1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox.
19 #include <sys/ioctl.h>
20 #include <sys/socket.h>
21 #include <sys/utsname.h>
22 #include <netinet/in.h>
23 #include <linux/ethtool.h>
24 #include <linux/sockios.h>
25 #include <linux/version.h>
30 #include <rte_atomic.h>
31 #include <rte_ethdev_driver.h>
32 #include <rte_bus_pci.h>
34 #include <rte_common.h>
35 #include <rte_interrupts.h>
36 #include <rte_alarm.h>
37 #include <rte_malloc.h>
40 #include "mlx5_glue.h"
41 #include "mlx5_rxtx.h"
42 #include "mlx5_utils.h"
44 /* Add defines in case the running kernel is not the same as user headers. */
45 #ifndef ETHTOOL_GLINKSETTINGS
46 struct ethtool_link_settings {
55 uint8_t eth_tp_mdix_ctrl;
56 int8_t link_mode_masks_nwords;
58 uint32_t link_mode_masks[];
61 #define ETHTOOL_GLINKSETTINGS 0x0000004c
62 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
63 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
64 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
65 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
66 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
67 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
68 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
69 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
70 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
71 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
72 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
73 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
74 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
75 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
76 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
77 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
79 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
80 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
81 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
82 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
84 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
85 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
86 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
88 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
89 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
90 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
91 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
92 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
96 * Get interface name from private structure.
99 * Pointer to private structure.
101 * Interface name output buffer.
104 * 0 on success, -1 on failure and errno is set.
107 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
111 unsigned int dev_type = 0;
112 unsigned int dev_port_prev = ~0u;
113 char match[IF_NAMESIZE] = "";
116 MKSTR(path, "%s/device/net", priv->ibdev_path);
122 while ((dent = readdir(dir)) != NULL) {
123 char *name = dent->d_name;
125 unsigned int dev_port;
128 if ((name[0] == '.') &&
129 ((name[1] == '\0') ||
130 ((name[1] == '.') && (name[2] == '\0'))))
133 MKSTR(path, "%s/device/net/%s/%s",
134 priv->ibdev_path, name,
135 (dev_type ? "dev_id" : "dev_port"));
137 file = fopen(path, "rb");
142 * Switch to dev_id when dev_port does not exist as
143 * is the case with Linux kernel versions < 3.15.
154 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
159 * Switch to dev_id when dev_port returns the same value for
160 * all ports. May happen when using a MOFED release older than
161 * 3.0 with a Linux kernel >= 3.15.
163 if (dev_port == dev_port_prev)
165 dev_port_prev = dev_port;
166 if (dev_port == (priv->port - 1u))
167 snprintf(match, sizeof(match), "%s", name);
170 if (match[0] == '\0')
172 strncpy(*ifname, match, sizeof(*ifname));
177 * Perform ifreq ioctl() on associated Ethernet device.
180 * Pointer to private structure.
182 * Request number to pass to ioctl().
184 * Interface request structure output buffer.
187 * 0 on success, -1 on failure and errno is set.
190 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
192 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
197 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
198 ret = ioctl(sock, req, ifr);
207 * Pointer to private structure.
209 * MTU value output buffer.
212 * 0 on success, -1 on failure and errno is set.
215 priv_get_mtu(struct priv *priv, uint16_t *mtu)
217 struct ifreq request;
218 int ret = priv_ifreq(priv, SIOCGIFMTU, &request);
222 *mtu = request.ifr_mtu;
230 * Pointer to private structure.
235 * 0 on success, -1 on failure and errno is set.
238 priv_set_mtu(struct priv *priv, uint16_t mtu)
240 struct ifreq request = { .ifr_mtu = mtu, };
242 return priv_ifreq(priv, SIOCSIFMTU, &request);
249 * Pointer to private structure.
251 * Bitmask for flags that must remain untouched.
253 * Bitmask for flags to modify.
256 * 0 on success, -1 on failure and errno is set.
259 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
261 struct ifreq request;
262 int ret = priv_ifreq(priv, SIOCGIFFLAGS, &request);
266 request.ifr_flags &= keep;
267 request.ifr_flags |= flags & ~keep;
268 return priv_ifreq(priv, SIOCSIFFLAGS, &request);
272 * Ethernet device configuration.
274 * Prepare the driver for a given number of TX and RX queues.
277 * Pointer to Ethernet device structure.
280 * 0 on success, errno value on failure.
283 dev_configure(struct rte_eth_dev *dev)
285 struct priv *priv = dev->data->dev_private;
286 unsigned int rxqs_n = dev->data->nb_rx_queues;
287 unsigned int txqs_n = dev->data->nb_tx_queues;
290 unsigned int reta_idx_n;
291 const uint8_t use_app_rss_key =
292 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
293 uint64_t supp_tx_offloads = mlx5_priv_get_tx_port_offloads(priv);
294 uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
295 uint64_t supp_rx_offloads =
296 (mlx5_priv_get_rx_port_offloads(priv) |
297 mlx5_priv_get_rx_queue_offloads(priv));
298 uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
300 if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
301 ERROR("Some Tx offloads are not supported "
302 "requested 0x%" PRIx64 " supported 0x%" PRIx64,
303 tx_offloads, supp_tx_offloads);
306 if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
307 ERROR("Some Rx offloads are not supported "
308 "requested 0x%" PRIx64 " supported 0x%" PRIx64,
309 rx_offloads, supp_rx_offloads);
312 if (use_app_rss_key &&
313 (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
314 rss_hash_default_key_len)) {
315 /* MLX5 RSS only support 40bytes key. */
318 priv->rss_conf.rss_key =
319 rte_realloc(priv->rss_conf.rss_key,
320 rss_hash_default_key_len, 0);
321 if (!priv->rss_conf.rss_key) {
322 ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
325 memcpy(priv->rss_conf.rss_key,
327 dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
328 rss_hash_default_key,
329 rss_hash_default_key_len);
330 priv->rss_conf.rss_key_len = rss_hash_default_key_len;
331 priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
332 priv->rxqs = (void *)dev->data->rx_queues;
333 priv->txqs = (void *)dev->data->tx_queues;
334 if (txqs_n != priv->txqs_n) {
335 INFO("%p: TX queues number update: %u -> %u",
336 (void *)dev, priv->txqs_n, txqs_n);
337 priv->txqs_n = txqs_n;
339 if (rxqs_n > priv->config.ind_table_max_size) {
340 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
343 if (rxqs_n == priv->rxqs_n)
345 INFO("%p: RX queues number update: %u -> %u",
346 (void *)dev, priv->rxqs_n, rxqs_n);
347 priv->rxqs_n = rxqs_n;
348 /* If the requested number of RX queues is not a power of two, use the
349 * maximum indirection table size for better balancing.
350 * The result is always rounded to the next power of two. */
351 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
352 priv->config.ind_table_max_size :
354 if (priv_rss_reta_index_resize(priv, reta_idx_n))
356 /* When the number of RX queues is not a power of two, the remaining
357 * table entries are padded with reused WQs and hashes are not spread
359 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
360 (*priv->reta_idx)[i] = j;
368 * DPDK callback for Ethernet device configuration.
371 * Pointer to Ethernet device structure.
374 * 0 on success, negative errno value on failure.
377 mlx5_dev_configure(struct rte_eth_dev *dev)
379 struct priv *priv = dev->data->dev_private;
383 ret = dev_configure(dev);
390 * DPDK callback to get information about the device.
393 * Pointer to Ethernet device structure.
395 * Info structure output buffer.
398 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
400 struct priv *priv = dev->data->dev_private;
401 struct mlx5_dev_config *config = &priv->config;
403 char ifname[IF_NAMESIZE];
405 info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
408 /* FIXME: we should ask the device for these values. */
409 info->min_rx_bufsize = 32;
410 info->max_rx_pktlen = 65536;
412 * Since we need one CQ per QP, the limit is the minimum number
413 * between the two values.
415 max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
416 priv->device_attr.orig_attr.max_qp);
417 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
420 info->max_rx_queues = max;
421 info->max_tx_queues = max;
422 info->max_mac_addrs = RTE_DIM(priv->mac);
423 info->rx_queue_offload_capa =
424 mlx5_priv_get_rx_queue_offloads(priv);
425 info->rx_offload_capa = (mlx5_priv_get_rx_port_offloads(priv) |
426 info->rx_queue_offload_capa);
427 info->tx_offload_capa = mlx5_priv_get_tx_port_offloads(priv);
428 if (priv_get_ifname(priv, &ifname) == 0)
429 info->if_index = if_nametoindex(ifname);
430 info->reta_size = priv->reta_idx_n ?
431 priv->reta_idx_n : config->ind_table_max_size;
432 info->hash_key_size = priv->rss_conf.rss_key_len;
433 info->speed_capa = priv->link_speed_capa;
434 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
439 * Get supported packet types.
442 * Pointer to Ethernet device structure.
445 * A pointer to the supported Packet types array.
448 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
450 static const uint32_t ptypes[] = {
451 /* refers to rxq_cq_to_pkt_type() */
453 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
454 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
455 RTE_PTYPE_L4_NONFRAG,
459 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
460 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
461 RTE_PTYPE_INNER_L4_NONFRAG,
462 RTE_PTYPE_INNER_L4_FRAG,
463 RTE_PTYPE_INNER_L4_TCP,
464 RTE_PTYPE_INNER_L4_UDP,
468 if (dev->rx_pkt_burst == mlx5_rx_burst ||
469 dev->rx_pkt_burst == mlx5_rx_burst_vec)
475 * DPDK callback to retrieve physical link information.
478 * Pointer to Ethernet device structure.
481 * 0 on success, -1 on error.
484 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev)
486 struct priv *priv = dev->data->dev_private;
487 struct ethtool_cmd edata = {
488 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
491 struct rte_eth_link dev_link;
494 /* priv_lock() is not taken to allow concurrent calls. */
496 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
497 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
500 memset(&dev_link, 0, sizeof(dev_link));
501 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
502 (ifr.ifr_flags & IFF_RUNNING));
503 ifr.ifr_data = (void *)&edata;
504 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
505 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
509 link_speed = ethtool_cmd_speed(&edata);
510 if (link_speed == -1)
511 dev_link.link_speed = 0;
513 dev_link.link_speed = link_speed;
514 priv->link_speed_capa = 0;
515 if (edata.supported & SUPPORTED_Autoneg)
516 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
517 if (edata.supported & (SUPPORTED_1000baseT_Full |
518 SUPPORTED_1000baseKX_Full))
519 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
520 if (edata.supported & SUPPORTED_10000baseKR_Full)
521 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
522 if (edata.supported & (SUPPORTED_40000baseKR4_Full |
523 SUPPORTED_40000baseCR4_Full |
524 SUPPORTED_40000baseSR4_Full |
525 SUPPORTED_40000baseLR4_Full))
526 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
527 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
528 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
529 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
530 ETH_LINK_SPEED_FIXED);
531 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
532 /* Link status changed. */
533 dev->data->dev_link = dev_link;
536 /* Link status is still the same. */
541 * Retrieve physical link information (unlocked version using new ioctl).
544 * Pointer to Ethernet device structure.
547 * 0 on success, -1 on error.
550 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev)
552 struct priv *priv = dev->data->dev_private;
553 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
555 struct rte_eth_link dev_link;
558 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
559 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
562 memset(&dev_link, 0, sizeof(dev_link));
563 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
564 (ifr.ifr_flags & IFF_RUNNING));
565 ifr.ifr_data = (void *)&gcmd;
566 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
567 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
571 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
573 alignas(struct ethtool_link_settings)
574 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
575 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
576 struct ethtool_link_settings *ecmd = (void *)data;
579 ifr.ifr_data = (void *)ecmd;
580 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
581 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
585 dev_link.link_speed = ecmd->speed;
586 sc = ecmd->link_mode_masks[0] |
587 ((uint64_t)ecmd->link_mode_masks[1] << 32);
588 priv->link_speed_capa = 0;
589 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
590 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
591 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
592 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
593 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
594 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
595 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
596 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
597 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
598 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
599 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
600 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
601 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
602 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
603 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
604 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
605 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
606 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
607 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
608 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
609 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
610 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
611 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
612 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
613 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
614 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
615 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
616 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
617 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
618 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
619 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
620 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
621 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
622 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
623 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
624 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
625 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
626 ETH_LINK_SPEED_FIXED);
627 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
628 /* Link status changed. */
629 dev->data->dev_link = dev_link;
632 /* Link status is still the same. */
637 * Enable receiving and transmitting traffic.
640 * Pointer to private structure.
643 priv_link_start(struct priv *priv)
645 struct rte_eth_dev *dev = priv->dev;
648 dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
649 dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
650 err = priv_dev_traffic_enable(priv, dev);
652 ERROR("%p: error occurred while configuring control flows: %s",
653 (void *)priv, strerror(err));
654 err = priv_flow_start(priv, &priv->flows);
656 ERROR("%p: error occurred while configuring flows: %s",
657 (void *)priv, strerror(err));
661 * Disable receiving and transmitting traffic.
664 * Pointer to private structure.
667 priv_link_stop(struct priv *priv)
669 struct rte_eth_dev *dev = priv->dev;
671 priv_flow_stop(priv, &priv->flows);
672 priv_dev_traffic_disable(priv, dev);
673 dev->rx_pkt_burst = removed_rx_burst;
674 dev->tx_pkt_burst = removed_tx_burst;
678 * Retrieve physical link information and update rx/tx_pkt_burst callbacks
682 * Pointer to private structure.
683 * @param wait_to_complete
684 * Wait for request completion (ignored).
687 priv_link_update(struct priv *priv, int wait_to_complete __rte_unused)
689 struct rte_eth_dev *dev = priv->dev;
690 struct utsname utsname;
693 struct rte_eth_link dev_link = dev->data->dev_link;
695 if (uname(&utsname) == -1 ||
696 sscanf(utsname.release, "%d.%d.%d",
697 &ver[0], &ver[1], &ver[2]) != 3 ||
698 KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
699 ret = mlx5_link_update_unlocked_gset(dev);
701 ret = mlx5_link_update_unlocked_gs(dev);
702 /* If lsc interrupt is disabled, should always be ready for traffic. */
703 if (!dev->data->dev_conf.intr_conf.lsc) {
704 priv_link_start(priv);
707 /* Re-select burst callbacks only if link status has been changed. */
708 if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
709 if (dev->data->dev_link.link_status == ETH_LINK_UP)
710 priv_link_start(priv);
712 priv_link_stop(priv);
718 * Querying the link status till it changes to the desired state.
719 * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
722 * Pointer to private structure.
724 * Link desired status.
727 * 0 on success, negative errno value on failure.
730 priv_force_link_status_change(struct priv *priv, int status)
734 while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
735 priv_link_update(priv, 0);
736 if (priv->dev->data->dev_link.link_status == status)
745 * DPDK callback to retrieve physical link information.
748 * Pointer to Ethernet device structure.
749 * @param wait_to_complete
750 * Wait for request completion (ignored).
753 * 0 on success, -1 on error.
756 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete __rte_unused)
758 struct priv *priv = dev->data->dev_private;
762 ret = priv_link_update(priv, wait_to_complete);
768 * DPDK callback to change the MTU.
771 * Pointer to Ethernet device structure.
776 * 0 on success, negative errno value on failure.
779 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
781 struct priv *priv = dev->data->dev_private;
786 ret = priv_get_mtu(priv, &kern_mtu);
789 /* Set kernel interface MTU first. */
790 ret = priv_set_mtu(priv, mtu);
793 ret = priv_get_mtu(priv, &kern_mtu);
796 if (kern_mtu == mtu) {
798 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
804 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
812 * DPDK callback to get flow control status.
815 * Pointer to Ethernet device structure.
816 * @param[out] fc_conf
817 * Flow control output buffer.
820 * 0 on success, negative errno value on failure.
823 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
825 struct priv *priv = dev->data->dev_private;
827 struct ethtool_pauseparam ethpause = {
828 .cmd = ETHTOOL_GPAUSEPARAM
832 ifr.ifr_data = (void *)ðpause;
834 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
836 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
842 fc_conf->autoneg = ethpause.autoneg;
843 if (ethpause.rx_pause && ethpause.tx_pause)
844 fc_conf->mode = RTE_FC_FULL;
845 else if (ethpause.rx_pause)
846 fc_conf->mode = RTE_FC_RX_PAUSE;
847 else if (ethpause.tx_pause)
848 fc_conf->mode = RTE_FC_TX_PAUSE;
850 fc_conf->mode = RTE_FC_NONE;
860 * DPDK callback to modify flow control parameters.
863 * Pointer to Ethernet device structure.
865 * Flow control parameters.
868 * 0 on success, negative errno value on failure.
871 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
873 struct priv *priv = dev->data->dev_private;
875 struct ethtool_pauseparam ethpause = {
876 .cmd = ETHTOOL_SPAUSEPARAM
880 ifr.ifr_data = (void *)ðpause;
881 ethpause.autoneg = fc_conf->autoneg;
882 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
883 (fc_conf->mode & RTE_FC_RX_PAUSE))
884 ethpause.rx_pause = 1;
886 ethpause.rx_pause = 0;
888 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
889 (fc_conf->mode & RTE_FC_TX_PAUSE))
890 ethpause.tx_pause = 1;
892 ethpause.tx_pause = 0;
895 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
897 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
911 * Get PCI information from struct ibv_device.
914 * Pointer to Ethernet device structure.
915 * @param[out] pci_addr
916 * PCI bus address output buffer.
919 * 0 on success, -1 on failure and errno is set.
922 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
923 struct rte_pci_addr *pci_addr)
927 MKSTR(path, "%s/device/uevent", device->ibdev_path);
929 file = fopen(path, "rb");
932 while (fgets(line, sizeof(line), file) == line) {
933 size_t len = strlen(line);
936 /* Truncate long lines. */
937 if (len == (sizeof(line) - 1))
938 while (line[(len - 1)] != '\n') {
942 line[(len - 1)] = ret;
944 /* Extract information. */
947 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
951 &pci_addr->function) == 4) {
961 * Update the link status.
964 * Pointer to private structure.
967 * Zero if the callback process can be called immediately.
970 priv_link_status_update(struct priv *priv)
972 struct rte_eth_link *link = &priv->dev->data->dev_link;
974 priv_link_update(priv, 0);
975 if (((link->link_speed == 0) && link->link_status) ||
976 ((link->link_speed != 0) && !link->link_status)) {
978 * Inconsistent status. Event likely occurred before the
979 * kernel netdevice exposes the new status.
981 if (!priv->pending_alarm) {
982 priv->pending_alarm = 1;
983 rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
984 mlx5_dev_link_status_handler,
988 } else if (unlikely(priv->pending_alarm)) {
989 /* Link interrupt occurred while alarm is already scheduled. */
990 priv->pending_alarm = 0;
991 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
997 * Device status handler.
1000 * Pointer to private structure.
1002 * Pointer to event flags holder.
1005 * Events bitmap of callback process which can be called immediately.
1008 priv_dev_status_handler(struct priv *priv)
1010 struct ibv_async_event event;
1013 /* Read all message and acknowledge them. */
1015 if (mlx5_glue->get_async_event(priv->ctx, &event))
1017 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1018 event.event_type == IBV_EVENT_PORT_ERR) &&
1019 (priv->dev->data->dev_conf.intr_conf.lsc == 1))
1020 ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1021 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1022 priv->dev->data->dev_conf.intr_conf.rmv == 1)
1023 ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1025 DEBUG("event type %d on port %d not handled",
1026 event.event_type, event.element.port_num);
1027 mlx5_glue->ack_async_event(&event);
1029 if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1030 if (priv_link_status_update(priv))
1031 ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1036 * Handle delayed link status event.
1039 * Registered argument.
1042 mlx5_dev_link_status_handler(void *arg)
1044 struct rte_eth_dev *dev = arg;
1045 struct priv *priv = dev->data->dev_private;
1048 while (!priv_trylock(priv)) {
1049 /* Alarm is being canceled. */
1050 if (priv->pending_alarm == 0)
1054 priv->pending_alarm = 0;
1055 ret = priv_link_status_update(priv);
1058 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1062 * Handle interrupts from the NIC.
1064 * @param[in] intr_handle
1065 * Interrupt handler.
1067 * Callback argument.
1070 mlx5_dev_interrupt_handler(void *cb_arg)
1072 struct rte_eth_dev *dev = cb_arg;
1073 struct priv *priv = dev->data->dev_private;
1077 events = priv_dev_status_handler(priv);
1079 if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1080 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1081 if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1082 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1086 * Handle interrupts from the socket.
1089 * Callback argument.
1092 mlx5_dev_handler_socket(void *cb_arg)
1094 struct rte_eth_dev *dev = cb_arg;
1095 struct priv *priv = dev->data->dev_private;
1098 priv_socket_handle(priv);
1103 * Uninstall interrupt handler.
1106 * Pointer to private structure.
1108 * Pointer to the rte_eth_dev structure.
1111 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1113 if (dev->data->dev_conf.intr_conf.lsc ||
1114 dev->data->dev_conf.intr_conf.rmv)
1115 rte_intr_callback_unregister(&priv->intr_handle,
1116 mlx5_dev_interrupt_handler, dev);
1117 if (priv->primary_socket)
1118 rte_intr_callback_unregister(&priv->intr_handle_socket,
1119 mlx5_dev_handler_socket, dev);
1120 if (priv->pending_alarm) {
1121 priv->pending_alarm = 0;
1122 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1124 priv->intr_handle.fd = 0;
1125 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1126 priv->intr_handle_socket.fd = 0;
1127 priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1131 * Install interrupt handler.
1134 * Pointer to private structure.
1136 * Pointer to the rte_eth_dev structure.
1139 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1143 assert(priv->ctx->async_fd > 0);
1144 flags = fcntl(priv->ctx->async_fd, F_GETFL);
1145 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1147 INFO("failed to change file descriptor async event queue");
1148 dev->data->dev_conf.intr_conf.lsc = 0;
1149 dev->data->dev_conf.intr_conf.rmv = 0;
1151 if (dev->data->dev_conf.intr_conf.lsc ||
1152 dev->data->dev_conf.intr_conf.rmv) {
1153 priv->intr_handle.fd = priv->ctx->async_fd;
1154 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1155 rte_intr_callback_register(&priv->intr_handle,
1156 mlx5_dev_interrupt_handler, dev);
1159 rc = priv_socket_init(priv);
1160 if (!rc && priv->primary_socket) {
1161 priv->intr_handle_socket.fd = priv->primary_socket;
1162 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1163 rte_intr_callback_register(&priv->intr_handle_socket,
1164 mlx5_dev_handler_socket, dev);
1169 * Change the link state (UP / DOWN).
1172 * Pointer to private data structure.
1174 * Nonzero for link up, otherwise link down.
1177 * 0 on success, errno value on failure.
1180 priv_dev_set_link(struct priv *priv, int up)
1182 return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1186 * DPDK callback to bring the link DOWN.
1189 * Pointer to Ethernet device structure.
1192 * 0 on success, errno value on failure.
1195 mlx5_set_link_down(struct rte_eth_dev *dev)
1197 struct priv *priv = dev->data->dev_private;
1201 err = priv_dev_set_link(priv, 0);
1207 * DPDK callback to bring the link UP.
1210 * Pointer to Ethernet device structure.
1213 * 0 on success, errno value on failure.
1216 mlx5_set_link_up(struct rte_eth_dev *dev)
1218 struct priv *priv = dev->data->dev_private;
1222 err = priv_dev_set_link(priv, 1);
1228 * Configure the TX function to use.
1231 * Pointer to private data structure.
1233 * Pointer to rte_eth_dev structure.
1236 * Pointer to selected Tx burst function.
1239 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1241 eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1242 struct mlx5_dev_config *config = &priv->config;
1243 uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1244 int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1245 DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1246 DEV_TX_OFFLOAD_GRE_TNL_TSO));
1247 int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1249 assert(priv != NULL);
1250 /* Select appropriate TX function. */
1251 if (vlan_insert || tso)
1252 return tx_pkt_burst;
1253 if (config->mps == MLX5_MPW_ENHANCED) {
1254 if (priv_check_vec_tx_support(priv, dev) > 0) {
1255 if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1256 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1258 tx_pkt_burst = mlx5_tx_burst_vec;
1259 DEBUG("selected Enhanced MPW TX vectorized function");
1261 tx_pkt_burst = mlx5_tx_burst_empw;
1262 DEBUG("selected Enhanced MPW TX function");
1264 } else if (config->mps && (config->txq_inline > 0)) {
1265 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1266 DEBUG("selected MPW inline TX function");
1267 } else if (config->mps) {
1268 tx_pkt_burst = mlx5_tx_burst_mpw;
1269 DEBUG("selected MPW TX function");
1271 return tx_pkt_burst;
1275 * Configure the RX function to use.
1278 * Pointer to private data structure.
1280 * Pointer to rte_eth_dev structure.
1283 * Pointer to selected Rx burst function.
1286 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1288 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1290 assert(priv != NULL);
1291 if (priv_check_vec_rx_support(priv) > 0) {
1292 rx_pkt_burst = mlx5_rx_burst_vec;
1293 DEBUG("selected RX vectorized function");
1295 return rx_pkt_burst;
1299 * Check if mlx5 device was removed.
1302 * Pointer to Ethernet device structure.
1305 * 1 when device is removed, otherwise 0.
1308 mlx5_is_removed(struct rte_eth_dev *dev)
1310 struct ibv_device_attr device_attr;
1311 struct priv *priv = dev->data->dev_private;
1313 if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)