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
935 event.event_type == IBV_EVENT_PORT_ERR)
938 DEBUG("event type %d on port %d not handled",
939 event.event_type, event.element.port_num);
940 ibv_ack_async_event(&event);
943 if (port_change ^ priv->pending_alarm) {
944 struct rte_eth_link *link = &dev->data->dev_link;
946 priv->pending_alarm = 0;
947 mlx5_link_update_unlocked(dev, 0);
948 if (((link->link_speed == 0) && link->link_status) ||
949 ((link->link_speed != 0) && !link->link_status)) {
950 /* Inconsistent status, check again later. */
951 priv->pending_alarm = 1;
952 rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
953 mlx5_dev_link_status_handler,
962 * Handle delayed link status event.
965 * Registered argument.
968 mlx5_dev_link_status_handler(void *arg)
970 struct rte_eth_dev *dev = arg;
971 struct priv *priv = dev->data->dev_private;
975 assert(priv->pending_alarm == 1);
976 ret = priv_dev_link_status_handler(priv, dev);
979 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
983 * Handle interrupts from the NIC.
985 * @param[in] intr_handle
991 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
993 struct rte_eth_dev *dev = cb_arg;
994 struct priv *priv = dev->data->dev_private;
999 ret = priv_dev_link_status_handler(priv, dev);
1002 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
1006 * Uninstall interrupt handler.
1009 * Pointer to private structure.
1011 * Pointer to the rte_eth_dev structure.
1014 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1016 if (!dev->data->dev_conf.intr_conf.lsc)
1018 rte_intr_callback_unregister(&priv->intr_handle,
1019 mlx5_dev_interrupt_handler,
1021 if (priv->pending_alarm)
1022 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1023 priv->pending_alarm = 0;
1024 priv->intr_handle.fd = 0;
1025 priv->intr_handle.type = 0;
1029 * Install interrupt handler.
1032 * Pointer to private structure.
1034 * Pointer to the rte_eth_dev structure.
1037 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1041 if (!dev->data->dev_conf.intr_conf.lsc)
1043 assert(priv->ctx->async_fd > 0);
1044 flags = fcntl(priv->ctx->async_fd, F_GETFL);
1045 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1047 INFO("failed to change file descriptor async event queue");
1048 dev->data->dev_conf.intr_conf.lsc = 0;
1050 priv->intr_handle.fd = priv->ctx->async_fd;
1051 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1052 rte_intr_callback_register(&priv->intr_handle,
1053 mlx5_dev_interrupt_handler,
1059 * Change the link state (UP / DOWN).
1062 * Pointer to Ethernet device structure.
1064 * Nonzero for link up, otherwise link down.
1067 * 0 on success, errno value on failure.
1070 priv_set_link(struct priv *priv, int up)
1072 struct rte_eth_dev *dev = priv->dev;
1077 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1080 for (i = 0; i < priv->rxqs_n; i++)
1081 if ((*priv->rxqs)[i]->sp)
1083 /* Check if an sp queue exists.
1084 * Note: Some old frames might be received.
1086 if (i == priv->rxqs_n)
1087 dev->rx_pkt_burst = mlx5_rx_burst;
1089 dev->rx_pkt_burst = mlx5_rx_burst_sp;
1090 dev->tx_pkt_burst = mlx5_tx_burst;
1092 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1095 dev->rx_pkt_burst = removed_rx_burst;
1096 dev->tx_pkt_burst = removed_tx_burst;
1102 * DPDK callback to bring the link DOWN.
1105 * Pointer to Ethernet device structure.
1108 * 0 on success, errno value on failure.
1111 mlx5_set_link_down(struct rte_eth_dev *dev)
1113 struct priv *priv = dev->data->dev_private;
1117 err = priv_set_link(priv, 0);
1123 * DPDK callback to bring the link UP.
1126 * Pointer to Ethernet device structure.
1129 * 0 on success, errno value on failure.
1132 mlx5_set_link_up(struct rte_eth_dev *dev)
1134 struct priv *priv = dev->data->dev_private;
1138 err = priv_set_link(priv, 1);
1144 * Configure secondary process queues from a private data pointer (primary
1145 * or secondary) and update burst callbacks. Can take place only once.
1147 * All queues must have been previously created by the primary process to
1148 * avoid undefined behavior.
1151 * Private data pointer from either primary or secondary process.
1154 * Private data pointer from secondary process, NULL in case of error.
1157 mlx5_secondary_data_setup(struct priv *priv)
1159 unsigned int port_id = 0;
1160 struct mlx5_secondary_data *sd;
1163 unsigned int nb_tx_queues;
1164 unsigned int nb_rx_queues;
1167 /* priv must be valid at this point. */
1168 assert(priv != NULL);
1169 /* priv->dev must also be valid but may point to local memory from
1170 * another process, possibly with the same address and must not
1171 * be dereferenced yet. */
1172 assert(priv->dev != NULL);
1173 /* Determine port ID by finding out where priv comes from. */
1175 sd = &mlx5_secondary_data[port_id];
1176 rte_spinlock_lock(&sd->lock);
1177 /* Primary process? */
1178 if (sd->primary_priv == priv)
1180 /* Secondary process? */
1181 if (sd->data.dev_private == priv)
1183 rte_spinlock_unlock(&sd->lock);
1184 if (++port_id == RTE_DIM(mlx5_secondary_data))
1187 /* Switch to secondary private structure. If private data has already
1188 * been updated by another thread, there is nothing else to do. */
1189 priv = sd->data.dev_private;
1190 if (priv->dev->data == &sd->data)
1192 /* Sanity checks. Secondary private structure is supposed to point
1193 * to local eth_dev, itself still pointing to the shared device data
1194 * structure allocated by the primary process. */
1195 assert(sd->shared_dev_data != &sd->data);
1196 assert(sd->data.nb_tx_queues == 0);
1197 assert(sd->data.tx_queues == NULL);
1198 assert(sd->data.nb_rx_queues == 0);
1199 assert(sd->data.rx_queues == NULL);
1200 assert(priv != sd->primary_priv);
1201 assert(priv->dev->data == sd->shared_dev_data);
1202 assert(priv->txqs_n == 0);
1203 assert(priv->txqs == NULL);
1204 assert(priv->rxqs_n == 0);
1205 assert(priv->rxqs == NULL);
1206 nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1207 nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1208 /* Allocate local storage for queues. */
1209 tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1210 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1211 RTE_CACHE_LINE_SIZE);
1212 rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1213 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1214 RTE_CACHE_LINE_SIZE);
1215 if (tx_queues == NULL || rx_queues == NULL)
1217 /* Lock to prevent control operations during setup. */
1220 for (i = 0; i != nb_tx_queues; ++i) {
1221 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1224 if (primary_txq == NULL)
1226 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0,
1227 primary_txq->socket);
1229 if (txq_setup(priv->dev,
1231 primary_txq->elts_n * MLX5_PMD_SGE_WR_N,
1232 primary_txq->socket,
1234 txq->stats.idx = primary_txq->stats.idx;
1241 txq = tx_queues[--i];
1248 for (i = 0; i != nb_rx_queues; ++i) {
1249 struct rxq *primary_rxq = (*sd->primary_priv->rxqs)[i];
1251 if (primary_rxq == NULL)
1253 /* Not supported yet. */
1254 rx_queues[i] = NULL;
1256 /* Update everything. */
1257 priv->txqs = (void *)tx_queues;
1258 priv->txqs_n = nb_tx_queues;
1259 priv->rxqs = (void *)rx_queues;
1260 priv->rxqs_n = nb_rx_queues;
1261 sd->data.rx_queues = rx_queues;
1262 sd->data.tx_queues = tx_queues;
1263 sd->data.nb_rx_queues = nb_rx_queues;
1264 sd->data.nb_tx_queues = nb_tx_queues;
1265 sd->data.dev_link = sd->shared_dev_data->dev_link;
1266 sd->data.mtu = sd->shared_dev_data->mtu;
1267 memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1268 sizeof(sd->data.rx_queue_state));
1269 memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1270 sizeof(sd->data.tx_queue_state));
1271 sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1272 /* Use local data from now on. */
1274 priv->dev->data = &sd->data;
1276 priv->dev->tx_pkt_burst = mlx5_tx_burst;
1277 priv->dev->rx_pkt_burst = removed_rx_burst;
1280 /* More sanity checks. */
1281 assert(priv->dev->tx_pkt_burst == mlx5_tx_burst);
1282 assert(priv->dev->rx_pkt_burst == removed_rx_burst);
1283 assert(priv->dev->data == &sd->data);
1284 rte_spinlock_unlock(&sd->lock);
1288 rte_free(tx_queues);
1289 rte_free(rx_queues);
1290 rte_spinlock_unlock(&sd->lock);