drivers/net/mlx5/mlx5_ethdev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2015 Mellanox.
   4  */
   5
   6 #define _GNU_SOURCE
   7
   8 #include <stddef.h>
   9 #include <assert.h>
  10 #include <inttypes.h>
  11 #include <unistd.h>
  12 #include <stdint.h>
  13 #include <stdio.h>
  14 #include <string.h>
  15 #include <stdlib.h>
  16 #include <errno.h>
  17 #include <dirent.h>
  18 #include <net/if.h>
  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>
  26 #include <fcntl.h>
  27 #include <stdalign.h>
  28 #include <sys/un.h>
  29
  30 #include <rte_atomic.h>
  31 #include <rte_ethdev_driver.h>
  32 #include <rte_bus_pci.h>
  33 #include <rte_mbuf.h>
  34 #include <rte_common.h>
  35 #include <rte_interrupts.h>
  36 #include <rte_alarm.h>
  37 #include <rte_malloc.h>
  38
  39 #include "mlx5.h"
  40 #include "mlx5_glue.h"
  41 #include "mlx5_rxtx.h"
  42 #include "mlx5_utils.h"
  43
  44 /* Add defines in case the running kernel is not the same as user headers. */
  45 #ifndef ETHTOOL_GLINKSETTINGS
  46 struct ethtool_link_settings {
  47         uint32_t cmd;
  48         uint32_t speed;
  49         uint8_t duplex;
  50         uint8_t port;
  51         uint8_t phy_address;
  52         uint8_t autoneg;
  53         uint8_t mdio_support;
  54         uint8_t eth_to_mdix;
  55         uint8_t eth_tp_mdix_ctrl;
  56         int8_t link_mode_masks_nwords;
  57         uint32_t reserved[8];
  58         uint32_t link_mode_masks[];
  59 };
  60
  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
  78 #endif
  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
  83 #endif
  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
  87 #endif
  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
  93 #endif
  94
  95 /**
  96  * Get interface name from private structure.
  97  *
  98  * @param[in] dev
  99  *   Pointer to Ethernet device.
 100  * @param[out] ifname
 101  *   Interface name output buffer.
 102  *
 103  * @return
 104  *   0 on success, a negative errno value otherwise and rte_errno is set.
 105  */
 106 int
 107 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
 108 {
 109         struct priv *priv = dev->data->dev_private;
 110         DIR *dir;
 111         struct dirent *dent;
 112         unsigned int dev_type = 0;
 113         unsigned int dev_port_prev = ~0u;
 114         char match[IF_NAMESIZE] = "";
 115
 116         {
 117                 MKSTR(path, "%s/device/net", priv->ibdev_path);
 118
 119                 dir = opendir(path);
 120                 if (dir == NULL) {
 121                         rte_errno = errno;
 122                         return -rte_errno;
 123                 }
 124         }
 125         while ((dent = readdir(dir)) != NULL) {
 126                 char *name = dent->d_name;
 127                 FILE *file;
 128                 unsigned int dev_port;
 129                 int r;
 130
 131                 if ((name[0] == '.') &&
 132                     ((name[1] == '\0') ||
 133                      ((name[1] == '.') && (name[2] == '\0'))))
 134                         continue;
 135
 136                 MKSTR(path, "%s/device/net/%s/%s",
 137                       priv->ibdev_path, name,
 138                       (dev_type ? "dev_id" : "dev_port"));
 139
 140                 file = fopen(path, "rb");
 141                 if (file == NULL) {
 142                         if (errno != ENOENT)
 143                                 continue;
 144                         /*
 145                          * Switch to dev_id when dev_port does not exist as
 146                          * is the case with Linux kernel versions < 3.15.
 147                          */
 148 try_dev_id:
 149                         match[0] = '\0';
 150                         if (dev_type)
 151                                 break;
 152                         dev_type = 1;
 153                         dev_port_prev = ~0u;
 154                         rewinddir(dir);
 155                         continue;
 156                 }
 157                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 158                 fclose(file);
 159                 if (r != 1)
 160                         continue;
 161                 /*
 162                  * Switch to dev_id when dev_port returns the same value for
 163                  * all ports. May happen when using a MOFED release older than
 164                  * 3.0 with a Linux kernel >= 3.15.
 165                  */
 166                 if (dev_port == dev_port_prev)
 167                         goto try_dev_id;
 168                 dev_port_prev = dev_port;
 169                 if (dev_port == (priv->port - 1u))
 170                         snprintf(match, sizeof(match), "%s", name);
 171         }
 172         closedir(dir);
 173         if (match[0] == '\0') {
 174                 rte_errno = ENOENT;
 175                 return -rte_errno;
 176         }
 177         strncpy(*ifname, match, sizeof(*ifname));
 178         return 0;
 179 }
 180
 181 /**
 182  * Perform ifreq ioctl() on associated Ethernet device.
 183  *
 184  * @param[in] dev
 185  *   Pointer to Ethernet device.
 186  * @param req
 187  *   Request number to pass to ioctl().
 188  * @param[out] ifr
 189  *   Interface request structure output buffer.
 190  *
 191  * @return
 192  *   0 on success, a negative errno value otherwise and rte_errno is set.
 193  */
 194 int
 195 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
 196 {
 197         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 198         int ret = 0;
 199
 200         if (sock == -1) {
 201                 rte_errno = errno;
 202                 return -rte_errno;
 203         }
 204         ret = mlx5_get_ifname(dev, &ifr->ifr_name);
 205         if (ret)
 206                 goto error;
 207         ret = ioctl(sock, req, ifr);
 208         if (ret == -1) {
 209                 rte_errno = errno;
 210                 goto error;
 211         }
 212         close(sock);
 213         return 0;
 214 error:
 215         close(sock);
 216         return -rte_errno;
 217 }
 218
 219 /**
 220  * Get device MTU.
 221  *
 222  * @param dev
 223  *   Pointer to Ethernet device.
 224  * @param[out] mtu
 225  *   MTU value output buffer.
 226  *
 227  * @return
 228  *   0 on success, a negative errno value otherwise and rte_errno is set.
 229  */
 230 int
 231 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
 232 {
 233         struct ifreq request;
 234         int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
 235
 236         if (ret)
 237                 return ret;
 238         *mtu = request.ifr_mtu;
 239         return 0;
 240 }
 241
 242 /**
 243  * Set device MTU.
 244  *
 245  * @param dev
 246  *   Pointer to Ethernet device.
 247  * @param mtu
 248  *   MTU value to set.
 249  *
 250  * @return
 251  *   0 on success, a negative errno value otherwise and rte_errno is set.
 252  */
 253 static int
 254 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 255 {
 256         struct ifreq request = { .ifr_mtu = mtu, };
 257
 258         return mlx5_ifreq(dev, SIOCSIFMTU, &request);
 259 }
 260
 261 /**
 262  * Set device flags.
 263  *
 264  * @param dev
 265  *   Pointer to Ethernet device.
 266  * @param keep
 267  *   Bitmask for flags that must remain untouched.
 268  * @param flags
 269  *   Bitmask for flags to modify.
 270  *
 271  * @return
 272  *   0 on success, a negative errno value otherwise and rte_errno is set.
 273  */
 274 int
 275 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
 276 {
 277         struct ifreq request;
 278         int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
 279
 280         if (ret)
 281                 return ret;
 282         request.ifr_flags &= keep;
 283         request.ifr_flags |= flags & ~keep;
 284         return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
 285 }
 286
 287 /**
 288  * DPDK callback for Ethernet device configuration.
 289  *
 290  * @param dev
 291  *   Pointer to Ethernet device structure.
 292  *
 293  * @return
 294  *   0 on success, a negative errno value otherwise and rte_errno is set.
 295  */
 296 int
 297 mlx5_dev_configure(struct rte_eth_dev *dev)
 298 {
 299         struct priv *priv = dev->data->dev_private;
 300         unsigned int rxqs_n = dev->data->nb_rx_queues;
 301         unsigned int txqs_n = dev->data->nb_tx_queues;
 302         unsigned int i;
 303         unsigned int j;
 304         unsigned int reta_idx_n;
 305         const uint8_t use_app_rss_key =
 306                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 307         uint64_t supp_tx_offloads = mlx5_get_tx_port_offloads(dev);
 308         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
 309         uint64_t supp_rx_offloads =
 310                 (mlx5_get_rx_port_offloads() |
 311                  mlx5_get_rx_queue_offloads(dev));
 312         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
 313         int ret = 0;
 314
 315         if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
 316                 DRV_LOG(ERR,
 317                         "port %u some Tx offloads are not supported requested"
 318                         " 0x%" PRIx64 " supported 0x%" PRIx64,
 319                         dev->data->port_id, tx_offloads, supp_tx_offloads);
 320                 rte_errno = ENOTSUP;
 321                 return -rte_errno;
 322         }
 323         if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
 324                 DRV_LOG(ERR,
 325                         "port %u some Rx offloads are not supported requested"
 326                         " 0x%" PRIx64 " supported 0x%" PRIx64,
 327                         dev->data->port_id, rx_offloads, supp_rx_offloads);
 328                 rte_errno = ENOTSUP;
 329                 return -rte_errno;
 330         }
 331         if (use_app_rss_key &&
 332             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 333              rss_hash_default_key_len)) {
 334                 /* MLX5 RSS only support 40bytes key. */
 335                 rte_errno = EINVAL;
 336                 return -rte_errno;
 337         }
 338         priv->rss_conf.rss_key =
 339                 rte_realloc(priv->rss_conf.rss_key,
 340                             rss_hash_default_key_len, 0);
 341         if (!priv->rss_conf.rss_key) {
 342                 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
 343                         dev->data->port_id, rxqs_n);
 344                 rte_errno = ENOMEM;
 345                 return -rte_errno;
 346         }
 347         memcpy(priv->rss_conf.rss_key,
 348                use_app_rss_key ?
 349                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 350                rss_hash_default_key,
 351                rss_hash_default_key_len);
 352         priv->rss_conf.rss_key_len = rss_hash_default_key_len;
 353         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 354         priv->rxqs = (void *)dev->data->rx_queues;
 355         priv->txqs = (void *)dev->data->tx_queues;
 356         if (txqs_n != priv->txqs_n) {
 357                 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
 358                         dev->data->port_id, priv->txqs_n, txqs_n);
 359                 priv->txqs_n = txqs_n;
 360         }
 361         if (rxqs_n > priv->config.ind_table_max_size) {
 362                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 363                         dev->data->port_id, rxqs_n);
 364                 rte_errno = EINVAL;
 365                 return -rte_errno;
 366         }
 367         if (rxqs_n == priv->rxqs_n)
 368                 return 0;
 369         DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 370                 dev->data->port_id, priv->rxqs_n, rxqs_n);
 371         priv->rxqs_n = rxqs_n;
 372         /* If the requested number of RX queues is not a power of two, use the
 373          * maximum indirection table size for better balancing.
 374          * The result is always rounded to the next power of two. */
 375         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 376                                      priv->config.ind_table_max_size :
 377                                      rxqs_n));
 378         ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
 379         if (ret)
 380                 return ret;
 381         /* When the number of RX queues is not a power of two, the remaining
 382          * table entries are padded with reused WQs and hashes are not spread
 383          * uniformly. */
 384         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 385                 (*priv->reta_idx)[i] = j;
 386                 if (++j == rxqs_n)
 387                         j = 0;
 388         }
 389         return 0;
 390 }
 391
 392 /**
 393  * DPDK callback to get information about the device.
 394  *
 395  * @param dev
 396  *   Pointer to Ethernet device structure.
 397  * @param[out] info
 398  *   Info structure output buffer.
 399  */
 400 void
 401 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 402 {
 403         struct priv *priv = dev->data->dev_private;
 404         struct mlx5_dev_config *config = &priv->config;
 405         unsigned int max;
 406         char ifname[IF_NAMESIZE];
 407
 408         info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
 409         /* FIXME: we should ask the device for these values. */
 410         info->min_rx_bufsize = 32;
 411         info->max_rx_pktlen = 65536;
 412         /*
 413          * Since we need one CQ per QP, the limit is the minimum number
 414          * between the two values.
 415          */
 416         max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
 417                       priv->device_attr.orig_attr.max_qp);
 418         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 419         if (max >= 65535)
 420                 max = 65535;
 421         info->max_rx_queues = max;
 422         info->max_tx_queues = max;
 423         info->max_mac_addrs = RTE_DIM(priv->mac);
 424         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
 425         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
 426                                  info->rx_queue_offload_capa);
 427         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
 428         if (mlx5_get_ifname(dev, &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;
 435 }
 436
 437 /**
 438  * Get supported packet types.
 439  *
 440  * @param dev
 441  *   Pointer to Ethernet device structure.
 442  *
 443  * @return
 444  *   A pointer to the supported Packet types array.
 445  */
 446 const uint32_t *
 447 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 448 {
 449         static const uint32_t ptypes[] = {
 450                 /* refers to rxq_cq_to_pkt_type() */
 451                 RTE_PTYPE_L2_ETHER,
 452                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 453                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 454                 RTE_PTYPE_L4_NONFRAG,
 455                 RTE_PTYPE_L4_FRAG,
 456                 RTE_PTYPE_L4_TCP,
 457                 RTE_PTYPE_L4_UDP,
 458                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 459                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 460                 RTE_PTYPE_INNER_L4_NONFRAG,
 461                 RTE_PTYPE_INNER_L4_FRAG,
 462                 RTE_PTYPE_INNER_L4_TCP,
 463                 RTE_PTYPE_INNER_L4_UDP,
 464                 RTE_PTYPE_UNKNOWN
 465         };
 466
 467         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 468             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 469                 return ptypes;
 470         return NULL;
 471 }
 472
 473 /**
 474  * DPDK callback to retrieve physical link information.
 475  *
 476  * @param dev
 477  *   Pointer to Ethernet device structure.
 478  *
 479  * @return
 480  *   0 on success, a negative errno value otherwise and rte_errno is set.
 481  */
 482 static int
 483 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev)
 484 {
 485         struct priv *priv = dev->data->dev_private;
 486         struct ethtool_cmd edata = {
 487                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 488         };
 489         struct ifreq ifr;
 490         struct rte_eth_link dev_link;
 491         int link_speed = 0;
 492         int ret;
 493
 494         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 495         if (ret) {
 496                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 497                         dev->data->port_id, strerror(rte_errno));
 498                 return ret;
 499         }
 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         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 505         if (ret) {
 506                 DRV_LOG(WARNING,
 507                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 508                         dev->data->port_id, strerror(rte_errno));
 509                 return ret;
 510         }
 511         link_speed = ethtool_cmd_speed(&edata);
 512         if (link_speed == -1)
 513                 dev_link.link_speed = 0;
 514         else
 515                 dev_link.link_speed = link_speed;
 516         priv->link_speed_capa = 0;
 517         if (edata.supported & SUPPORTED_Autoneg)
 518                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 519         if (edata.supported & (SUPPORTED_1000baseT_Full |
 520                                SUPPORTED_1000baseKX_Full))
 521                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 522         if (edata.supported & SUPPORTED_10000baseKR_Full)
 523                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 524         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 525                                SUPPORTED_40000baseCR4_Full |
 526                                SUPPORTED_40000baseSR4_Full |
 527                                SUPPORTED_40000baseLR4_Full))
 528                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 529         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 530                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 531         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 532                         ETH_LINK_SPEED_FIXED);
 533         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 534                 /* Link status changed. */
 535                 dev->data->dev_link = dev_link;
 536                 return 0;
 537         }
 538         /* Link status is still the same. */
 539         rte_errno = EAGAIN;
 540         return -rte_errno;
 541 }
 542
 543 /**
 544  * Retrieve physical link information (unlocked version using new ioctl).
 545  *
 546  * @param dev
 547  *   Pointer to Ethernet device structure.
 548  *
 549  * @return
 550  *   0 on success, a negative errno value otherwise and rte_errno is set.
 551  */
 552 static int
 553 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev)
 554 {
 555         struct priv *priv = dev->data->dev_private;
 556         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 557         struct ifreq ifr;
 558         struct rte_eth_link dev_link;
 559         uint64_t sc;
 560         int ret;
 561
 562         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 563         if (ret) {
 564                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 565                         dev->data->port_id, strerror(rte_errno));
 566                 return ret;
 567         }
 568         memset(&dev_link, 0, sizeof(dev_link));
 569         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 570                                 (ifr.ifr_flags & IFF_RUNNING));
 571         ifr.ifr_data = (void *)&gcmd;
 572         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 573         if (ret) {
 574                 DRV_LOG(DEBUG,
 575                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
 576                         " failed: %s",
 577                         dev->data->port_id, strerror(rte_errno));
 578                 return ret;
 579         }
 580         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 581
 582         alignas(struct ethtool_link_settings)
 583         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 584                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 585         struct ethtool_link_settings *ecmd = (void *)data;
 586
 587         *ecmd = gcmd;
 588         ifr.ifr_data = (void *)ecmd;
 589         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 590         if (ret) {
 591                 DRV_LOG(DEBUG,
 592                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
 593                         " failed: %s",
 594                         dev->data->port_id, strerror(rte_errno));
 595                 return ret;
 596         }
 597         dev_link.link_speed = ecmd->speed;
 598         sc = ecmd->link_mode_masks[0] |
 599                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 600         priv->link_speed_capa = 0;
 601         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 602                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 603         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 604                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
 605                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 606         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
 607                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
 608                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
 609                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 610         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
 611                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
 612                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 613         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
 614                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
 615                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
 616                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
 617                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 618         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
 619                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
 620                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
 621                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
 622                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 623         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
 624                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
 625                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
 626                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 627         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
 628                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
 629                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 630         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
 631                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
 632                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
 633                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
 634                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 635         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
 636                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 637         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 638                                   ETH_LINK_SPEED_FIXED);
 639         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 640                 /* Link status changed. */
 641                 dev->data->dev_link = dev_link;
 642                 return 0;
 643         }
 644         /* Link status is still the same. */
 645         rte_errno = EAGAIN;
 646         return -rte_errno;
 647 }
 648
 649 /**
 650  * Enable receiving and transmitting traffic.
 651  *
 652  * @param dev
 653  *   Pointer to Ethernet device.
 654  */
 655 static void
 656 mlx5_link_start(struct rte_eth_dev *dev)
 657 {
 658         struct priv *priv = dev->data->dev_private;
 659         int ret;
 660
 661         dev->tx_pkt_burst = mlx5_select_tx_function(dev);
 662         dev->rx_pkt_burst = mlx5_select_rx_function(dev);
 663         ret = mlx5_traffic_enable(dev);
 664         if (ret) {
 665                 DRV_LOG(ERR,
 666                         "port %u error occurred while configuring control"
 667                         " flows: %s",
 668                         dev->data->port_id, strerror(rte_errno));
 669                 return;
 670         }
 671         ret = mlx5_flow_start(dev, &priv->flows);
 672         if (ret)
 673                 DRV_LOG(ERR,
 674                         "port %u error occurred while configuring flows: %s",
 675                         dev->data->port_id, strerror(rte_errno));
 676 }
 677
 678 /**
 679  * Disable receiving and transmitting traffic.
 680  *
 681  * @param dev
 682  *   Pointer to Ethernet device.
 683  */
 684 static void
 685 mlx5_link_stop(struct rte_eth_dev *dev)
 686 {
 687         struct priv *priv = dev->data->dev_private;
 688
 689         mlx5_flow_stop(dev, &priv->flows);
 690         mlx5_traffic_disable(dev);
 691         dev->rx_pkt_burst = removed_rx_burst;
 692         dev->tx_pkt_burst = removed_tx_burst;
 693 }
 694
 695 /**
 696  * Querying the link status till it changes to the desired state.
 697  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
 698  *
 699  * @param dev
 700  *   Pointer to Ethernet device.
 701  * @param status
 702  *   Link desired status.
 703  *
 704  * @return
 705  *   0 on success, a negative errno value otherwise and rte_errno is set.
 706  */
 707 int
 708 mlx5_force_link_status_change(struct rte_eth_dev *dev, int status)
 709 {
 710         int try = 0;
 711
 712         while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
 713                 mlx5_link_update(dev, 0);
 714                 if (dev->data->dev_link.link_status == status)
 715                         return 0;
 716                 try++;
 717                 sleep(1);
 718         }
 719         rte_errno = EAGAIN;
 720         return -rte_errno;
 721 }
 722
 723 /**
 724  * DPDK callback to retrieve physical link information.
 725  *
 726  * @param dev
 727  *   Pointer to Ethernet device structure.
 728  * @param wait_to_complete
 729  *   Wait for request completion (ignored).
 730  *
 731  * @return
 732  *   0 on success, a negative errno value otherwise and rte_errno is set.
 733  */
 734 int
 735 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete __rte_unused)
 736 {
 737         struct utsname utsname;
 738         int ver[3];
 739         int ret;
 740         struct rte_eth_link dev_link = dev->data->dev_link;
 741
 742         if (uname(&utsname) == -1 ||
 743             sscanf(utsname.release, "%d.%d.%d",
 744                    &ver[0], &ver[1], &ver[2]) != 3 ||
 745             KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
 746                 ret = mlx5_link_update_unlocked_gset(dev);
 747         else
 748                 ret = mlx5_link_update_unlocked_gs(dev);
 749         if (ret)
 750                 return ret;
 751         /* If lsc interrupt is disabled, should always be ready for traffic. */
 752         if (!dev->data->dev_conf.intr_conf.lsc) {
 753                 mlx5_link_start(dev);
 754                 return 0;
 755         }
 756         /* Re-select burst callbacks only if link status has been changed. */
 757         if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
 758                 if (dev->data->dev_link.link_status == ETH_LINK_UP)
 759                         mlx5_link_start(dev);
 760                 else
 761                         mlx5_link_stop(dev);
 762         }
 763         return 0;
 764 }
 765
 766 /**
 767  * DPDK callback to change the MTU.
 768  *
 769  * @param dev
 770  *   Pointer to Ethernet device structure.
 771  * @param in_mtu
 772  *   New MTU.
 773  *
 774  * @return
 775  *   0 on success, a negative errno value otherwise and rte_errno is set.
 776  */
 777 int
 778 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 779 {
 780         struct priv *priv = dev->data->dev_private;
 781         uint16_t kern_mtu = 0;
 782         int ret;
 783
 784         ret = mlx5_get_mtu(dev, &kern_mtu);
 785         if (ret)
 786                 return ret;
 787         /* Set kernel interface MTU first. */
 788         ret = mlx5_set_mtu(dev, mtu);
 789         if (ret)
 790                 return ret;
 791         ret = mlx5_get_mtu(dev, &kern_mtu);
 792         if (ret)
 793                 return ret;
 794         if (kern_mtu == mtu) {
 795                 priv->mtu = mtu;
 796                 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
 797                         dev->data->port_id, mtu);
 798                 return 0;
 799         }
 800         rte_errno = EAGAIN;
 801         return -rte_errno;
 802 }
 803
 804 /**
 805  * DPDK callback to get flow control status.
 806  *
 807  * @param dev
 808  *   Pointer to Ethernet device structure.
 809  * @param[out] fc_conf
 810  *   Flow control output buffer.
 811  *
 812  * @return
 813  *   0 on success, a negative errno value otherwise and rte_errno is set.
 814  */
 815 int
 816 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 817 {
 818         struct ifreq ifr;
 819         struct ethtool_pauseparam ethpause = {
 820                 .cmd = ETHTOOL_GPAUSEPARAM
 821         };
 822         int ret;
 823
 824         ifr.ifr_data = (void *)&ethpause;
 825         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 826         if (ret) {
 827                 DRV_LOG(WARNING,
 828                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
 829                         " %s",
 830                         dev->data->port_id, strerror(rte_errno));
 831                 return ret;
 832         }
 833         fc_conf->autoneg = ethpause.autoneg;
 834         if (ethpause.rx_pause && ethpause.tx_pause)
 835                 fc_conf->mode = RTE_FC_FULL;
 836         else if (ethpause.rx_pause)
 837                 fc_conf->mode = RTE_FC_RX_PAUSE;
 838         else if (ethpause.tx_pause)
 839                 fc_conf->mode = RTE_FC_TX_PAUSE;
 840         else
 841                 fc_conf->mode = RTE_FC_NONE;
 842         return 0;
 843 }
 844
 845 /**
 846  * DPDK callback to modify flow control parameters.
 847  *
 848  * @param dev
 849  *   Pointer to Ethernet device structure.
 850  * @param[in] fc_conf
 851  *   Flow control parameters.
 852  *
 853  * @return
 854  *   0 on success, a negative errno value otherwise and rte_errno is set.
 855  */
 856 int
 857 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 858 {
 859         struct ifreq ifr;
 860         struct ethtool_pauseparam ethpause = {
 861                 .cmd = ETHTOOL_SPAUSEPARAM
 862         };
 863         int ret;
 864
 865         ifr.ifr_data = (void *)&ethpause;
 866         ethpause.autoneg = fc_conf->autoneg;
 867         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 868             (fc_conf->mode & RTE_FC_RX_PAUSE))
 869                 ethpause.rx_pause = 1;
 870         else
 871                 ethpause.rx_pause = 0;
 872
 873         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 874             (fc_conf->mode & RTE_FC_TX_PAUSE))
 875                 ethpause.tx_pause = 1;
 876         else
 877                 ethpause.tx_pause = 0;
 878         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 879         if (ret) {
 880                 DRV_LOG(WARNING,
 881                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
 882                         " failed: %s",
 883                         dev->data->port_id, strerror(rte_errno));
 884                 return ret;
 885         }
 886         return 0;
 887 }
 888
 889 /**
 890  * Get PCI information from struct ibv_device.
 891  *
 892  * @param device
 893  *   Pointer to Ethernet device structure.
 894  * @param[out] pci_addr
 895  *   PCI bus address output buffer.
 896  *
 897  * @return
 898  *   0 on success, a negative errno value otherwise and rte_errno is set.
 899  */
 900 int
 901 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
 902                             struct rte_pci_addr *pci_addr)
 903 {
 904         FILE *file;
 905         char line[32];
 906         MKSTR(path, "%s/device/uevent", device->ibdev_path);
 907
 908         file = fopen(path, "rb");
 909         if (file == NULL) {
 910                 rte_errno = errno;
 911                 return -rte_errno;
 912         }
 913         while (fgets(line, sizeof(line), file) == line) {
 914                 size_t len = strlen(line);
 915                 int ret;
 916
 917                 /* Truncate long lines. */
 918                 if (len == (sizeof(line) - 1))
 919                         while (line[(len - 1)] != '\n') {
 920                                 ret = fgetc(file);
 921                                 if (ret == EOF)
 922                                         break;
 923                                 line[(len - 1)] = ret;
 924                         }
 925                 /* Extract information. */
 926                 if (sscanf(line,
 927                            "PCI_SLOT_NAME="
 928                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
 929                            &pci_addr->domain,
 930                            &pci_addr->bus,
 931                            &pci_addr->devid,
 932                            &pci_addr->function) == 4) {
 933                         ret = 0;
 934                         break;
 935                 }
 936         }
 937         fclose(file);
 938         return 0;
 939 }
 940
 941 /**
 942  * Update the link status.
 943  *
 944  * @param dev
 945  *   Pointer to Ethernet device.
 946  *
 947  * @return
 948  *   Zero if the callback process can be called immediately, negative errno
 949  *   value otherwise and rte_errno is set.
 950  */
 951 static int
 952 mlx5_link_status_update(struct rte_eth_dev *dev)
 953 {
 954         struct priv *priv = dev->data->dev_private;
 955         struct rte_eth_link *link = &dev->data->dev_link;
 956         int ret;
 957
 958         ret = mlx5_link_update(dev, 0);
 959         if (ret)
 960                 return ret;
 961         if (((link->link_speed == 0) && link->link_status) ||
 962                 ((link->link_speed != 0) && !link->link_status)) {
 963                 /*
 964                  * Inconsistent status. Event likely occurred before the
 965                  * kernel netdevice exposes the new status.
 966                  */
 967                 if (!priv->pending_alarm) {
 968                         priv->pending_alarm = 1;
 969                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
 970                                           mlx5_dev_link_status_handler,
 971                                           priv->dev);
 972                 }
 973                 return 1;
 974         } else if (unlikely(priv->pending_alarm)) {
 975                 /* Link interrupt occurred while alarm is already scheduled. */
 976                 priv->pending_alarm = 0;
 977                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
 978         }
 979         return 0;
 980 }
 981
 982 /**
 983  * Device status handler.
 984  *
 985  * @param dev
 986  *   Pointer to Ethernet device.
 987  * @param events
 988  *   Pointer to event flags holder.
 989  *
 990  * @return
 991  *   Events bitmap of callback process which can be called immediately.
 992  */
 993 static uint32_t
 994 mlx5_dev_status_handler(struct rte_eth_dev *dev)
 995 {
 996         struct priv *priv = dev->data->dev_private;
 997         struct ibv_async_event event;
 998         uint32_t ret = 0;
 999
1000         /* Read all message and acknowledge them. */
1001         for (;;) {
1002                 if (mlx5_glue->get_async_event(priv->ctx, &event))
1003                         break;
1004                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1005                         event.event_type == IBV_EVENT_PORT_ERR) &&
1006                         (dev->data->dev_conf.intr_conf.lsc == 1))
1007                         ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1008                 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1009                         dev->data->dev_conf.intr_conf.rmv == 1)
1010                         ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1011                 else
1012                         DRV_LOG(DEBUG,
1013                                 "port %u event type %d on not handled",
1014                                 dev->data->port_id, event.event_type);
1015                 mlx5_glue->ack_async_event(&event);
1016         }
1017         if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1018                 if (mlx5_link_status_update(dev))
1019                         ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1020         return ret;
1021 }
1022
1023 /**
1024  * Handle delayed link status event.
1025  *
1026  * @param arg
1027  *   Registered argument.
1028  */
1029 void
1030 mlx5_dev_link_status_handler(void *arg)
1031 {
1032         struct rte_eth_dev *dev = arg;
1033         struct priv *priv = dev->data->dev_private;
1034         int ret;
1035
1036         priv->pending_alarm = 0;
1037         ret = mlx5_link_status_update(dev);
1038         if (!ret)
1039                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1040 }
1041
1042 /**
1043  * Handle interrupts from the NIC.
1044  *
1045  * @param[in] intr_handle
1046  *   Interrupt handler.
1047  * @param cb_arg
1048  *   Callback argument.
1049  */
1050 void
1051 mlx5_dev_interrupt_handler(void *cb_arg)
1052 {
1053         struct rte_eth_dev *dev = cb_arg;
1054         uint32_t events;
1055
1056         events = mlx5_dev_status_handler(dev);
1057         if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1058                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1059         if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1060                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1061 }
1062
1063 /**
1064  * Handle interrupts from the socket.
1065  *
1066  * @param cb_arg
1067  *   Callback argument.
1068  */
1069 static void
1070 mlx5_dev_handler_socket(void *cb_arg)
1071 {
1072         struct rte_eth_dev *dev = cb_arg;
1073
1074         mlx5_socket_handle(dev);
1075 }
1076
1077 /**
1078  * Uninstall interrupt handler.
1079  *
1080  * @param dev
1081  *   Pointer to Ethernet device.
1082  */
1083 void
1084 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1085 {
1086         struct priv *priv = dev->data->dev_private;
1087
1088         if (dev->data->dev_conf.intr_conf.lsc ||
1089             dev->data->dev_conf.intr_conf.rmv)
1090                 rte_intr_callback_unregister(&priv->intr_handle,
1091                                              mlx5_dev_interrupt_handler, dev);
1092         if (priv->primary_socket)
1093                 rte_intr_callback_unregister(&priv->intr_handle_socket,
1094                                              mlx5_dev_handler_socket, dev);
1095         if (priv->pending_alarm) {
1096                 priv->pending_alarm = 0;
1097                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1098         }
1099         priv->intr_handle.fd = 0;
1100         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1101         priv->intr_handle_socket.fd = 0;
1102         priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1103 }
1104
1105 /**
1106  * Install interrupt handler.
1107  *
1108  * @param dev
1109  *   Pointer to Ethernet device.
1110  */
1111 void
1112 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1113 {
1114         struct priv *priv = dev->data->dev_private;
1115         int ret;
1116         int flags;
1117
1118         assert(priv->ctx->async_fd > 0);
1119         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1120         ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1121         if (ret) {
1122                 DRV_LOG(INFO,
1123                         "port %u failed to change file descriptor async event"
1124                         " queue",
1125                         dev->data->port_id);
1126                 dev->data->dev_conf.intr_conf.lsc = 0;
1127                 dev->data->dev_conf.intr_conf.rmv = 0;
1128         }
1129         if (dev->data->dev_conf.intr_conf.lsc ||
1130             dev->data->dev_conf.intr_conf.rmv) {
1131                 priv->intr_handle.fd = priv->ctx->async_fd;
1132                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1133                 rte_intr_callback_register(&priv->intr_handle,
1134                                            mlx5_dev_interrupt_handler, dev);
1135         }
1136         ret = mlx5_socket_init(dev);
1137         if (ret)
1138                 DRV_LOG(ERR, "port %u cannot initialise socket: %s",
1139                         dev->data->port_id, strerror(rte_errno));
1140         else if (priv->primary_socket) {
1141                 priv->intr_handle_socket.fd = priv->primary_socket;
1142                 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1143                 rte_intr_callback_register(&priv->intr_handle_socket,
1144                                            mlx5_dev_handler_socket, dev);
1145         }
1146 }
1147
1148 /**
1149  * DPDK callback to bring the link DOWN.
1150  *
1151  * @param dev
1152  *   Pointer to Ethernet device structure.
1153  *
1154  * @return
1155  *   0 on success, a negative errno value otherwise and rte_errno is set.
1156  */
1157 int
1158 mlx5_set_link_down(struct rte_eth_dev *dev)
1159 {
1160         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1161 }
1162
1163 /**
1164  * DPDK callback to bring the link UP.
1165  *
1166  * @param dev
1167  *   Pointer to Ethernet device structure.
1168  *
1169  * @return
1170  *   0 on success, a negative errno value otherwise and rte_errno is set.
1171  */
1172 int
1173 mlx5_set_link_up(struct rte_eth_dev *dev)
1174 {
1175         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1176 }
1177
1178 /**
1179  * Configure the TX function to use.
1180  *
1181  * @param dev
1182  *   Pointer to private data structure.
1183  *
1184  * @return
1185  *   Pointer to selected Tx burst function.
1186  */
1187 eth_tx_burst_t
1188 mlx5_select_tx_function(struct rte_eth_dev *dev)
1189 {
1190         struct priv *priv = dev->data->dev_private;
1191         eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1192         struct mlx5_dev_config *config = &priv->config;
1193         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1194         int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1195                                     DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1196                                     DEV_TX_OFFLOAD_GRE_TNL_TSO));
1197         int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1198
1199         assert(priv != NULL);
1200         /* Select appropriate TX function. */
1201         if (vlan_insert || tso)
1202                 return tx_pkt_burst;
1203         if (config->mps == MLX5_MPW_ENHANCED) {
1204                 if (mlx5_check_vec_tx_support(dev) > 0) {
1205                         if (mlx5_check_raw_vec_tx_support(dev) > 0)
1206                                 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1207                         else
1208                                 tx_pkt_burst = mlx5_tx_burst_vec;
1209                         DRV_LOG(DEBUG,
1210                                 "port %u selected enhanced MPW Tx vectorized"
1211                                 " function",
1212                                 dev->data->port_id);
1213                 } else {
1214                         tx_pkt_burst = mlx5_tx_burst_empw;
1215                         DRV_LOG(DEBUG,
1216                                 "port %u selected enhanced MPW Tx function",
1217                                 dev->data->port_id);
1218                 }
1219         } else if (config->mps && (config->txq_inline > 0)) {
1220                 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1221                 DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1222                         dev->data->port_id);
1223         } else if (config->mps) {
1224                 tx_pkt_burst = mlx5_tx_burst_mpw;
1225                 DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1226                         dev->data->port_id);
1227         }
1228         return tx_pkt_burst;
1229 }
1230
1231 /**
1232  * Configure the RX function to use.
1233  *
1234  * @param dev
1235  *   Pointer to private data structure.
1236  *
1237  * @return
1238  *   Pointer to selected Rx burst function.
1239  */
1240 eth_rx_burst_t
1241 mlx5_select_rx_function(struct rte_eth_dev *dev)
1242 {
1243         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1244
1245         assert(dev != NULL);
1246         if (mlx5_check_vec_rx_support(dev) > 0) {
1247                 rx_pkt_burst = mlx5_rx_burst_vec;
1248                 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1249                         dev->data->port_id);
1250         }
1251         return rx_pkt_burst;
1252 }
1253
1254 /**
1255  * Check if mlx5 device was removed.
1256  *
1257  * @param dev
1258  *   Pointer to Ethernet device structure.
1259  *
1260  * @return
1261  *   1 when device is removed, otherwise 0.
1262  */
1263 int
1264 mlx5_is_removed(struct rte_eth_dev *dev)
1265 {
1266         struct ibv_device_attr device_attr;
1267         struct priv *priv = dev->data->dev_private;
1268
1269         if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1270                 return 1;
1271         return 0;
1272 }