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>
48 #include <linux/ethtool.h>
49 #include <linux/sockios.h>
52 /* DPDK headers don't like -pedantic. */
54 #pragma GCC diagnostic ignored "-pedantic"
56 #include <rte_atomic.h>
57 #include <rte_ethdev.h>
59 #include <rte_common.h>
60 #include <rte_interrupts.h>
61 #include <rte_alarm.h>
63 #pragma GCC diagnostic error "-pedantic"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_utils.h"
71 * Get interface name from private structure.
74 * Pointer to private structure.
76 * Interface name output buffer.
79 * 0 on success, -1 on failure and errno is set.
82 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
86 unsigned int dev_type = 0;
87 unsigned int dev_port_prev = ~0u;
88 char match[IF_NAMESIZE] = "";
91 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
97 while ((dent = readdir(dir)) != NULL) {
98 char *name = dent->d_name;
100 unsigned int dev_port;
103 if ((name[0] == '.') &&
104 ((name[1] == '\0') ||
105 ((name[1] == '.') && (name[2] == '\0'))))
108 MKSTR(path, "%s/device/net/%s/%s",
109 priv->ctx->device->ibdev_path, name,
110 (dev_type ? "dev_id" : "dev_port"));
112 file = fopen(path, "rb");
117 * Switch to dev_id when dev_port does not exist as
118 * is the case with Linux kernel versions < 3.15.
129 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
134 * Switch to dev_id when dev_port returns the same value for
135 * all ports. May happen when using a MOFED release older than
136 * 3.0 with a Linux kernel >= 3.15.
138 if (dev_port == dev_port_prev)
140 dev_port_prev = dev_port;
141 if (dev_port == (priv->port - 1u))
142 snprintf(match, sizeof(match), "%s", name);
145 if (match[0] == '\0')
147 strncpy(*ifname, match, sizeof(*ifname));
152 * Read from sysfs entry.
155 * Pointer to private structure.
157 * Entry name relative to sysfs path.
159 * Data output buffer.
164 * 0 on success, -1 on failure and errno is set.
167 priv_sysfs_read(const struct priv *priv, const char *entry,
168 char *buf, size_t size)
170 char ifname[IF_NAMESIZE];
175 if (priv_get_ifname(priv, &ifname))
178 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
181 file = fopen(path, "rb");
184 ret = fread(buf, 1, size, file);
186 if (((size_t)ret < size) && (ferror(file)))
196 * Write to sysfs entry.
199 * Pointer to private structure.
201 * Entry name relative to sysfs path.
208 * 0 on success, -1 on failure and errno is set.
211 priv_sysfs_write(const struct priv *priv, const char *entry,
212 char *buf, size_t size)
214 char ifname[IF_NAMESIZE];
219 if (priv_get_ifname(priv, &ifname))
222 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
225 file = fopen(path, "wb");
228 ret = fwrite(buf, 1, size, file);
230 if (((size_t)ret < size) || (ferror(file)))
240 * Get unsigned long sysfs property.
243 * Pointer to private structure.
245 * Entry name relative to sysfs path.
247 * Value output buffer.
250 * 0 on success, -1 on failure and errno is set.
253 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
256 unsigned long value_ret;
259 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
261 DEBUG("cannot read %s value from sysfs: %s",
262 name, strerror(errno));
265 value_str[ret] = '\0';
267 value_ret = strtoul(value_str, NULL, 0);
269 DEBUG("invalid %s value `%s': %s", name, value_str,
278 * Set unsigned long sysfs property.
281 * Pointer to private structure.
283 * Entry name relative to sysfs path.
288 * 0 on success, -1 on failure and errno is set.
291 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
294 MKSTR(value_str, "%lu", value);
296 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
298 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
299 name, value_str, value, strerror(errno));
306 * Perform ifreq ioctl() on associated Ethernet device.
309 * Pointer to private structure.
311 * Request number to pass to ioctl().
313 * Interface request structure output buffer.
316 * 0 on success, -1 on failure and errno is set.
319 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
321 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
326 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
327 ret = ioctl(sock, req, ifr);
336 * Pointer to private structure.
338 * MTU value output buffer.
341 * 0 on success, -1 on failure and errno is set.
344 priv_get_mtu(struct priv *priv, uint16_t *mtu)
346 unsigned long ulong_mtu;
348 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
358 * Pointer to private structure.
363 * 0 on success, -1 on failure and errno is set.
366 priv_set_mtu(struct priv *priv, uint16_t mtu)
368 return priv_set_sysfs_ulong(priv, "mtu", mtu);
375 * Pointer to private structure.
377 * Bitmask for flags that must remain untouched.
379 * Bitmask for flags to modify.
382 * 0 on success, -1 on failure and errno is set.
385 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
389 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
393 return priv_set_sysfs_ulong(priv, "flags", tmp);
397 * Ethernet device configuration.
399 * Prepare the driver for a given number of TX and RX queues.
402 * Pointer to Ethernet device structure.
405 * 0 on success, errno value on failure.
408 dev_configure(struct rte_eth_dev *dev)
410 struct priv *priv = dev->data->dev_private;
411 unsigned int rxqs_n = dev->data->nb_rx_queues;
412 unsigned int txqs_n = dev->data->nb_tx_queues;
414 priv->rxqs = (void *)dev->data->rx_queues;
415 priv->txqs = (void *)dev->data->tx_queues;
416 if (txqs_n != priv->txqs_n) {
417 INFO("%p: TX queues number update: %u -> %u",
418 (void *)dev, priv->txqs_n, txqs_n);
419 priv->txqs_n = txqs_n;
421 if (rxqs_n == priv->rxqs_n)
423 INFO("%p: RX queues number update: %u -> %u",
424 (void *)dev, priv->rxqs_n, rxqs_n);
425 priv->rxqs_n = rxqs_n;
430 * DPDK callback for Ethernet device configuration.
433 * Pointer to Ethernet device structure.
436 * 0 on success, negative errno value on failure.
439 mlx5_dev_configure(struct rte_eth_dev *dev)
441 struct priv *priv = dev->data->dev_private;
445 ret = dev_configure(dev);
452 * DPDK callback to get information about the device.
455 * Pointer to Ethernet device structure.
457 * Info structure output buffer.
460 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
462 struct priv *priv = dev->data->dev_private;
464 char ifname[IF_NAMESIZE];
467 /* FIXME: we should ask the device for these values. */
468 info->min_rx_bufsize = 32;
469 info->max_rx_pktlen = 65536;
471 * Since we need one CQ per QP, the limit is the minimum number
472 * between the two values.
474 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
475 priv->device_attr.max_qp : priv->device_attr.max_cq);
476 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
479 info->max_rx_queues = max;
480 info->max_tx_queues = max;
481 /* Last array entry is reserved for broadcast. */
482 info->max_mac_addrs = (RTE_DIM(priv->mac) - 1);
483 info->rx_offload_capa =
485 (DEV_RX_OFFLOAD_IPV4_CKSUM |
486 DEV_RX_OFFLOAD_UDP_CKSUM |
487 DEV_RX_OFFLOAD_TCP_CKSUM) :
489 info->tx_offload_capa =
491 (DEV_TX_OFFLOAD_IPV4_CKSUM |
492 DEV_TX_OFFLOAD_UDP_CKSUM |
493 DEV_TX_OFFLOAD_TCP_CKSUM) :
495 if (priv_get_ifname(priv, &ifname) == 0)
496 info->if_index = if_nametoindex(ifname);
501 * DPDK callback to retrieve physical link information (unlocked version).
504 * Pointer to Ethernet device structure.
505 * @param wait_to_complete
506 * Wait for request completion (ignored).
509 mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
511 struct priv *priv = dev->data->dev_private;
512 struct ethtool_cmd edata = {
516 struct rte_eth_link dev_link;
519 (void)wait_to_complete;
520 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
521 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
524 memset(&dev_link, 0, sizeof(dev_link));
525 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
526 (ifr.ifr_flags & IFF_RUNNING));
527 ifr.ifr_data = &edata;
528 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
529 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
533 link_speed = ethtool_cmd_speed(&edata);
534 if (link_speed == -1)
535 dev_link.link_speed = 0;
537 dev_link.link_speed = link_speed;
538 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
539 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
540 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
541 /* Link status changed. */
542 dev->data->dev_link = dev_link;
545 /* Link status is still the same. */
550 * DPDK callback to retrieve physical link information.
553 * Pointer to Ethernet device structure.
554 * @param wait_to_complete
555 * Wait for request completion (ignored).
558 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
560 struct priv *priv = dev->data->dev_private;
564 ret = mlx5_link_update_unlocked(dev, wait_to_complete);
570 * DPDK callback to change the MTU.
572 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
573 * received). Use this as a hint to enable/disable scattered packets support
574 * and improve performance when not needed.
575 * Since failure is not an option, reconfiguring queues on the fly is not
579 * Pointer to Ethernet device structure.
584 * 0 on success, negative errno value on failure.
587 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
589 struct priv *priv = dev->data->dev_private;
592 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
596 /* Set kernel interface MTU first. */
597 if (priv_set_mtu(priv, mtu)) {
599 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
603 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
605 /* Temporarily replace RX handler with a fake one, assuming it has not
606 * been copied elsewhere. */
607 dev->rx_pkt_burst = removed_rx_burst;
608 /* Make sure everyone has left mlx5_rx_burst() and uses
609 * removed_rx_burst() instead. */
612 /* Reconfigure each RX queue. */
613 for (i = 0; (i != priv->rxqs_n); ++i) {
614 struct rxq *rxq = (*priv->rxqs)[i];
615 unsigned int max_frame_len;
620 /* Calculate new maximum frame length according to MTU and
621 * toggle scattered support (sp) if necessary. */
622 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
623 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
624 sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
625 /* Provide new values to rxq_setup(). */
626 dev->data->dev_conf.rxmode.jumbo_frame = sp;
627 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
628 ret = rxq_rehash(dev, rxq);
630 /* Force SP RX if that queue requires it and abort. */
632 rx_func = mlx5_rx_burst_sp;
635 /* Scattered burst function takes priority. */
637 rx_func = mlx5_rx_burst_sp;
639 /* Burst functions can now be called again. */
641 dev->rx_pkt_burst = rx_func;
649 * DPDK callback to get flow control status.
652 * Pointer to Ethernet device structure.
653 * @param[out] fc_conf
654 * Flow control output buffer.
657 * 0 on success, negative errno value on failure.
660 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
662 struct priv *priv = dev->data->dev_private;
664 struct ethtool_pauseparam ethpause = {
665 .cmd = ETHTOOL_GPAUSEPARAM
669 ifr.ifr_data = ðpause;
671 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
673 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
679 fc_conf->autoneg = ethpause.autoneg;
680 if (ethpause.rx_pause && ethpause.tx_pause)
681 fc_conf->mode = RTE_FC_FULL;
682 else if (ethpause.rx_pause)
683 fc_conf->mode = RTE_FC_RX_PAUSE;
684 else if (ethpause.tx_pause)
685 fc_conf->mode = RTE_FC_TX_PAUSE;
687 fc_conf->mode = RTE_FC_NONE;
697 * DPDK callback to modify flow control parameters.
700 * Pointer to Ethernet device structure.
702 * Flow control parameters.
705 * 0 on success, negative errno value on failure.
708 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
710 struct priv *priv = dev->data->dev_private;
712 struct ethtool_pauseparam ethpause = {
713 .cmd = ETHTOOL_SPAUSEPARAM
717 ifr.ifr_data = ðpause;
718 ethpause.autoneg = fc_conf->autoneg;
719 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
720 (fc_conf->mode & RTE_FC_RX_PAUSE))
721 ethpause.rx_pause = 1;
723 ethpause.rx_pause = 0;
725 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
726 (fc_conf->mode & RTE_FC_TX_PAUSE))
727 ethpause.tx_pause = 1;
729 ethpause.tx_pause = 0;
732 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
734 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
748 * Get PCI information from struct ibv_device.
751 * Pointer to Ethernet device structure.
752 * @param[out] pci_addr
753 * PCI bus address output buffer.
756 * 0 on success, -1 on failure and errno is set.
759 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
760 struct rte_pci_addr *pci_addr)
764 MKSTR(path, "%s/device/uevent", device->ibdev_path);
766 file = fopen(path, "rb");
769 while (fgets(line, sizeof(line), file) == line) {
770 size_t len = strlen(line);
773 /* Truncate long lines. */
774 if (len == (sizeof(line) - 1))
775 while (line[(len - 1)] != '\n') {
779 line[(len - 1)] = ret;
781 /* Extract information. */
784 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
788 &pci_addr->function) == 4) {
798 * Link status handler.
801 * Pointer to private structure.
803 * Pointer to the rte_eth_dev structure.
806 * Nonzero if the callback process can be called immediately.
809 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
811 struct ibv_async_event event;
815 /* Read all message and acknowledge them. */
817 if (ibv_get_async_event(priv->ctx, &event))
820 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
821 event.event_type == IBV_EVENT_PORT_ERR)
824 DEBUG("event type %d on port %d not handled",
825 event.event_type, event.element.port_num);
826 ibv_ack_async_event(&event);
829 if (port_change ^ priv->pending_alarm) {
830 struct rte_eth_link *link = &dev->data->dev_link;
832 priv->pending_alarm = 0;
833 mlx5_link_update_unlocked(dev, 0);
834 if (((link->link_speed == 0) && link->link_status) ||
835 ((link->link_speed != 0) && !link->link_status)) {
836 /* Inconsistent status, check again later. */
837 priv->pending_alarm = 1;
838 rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
839 mlx5_dev_link_status_handler,
848 * Handle delayed link status event.
851 * Registered argument.
854 mlx5_dev_link_status_handler(void *arg)
856 struct rte_eth_dev *dev = arg;
857 struct priv *priv = dev->data->dev_private;
861 assert(priv->pending_alarm == 1);
862 ret = priv_dev_link_status_handler(priv, dev);
865 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
869 * Handle interrupts from the NIC.
871 * @param[in] intr_handle
877 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
879 struct rte_eth_dev *dev = cb_arg;
880 struct priv *priv = dev->data->dev_private;
885 ret = priv_dev_link_status_handler(priv, dev);
888 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
892 * Uninstall interrupt handler.
895 * Pointer to private structure.
897 * Pointer to the rte_eth_dev structure.
900 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
902 if (!dev->data->dev_conf.intr_conf.lsc)
904 rte_intr_callback_unregister(&priv->intr_handle,
905 mlx5_dev_interrupt_handler,
907 if (priv->pending_alarm)
908 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
909 priv->pending_alarm = 0;
910 priv->intr_handle.fd = 0;
911 priv->intr_handle.type = 0;
915 * Install interrupt handler.
918 * Pointer to private structure.
920 * Pointer to the rte_eth_dev structure.
923 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
927 if (!dev->data->dev_conf.intr_conf.lsc)
929 assert(priv->ctx->async_fd > 0);
930 flags = fcntl(priv->ctx->async_fd, F_GETFL);
931 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
933 INFO("failed to change file descriptor async event queue");
934 dev->data->dev_conf.intr_conf.lsc = 0;
936 priv->intr_handle.fd = priv->ctx->async_fd;
937 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
938 rte_intr_callback_register(&priv->intr_handle,
939 mlx5_dev_interrupt_handler,