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] priv
  99  *   Pointer to private structure.
 100  * @param[out] ifname
 101  *   Interface name output buffer.
 102  *
 103  * @return
 104  *   0 on success, -1 on failure and errno is set.
 105  */
 106 int
 107 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
 108 {
 109         DIR *dir;
 110         struct dirent *dent;
 111         unsigned int dev_type = 0;
 112         unsigned int dev_port_prev = ~0u;
 113         char match[IF_NAMESIZE] = "";
 114
 115         {
 116                 MKSTR(path, "%s/device/net", priv->ibdev_path);
 117
 118                 dir = opendir(path);
 119                 if (dir == NULL)
 120                         return -1;
 121         }
 122         while ((dent = readdir(dir)) != NULL) {
 123                 char *name = dent->d_name;
 124                 FILE *file;
 125                 unsigned int dev_port;
 126                 int r;
 127
 128                 if ((name[0] == '.') &&
 129                     ((name[1] == '\0') ||
 130                      ((name[1] == '.') && (name[2] == '\0'))))
 131                         continue;
 132
 133                 MKSTR(path, "%s/device/net/%s/%s",
 134                       priv->ibdev_path, name,
 135                       (dev_type ? "dev_id" : "dev_port"));
 136
 137                 file = fopen(path, "rb");
 138                 if (file == NULL) {
 139                         if (errno != ENOENT)
 140                                 continue;
 141                         /*
 142                          * Switch to dev_id when dev_port does not exist as
 143                          * is the case with Linux kernel versions < 3.15.
 144                          */
 145 try_dev_id:
 146                         match[0] = '\0';
 147                         if (dev_type)
 148                                 break;
 149                         dev_type = 1;
 150                         dev_port_prev = ~0u;
 151                         rewinddir(dir);
 152                         continue;
 153                 }
 154                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 155                 fclose(file);
 156                 if (r != 1)
 157                         continue;
 158                 /*
 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.
 162                  */
 163                 if (dev_port == dev_port_prev)
 164                         goto try_dev_id;
 165                 dev_port_prev = dev_port;
 166                 if (dev_port == (priv->port - 1u))
 167                         snprintf(match, sizeof(match), "%s", name);
 168         }
 169         closedir(dir);
 170         if (match[0] == '\0')
 171                 return -1;
 172         strncpy(*ifname, match, sizeof(*ifname));
 173         return 0;
 174 }
 175
 176 /**
 177  * Perform ifreq ioctl() on associated Ethernet device.
 178  *
 179  * @param[in] priv
 180  *   Pointer to private structure.
 181  * @param req
 182  *   Request number to pass to ioctl().
 183  * @param[out] ifr
 184  *   Interface request structure output buffer.
 185  *
 186  * @return
 187  *   0 on success, -1 on failure and errno is set.
 188  */
 189 int
 190 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
 191 {
 192         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 193         int ret = -1;
 194
 195         if (sock == -1)
 196                 return ret;
 197         if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
 198                 ret = ioctl(sock, req, ifr);
 199         close(sock);
 200         return ret;
 201 }
 202
 203 /**
 204  * Return the number of active VFs for the current device.
 205  *
 206  * @param[in] priv
 207  *   Pointer to private structure.
 208  * @param[out] num_vfs
 209  *   Number of active VFs.
 210  *
 211  * @return
 212  *   0 on success, -1 on failure and errno is set.
 213  */
 214 int
 215 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
 216 {
 217         /* The sysfs entry name depends on the operating system. */
 218         const char **name = (const char *[]){
 219                 "sriov_numvfs",
 220                 "mlx5_num_vfs",
 221                 NULL,
 222         };
 223
 224         do {
 225                 int n;
 226                 FILE *file;
 227                 MKSTR(path, "%s/device/%s", priv->ibdev_path, *name);
 228
 229                 file = fopen(path, "rb");
 230                 if (!file)
 231                         continue;
 232                 n = fscanf(file, "%" SCNu16, num_vfs);
 233                 fclose(file);
 234                 if (n == 1)
 235                         return 0;
 236         } while (*(++name));
 237         return -1;
 238 }
 239
 240 /**
 241  * Get device MTU.
 242  *
 243  * @param priv
 244  *   Pointer to private structure.
 245  * @param[out] mtu
 246  *   MTU value output buffer.
 247  *
 248  * @return
 249  *   0 on success, -1 on failure and errno is set.
 250  */
 251 int
 252 priv_get_mtu(struct priv *priv, uint16_t *mtu)
 253 {
 254         struct ifreq request;
 255         int ret = priv_ifreq(priv, SIOCGIFMTU, &request);
 256
 257         if (ret)
 258                 return ret;
 259         *mtu = request.ifr_mtu;
 260         return 0;
 261 }
 262
 263 /**
 264  * Set device MTU.
 265  *
 266  * @param priv
 267  *   Pointer to private structure.
 268  * @param mtu
 269  *   MTU value to set.
 270  *
 271  * @return
 272  *   0 on success, -1 on failure and errno is set.
 273  */
 274 static int
 275 priv_set_mtu(struct priv *priv, uint16_t mtu)
 276 {
 277         struct ifreq request = { .ifr_mtu = mtu, };
 278
 279         return priv_ifreq(priv, SIOCSIFMTU, &request);
 280 }
 281
 282 /**
 283  * Set device flags.
 284  *
 285  * @param priv
 286  *   Pointer to private structure.
 287  * @param keep
 288  *   Bitmask for flags that must remain untouched.
 289  * @param flags
 290  *   Bitmask for flags to modify.
 291  *
 292  * @return
 293  *   0 on success, -1 on failure and errno is set.
 294  */
 295 int
 296 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
 297 {
 298         struct ifreq request;
 299         int ret = priv_ifreq(priv, SIOCGIFFLAGS, &request);
 300
 301         if (ret)
 302                 return ret;
 303         request.ifr_flags &= keep;
 304         request.ifr_flags |= flags & ~keep;
 305         return priv_ifreq(priv, SIOCSIFFLAGS, &request);
 306 }
 307
 308 /**
 309  * Ethernet device configuration.
 310  *
 311  * Prepare the driver for a given number of TX and RX queues.
 312  *
 313  * @param dev
 314  *   Pointer to Ethernet device structure.
 315  *
 316  * @return
 317  *   0 on success, errno value on failure.
 318  */
 319 static int
 320 dev_configure(struct rte_eth_dev *dev)
 321 {
 322         struct priv *priv = dev->data->dev_private;
 323         unsigned int rxqs_n = dev->data->nb_rx_queues;
 324         unsigned int txqs_n = dev->data->nb_tx_queues;
 325         unsigned int i;
 326         unsigned int j;
 327         unsigned int reta_idx_n;
 328         const uint8_t use_app_rss_key =
 329                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 330         uint64_t supp_tx_offloads = mlx5_priv_get_tx_port_offloads(priv);
 331         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
 332         uint64_t supp_rx_offloads =
 333                 (mlx5_priv_get_rx_port_offloads(priv) |
 334                  mlx5_priv_get_rx_queue_offloads(priv));
 335         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
 336
 337         if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
 338                 ERROR("Some Tx offloads are not supported "
 339                       "requested 0x%" PRIx64 " supported 0x%" PRIx64,
 340                       tx_offloads, supp_tx_offloads);
 341                 return ENOTSUP;
 342         }
 343         if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
 344                 ERROR("Some Rx offloads are not supported "
 345                       "requested 0x%" PRIx64 " supported 0x%" PRIx64,
 346                       rx_offloads, supp_rx_offloads);
 347                 return ENOTSUP;
 348         }
 349         if (use_app_rss_key &&
 350             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 351              rss_hash_default_key_len)) {
 352                 /* MLX5 RSS only support 40bytes key. */
 353                 return EINVAL;
 354         }
 355         priv->rss_conf.rss_key =
 356                 rte_realloc(priv->rss_conf.rss_key,
 357                             rss_hash_default_key_len, 0);
 358         if (!priv->rss_conf.rss_key) {
 359                 ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
 360                 return ENOMEM;
 361         }
 362         memcpy(priv->rss_conf.rss_key,
 363                use_app_rss_key ?
 364                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 365                rss_hash_default_key,
 366                rss_hash_default_key_len);
 367         priv->rss_conf.rss_key_len = rss_hash_default_key_len;
 368         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 369         priv->rxqs = (void *)dev->data->rx_queues;
 370         priv->txqs = (void *)dev->data->tx_queues;
 371         if (txqs_n != priv->txqs_n) {
 372                 INFO("%p: TX queues number update: %u -> %u",
 373                      (void *)dev, priv->txqs_n, txqs_n);
 374                 priv->txqs_n = txqs_n;
 375         }
 376         if (rxqs_n > priv->config.ind_table_max_size) {
 377                 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
 378                 return EINVAL;
 379         }
 380         if (rxqs_n == priv->rxqs_n)
 381                 return 0;
 382         INFO("%p: RX queues number update: %u -> %u",
 383              (void *)dev, priv->rxqs_n, rxqs_n);
 384         priv->rxqs_n = rxqs_n;
 385         /* If the requested number of RX queues is not a power of two, use the
 386          * maximum indirection table size for better balancing.
 387          * The result is always rounded to the next power of two. */
 388         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 389                                      priv->config.ind_table_max_size :
 390                                      rxqs_n));
 391         if (priv_rss_reta_index_resize(priv, reta_idx_n))
 392                 return ENOMEM;
 393         /* When the number of RX queues is not a power of two, the remaining
 394          * table entries are padded with reused WQs and hashes are not spread
 395          * uniformly. */
 396         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 397                 (*priv->reta_idx)[i] = j;
 398                 if (++j == rxqs_n)
 399                         j = 0;
 400         }
 401         return 0;
 402 }
 403
 404 /**
 405  * DPDK callback for Ethernet device configuration.
 406  *
 407  * @param dev
 408  *   Pointer to Ethernet device structure.
 409  *
 410  * @return
 411  *   0 on success, negative errno value on failure.
 412  */
 413 int
 414 mlx5_dev_configure(struct rte_eth_dev *dev)
 415 {
 416         struct priv *priv = dev->data->dev_private;
 417         int ret;
 418
 419         priv_lock(priv);
 420         ret = dev_configure(dev);
 421         assert(ret >= 0);
 422         priv_unlock(priv);
 423         return -ret;
 424 }
 425
 426 /**
 427  * DPDK callback to get information about the device.
 428  *
 429  * @param dev
 430  *   Pointer to Ethernet device structure.
 431  * @param[out] info
 432  *   Info structure output buffer.
 433  */
 434 void
 435 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 436 {
 437         struct priv *priv = dev->data->dev_private;
 438         struct mlx5_dev_config *config = &priv->config;
 439         unsigned int max;
 440         char ifname[IF_NAMESIZE];
 441
 442         info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
 443
 444         priv_lock(priv);
 445         /* FIXME: we should ask the device for these values. */
 446         info->min_rx_bufsize = 32;
 447         info->max_rx_pktlen = 65536;
 448         /*
 449          * Since we need one CQ per QP, the limit is the minimum number
 450          * between the two values.
 451          */
 452         max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
 453                       priv->device_attr.orig_attr.max_qp);
 454         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 455         if (max >= 65535)
 456                 max = 65535;
 457         info->max_rx_queues = max;
 458         info->max_tx_queues = max;
 459         info->max_mac_addrs = RTE_DIM(priv->mac);
 460         info->rx_queue_offload_capa =
 461                 mlx5_priv_get_rx_queue_offloads(priv);
 462         info->rx_offload_capa = (mlx5_priv_get_rx_port_offloads(priv) |
 463                                  info->rx_queue_offload_capa);
 464         info->tx_offload_capa = mlx5_priv_get_tx_port_offloads(priv);
 465         if (priv_get_ifname(priv, &ifname) == 0)
 466                 info->if_index = if_nametoindex(ifname);
 467         info->reta_size = priv->reta_idx_n ?
 468                 priv->reta_idx_n : config->ind_table_max_size;
 469         info->hash_key_size = priv->rss_conf.rss_key_len;
 470         info->speed_capa = priv->link_speed_capa;
 471         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
 472         priv_unlock(priv);
 473 }
 474
 475 const uint32_t *
 476 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 477 {
 478         static const uint32_t ptypes[] = {
 479                 /* refers to rxq_cq_to_pkt_type() */
 480                 RTE_PTYPE_L2_ETHER,
 481                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 482                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 483                 RTE_PTYPE_L4_NONFRAG,
 484                 RTE_PTYPE_L4_FRAG,
 485                 RTE_PTYPE_L4_TCP,
 486                 RTE_PTYPE_L4_UDP,
 487                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 488                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 489                 RTE_PTYPE_INNER_L4_NONFRAG,
 490                 RTE_PTYPE_INNER_L4_FRAG,
 491                 RTE_PTYPE_INNER_L4_TCP,
 492                 RTE_PTYPE_INNER_L4_UDP,
 493                 RTE_PTYPE_UNKNOWN
 494         };
 495
 496         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 497             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 498                 return ptypes;
 499         return NULL;
 500 }
 501
 502 /**
 503  * DPDK callback to retrieve physical link information.
 504  *
 505  * @param dev
 506  *   Pointer to Ethernet device structure.
 507  * @param wait_to_complete
 508  *   Wait for request completion (ignored).
 509  */
 510 static int
 511 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
 512 {
 513         struct priv *priv = dev->data->dev_private;
 514         struct ethtool_cmd edata = {
 515                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 516         };
 517         struct ifreq ifr;
 518         struct rte_eth_link dev_link;
 519         int link_speed = 0;
 520
 521         /* priv_lock() is not taken to allow concurrent calls. */
 522
 523         (void)wait_to_complete;
 524         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 525                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 526                 return -1;
 527         }
 528         memset(&dev_link, 0, sizeof(dev_link));
 529         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 530                                 (ifr.ifr_flags & IFF_RUNNING));
 531         ifr.ifr_data = (void *)&edata;
 532         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 533                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 534                      strerror(errno));
 535                 return -1;
 536         }
 537         link_speed = ethtool_cmd_speed(&edata);
 538         if (link_speed == -1)
 539                 dev_link.link_speed = 0;
 540         else
 541                 dev_link.link_speed = link_speed;
 542         priv->link_speed_capa = 0;
 543         if (edata.supported & SUPPORTED_Autoneg)
 544                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 545         if (edata.supported & (SUPPORTED_1000baseT_Full |
 546                                SUPPORTED_1000baseKX_Full))
 547                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 548         if (edata.supported & SUPPORTED_10000baseKR_Full)
 549                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 550         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 551                                SUPPORTED_40000baseCR4_Full |
 552                                SUPPORTED_40000baseSR4_Full |
 553                                SUPPORTED_40000baseLR4_Full))
 554                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 555         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 556                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 557         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 558                         ETH_LINK_SPEED_FIXED);
 559         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 560                 /* Link status changed. */
 561                 dev->data->dev_link = dev_link;
 562                 return 0;
 563         }
 564         /* Link status is still the same. */
 565         return -1;
 566 }
 567
 568 /**
 569  * Retrieve physical link information (unlocked version using new ioctl).
 570  *
 571  * @param dev
 572  *   Pointer to Ethernet device structure.
 573  * @param wait_to_complete
 574  *   Wait for request completion (ignored).
 575  */
 576 static int
 577 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
 578 {
 579         struct priv *priv = dev->data->dev_private;
 580         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 581         struct ifreq ifr;
 582         struct rte_eth_link dev_link;
 583         uint64_t sc;
 584
 585         (void)wait_to_complete;
 586         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 587                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 588                 return -1;
 589         }
 590         memset(&dev_link, 0, sizeof(dev_link));
 591         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 592                                 (ifr.ifr_flags & IFF_RUNNING));
 593         ifr.ifr_data = (void *)&gcmd;
 594         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 595                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 596                       strerror(errno));
 597                 return -1;
 598         }
 599         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 600
 601         alignas(struct ethtool_link_settings)
 602         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 603                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 604         struct ethtool_link_settings *ecmd = (void *)data;
 605
 606         *ecmd = gcmd;
 607         ifr.ifr_data = (void *)ecmd;
 608         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 609                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 610                       strerror(errno));
 611                 return -1;
 612         }
 613         dev_link.link_speed = ecmd->speed;
 614         sc = ecmd->link_mode_masks[0] |
 615                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 616         priv->link_speed_capa = 0;
 617         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 618                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 619         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 620                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
 621                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 622         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
 623                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
 624                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
 625                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 626         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
 627                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
 628                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 629         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
 630                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
 631                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
 632                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
 633                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 634         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
 635                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
 636                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
 637                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
 638                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 639         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
 640                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
 641                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
 642                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 643         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
 644                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
 645                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 646         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
 647                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
 648                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
 649                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
 650                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 651         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
 652                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 653         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 654                                   ETH_LINK_SPEED_FIXED);
 655         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 656                 /* Link status changed. */
 657                 dev->data->dev_link = dev_link;
 658                 return 0;
 659         }
 660         /* Link status is still the same. */
 661         return -1;
 662 }
 663
 664 /**
 665  * Enable receiving and transmitting traffic.
 666  *
 667  * @param priv
 668  *   Pointer to private structure.
 669  */
 670 static void
 671 priv_link_start(struct priv *priv)
 672 {
 673         struct rte_eth_dev *dev = priv->dev;
 674         int err;
 675
 676         dev->tx_pkt_burst = priv_select_tx_function(priv, dev);
 677         dev->rx_pkt_burst = priv_select_rx_function(priv, dev);
 678         err = priv_dev_traffic_enable(priv, dev);
 679         if (err)
 680                 ERROR("%p: error occurred while configuring control flows: %s",
 681                       (void *)priv, strerror(err));
 682         err = priv_flow_start(priv, &priv->flows);
 683         if (err)
 684                 ERROR("%p: error occurred while configuring flows: %s",
 685                       (void *)priv, strerror(err));
 686 }
 687
 688 /**
 689  * Disable receiving and transmitting traffic.
 690  *
 691  * @param priv
 692  *   Pointer to private structure.
 693  */
 694 static void
 695 priv_link_stop(struct priv *priv)
 696 {
 697         struct rte_eth_dev *dev = priv->dev;
 698
 699         priv_flow_stop(priv, &priv->flows);
 700         priv_dev_traffic_disable(priv, dev);
 701         dev->rx_pkt_burst = removed_rx_burst;
 702         dev->tx_pkt_burst = removed_tx_burst;
 703 }
 704
 705 /**
 706  * Retrieve physical link information and update rx/tx_pkt_burst callbacks
 707  * accordingly.
 708  *
 709  * @param priv
 710  *   Pointer to private structure.
 711  * @param wait_to_complete
 712  *   Wait for request completion (ignored).
 713  */
 714 int
 715 priv_link_update(struct priv *priv, int wait_to_complete)
 716 {
 717         struct rte_eth_dev *dev = priv->dev;
 718         struct utsname utsname;
 719         int ver[3];
 720         int ret;
 721         struct rte_eth_link dev_link = dev->data->dev_link;
 722
 723         if (uname(&utsname) == -1 ||
 724             sscanf(utsname.release, "%d.%d.%d",
 725                    &ver[0], &ver[1], &ver[2]) != 3 ||
 726             KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
 727                 ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
 728         else
 729                 ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
 730         /* If lsc interrupt is disabled, should always be ready for traffic. */
 731         if (!dev->data->dev_conf.intr_conf.lsc) {
 732                 priv_link_start(priv);
 733                 return ret;
 734         }
 735         /* Re-select burst callbacks only if link status has been changed. */
 736         if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
 737                 if (dev->data->dev_link.link_status == ETH_LINK_UP)
 738                         priv_link_start(priv);
 739                 else
 740                         priv_link_stop(priv);
 741         }
 742         return ret;
 743 }
 744
 745 /**
 746  * Querying the link status till it changes to the desired state.
 747  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
 748  *
 749  * @param priv
 750  *   Pointer to private structure.
 751  * @param status
 752  *   Link desired status.
 753  *
 754  * @return
 755  *   0 on success, negative errno value on failure.
 756  */
 757 int
 758 priv_force_link_status_change(struct priv *priv, int status)
 759 {
 760         int try = 0;
 761
 762         while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
 763                 priv_link_update(priv, 0);
 764                 if (priv->dev->data->dev_link.link_status == status)
 765                         return 0;
 766                 try++;
 767                 sleep(1);
 768         }
 769         return -EAGAIN;
 770 }
 771
 772 /**
 773  * DPDK callback to retrieve physical link information.
 774  *
 775  * @param dev
 776  *   Pointer to Ethernet device structure.
 777  * @param wait_to_complete
 778  *   Wait for request completion (ignored).
 779  */
 780 int
 781 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 782 {
 783         struct priv *priv = dev->data->dev_private;
 784         int ret;
 785
 786         priv_lock(priv);
 787         ret = priv_link_update(priv, wait_to_complete);
 788         priv_unlock(priv);
 789         return ret;
 790 }
 791
 792 /**
 793  * DPDK callback to change the MTU.
 794  *
 795  * @param dev
 796  *   Pointer to Ethernet device structure.
 797  * @param in_mtu
 798  *   New MTU.
 799  *
 800  * @return
 801  *   0 on success, negative errno value on failure.
 802  */
 803 int
 804 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 805 {
 806         struct priv *priv = dev->data->dev_private;
 807         uint16_t kern_mtu;
 808         int ret = 0;
 809
 810         priv_lock(priv);
 811         ret = priv_get_mtu(priv, &kern_mtu);
 812         if (ret)
 813                 goto out;
 814         /* Set kernel interface MTU first. */
 815         ret = priv_set_mtu(priv, mtu);
 816         if (ret)
 817                 goto out;
 818         ret = priv_get_mtu(priv, &kern_mtu);
 819         if (ret)
 820                 goto out;
 821         if (kern_mtu == mtu) {
 822                 priv->mtu = mtu;
 823                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
 824         }
 825         priv_unlock(priv);
 826         return 0;
 827 out:
 828         ret = errno;
 829         WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
 830              strerror(ret));
 831         priv_unlock(priv);
 832         assert(ret >= 0);
 833         return -ret;
 834 }
 835
 836 /**
 837  * DPDK callback to get flow control status.
 838  *
 839  * @param dev
 840  *   Pointer to Ethernet device structure.
 841  * @param[out] fc_conf
 842  *   Flow control output buffer.
 843  *
 844  * @return
 845  *   0 on success, negative errno value on failure.
 846  */
 847 int
 848 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 849 {
 850         struct priv *priv = dev->data->dev_private;
 851         struct ifreq ifr;
 852         struct ethtool_pauseparam ethpause = {
 853                 .cmd = ETHTOOL_GPAUSEPARAM
 854         };
 855         int ret;
 856
 857         ifr.ifr_data = (void *)&ethpause;
 858         priv_lock(priv);
 859         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 860                 ret = errno;
 861                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
 862                      " failed: %s",
 863                      strerror(ret));
 864                 goto out;
 865         }
 866
 867         fc_conf->autoneg = ethpause.autoneg;
 868         if (ethpause.rx_pause && ethpause.tx_pause)
 869                 fc_conf->mode = RTE_FC_FULL;
 870         else if (ethpause.rx_pause)
 871                 fc_conf->mode = RTE_FC_RX_PAUSE;
 872         else if (ethpause.tx_pause)
 873                 fc_conf->mode = RTE_FC_TX_PAUSE;
 874         else
 875                 fc_conf->mode = RTE_FC_NONE;
 876         ret = 0;
 877
 878 out:
 879         priv_unlock(priv);
 880         assert(ret >= 0);
 881         return -ret;
 882 }
 883
 884 /**
 885  * DPDK callback to modify flow control parameters.
 886  *
 887  * @param dev
 888  *   Pointer to Ethernet device structure.
 889  * @param[in] fc_conf
 890  *   Flow control parameters.
 891  *
 892  * @return
 893  *   0 on success, negative errno value on failure.
 894  */
 895 int
 896 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 897 {
 898         struct priv *priv = dev->data->dev_private;
 899         struct ifreq ifr;
 900         struct ethtool_pauseparam ethpause = {
 901                 .cmd = ETHTOOL_SPAUSEPARAM
 902         };
 903         int ret;
 904
 905         ifr.ifr_data = (void *)&ethpause;
 906         ethpause.autoneg = fc_conf->autoneg;
 907         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 908             (fc_conf->mode & RTE_FC_RX_PAUSE))
 909                 ethpause.rx_pause = 1;
 910         else
 911                 ethpause.rx_pause = 0;
 912
 913         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 914             (fc_conf->mode & RTE_FC_TX_PAUSE))
 915                 ethpause.tx_pause = 1;
 916         else
 917                 ethpause.tx_pause = 0;
 918
 919         priv_lock(priv);
 920         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 921                 ret = errno;
 922                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
 923                      " failed: %s",
 924                      strerror(ret));
 925                 goto out;
 926         }
 927         ret = 0;
 928
 929 out:
 930         priv_unlock(priv);
 931         assert(ret >= 0);
 932         return -ret;
 933 }
 934
 935 /**
 936  * Get PCI information from struct ibv_device.
 937  *
 938  * @param device
 939  *   Pointer to Ethernet device structure.
 940  * @param[out] pci_addr
 941  *   PCI bus address output buffer.
 942  *
 943  * @return
 944  *   0 on success, -1 on failure and errno is set.
 945  */
 946 int
 947 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
 948                             struct rte_pci_addr *pci_addr)
 949 {
 950         FILE *file;
 951         char line[32];
 952         MKSTR(path, "%s/device/uevent", device->ibdev_path);
 953
 954         file = fopen(path, "rb");
 955         if (file == NULL)
 956                 return -1;
 957         while (fgets(line, sizeof(line), file) == line) {
 958                 size_t len = strlen(line);
 959                 int ret;
 960
 961                 /* Truncate long lines. */
 962                 if (len == (sizeof(line) - 1))
 963                         while (line[(len - 1)] != '\n') {
 964                                 ret = fgetc(file);
 965                                 if (ret == EOF)
 966                                         break;
 967                                 line[(len - 1)] = ret;
 968                         }
 969                 /* Extract information. */
 970                 if (sscanf(line,
 971                            "PCI_SLOT_NAME="
 972                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
 973                            &pci_addr->domain,
 974                            &pci_addr->bus,
 975                            &pci_addr->devid,
 976                            &pci_addr->function) == 4) {
 977                         ret = 0;
 978                         break;
 979                 }
 980         }
 981         fclose(file);
 982         return 0;
 983 }
 984
 985 /**
 986  * Update the link status.
 987  *
 988  * @param priv
 989  *   Pointer to private structure.
 990  *
 991  * @return
 992  *   Zero if the callback process can be called immediately.
 993  */
 994 static int
 995 priv_link_status_update(struct priv *priv)
 996 {
 997         struct rte_eth_link *link = &priv->dev->data->dev_link;
 998
 999         priv_link_update(priv, 0);
1000         if (((link->link_speed == 0) && link->link_status) ||
1001                 ((link->link_speed != 0) && !link->link_status)) {
1002                 /*
1003                  * Inconsistent status. Event likely occurred before the
1004                  * kernel netdevice exposes the new status.
1005                  */
1006                 if (!priv->pending_alarm) {
1007                         priv->pending_alarm = 1;
1008                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1009                                           mlx5_dev_link_status_handler,
1010                                           priv->dev);
1011                 }
1012                 return 1;
1013         } else if (unlikely(priv->pending_alarm)) {
1014                 /* Link interrupt occurred while alarm is already scheduled. */
1015                 priv->pending_alarm = 0;
1016                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
1017         }
1018         return 0;
1019 }
1020
1021 /**
1022  * Device status handler.
1023  *
1024  * @param priv
1025  *   Pointer to private structure.
1026  * @param events
1027  *   Pointer to event flags holder.
1028  *
1029  * @return
1030  *   Events bitmap of callback process which can be called immediately.
1031  */
1032 static uint32_t
1033 priv_dev_status_handler(struct priv *priv)
1034 {
1035         struct ibv_async_event event;
1036         uint32_t ret = 0;
1037
1038         /* Read all message and acknowledge them. */
1039         for (;;) {
1040                 if (mlx5_glue->get_async_event(priv->ctx, &event))
1041                         break;
1042                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1043                         event.event_type == IBV_EVENT_PORT_ERR) &&
1044                         (priv->dev->data->dev_conf.intr_conf.lsc == 1))
1045                         ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1046                 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1047                         priv->dev->data->dev_conf.intr_conf.rmv == 1)
1048                         ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1049                 else
1050                         DEBUG("event type %d on port %d not handled",
1051                               event.event_type, event.element.port_num);
1052                 mlx5_glue->ack_async_event(&event);
1053         }
1054         if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1055                 if (priv_link_status_update(priv))
1056                         ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1057         return ret;
1058 }
1059
1060 /**
1061  * Handle delayed link status event.
1062  *
1063  * @param arg
1064  *   Registered argument.
1065  */
1066 void
1067 mlx5_dev_link_status_handler(void *arg)
1068 {
1069         struct rte_eth_dev *dev = arg;
1070         struct priv *priv = dev->data->dev_private;
1071         int ret;
1072
1073         while (!priv_trylock(priv)) {
1074                 /* Alarm is being canceled. */
1075                 if (priv->pending_alarm == 0)
1076                         return;
1077                 rte_pause();
1078         }
1079         priv->pending_alarm = 0;
1080         ret = priv_link_status_update(priv);
1081         priv_unlock(priv);
1082         if (!ret)
1083                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1084 }
1085
1086 /**
1087  * Handle interrupts from the NIC.
1088  *
1089  * @param[in] intr_handle
1090  *   Interrupt handler.
1091  * @param cb_arg
1092  *   Callback argument.
1093  */
1094 void
1095 mlx5_dev_interrupt_handler(void *cb_arg)
1096 {
1097         struct rte_eth_dev *dev = cb_arg;
1098         struct priv *priv = dev->data->dev_private;
1099         uint32_t events;
1100
1101         priv_lock(priv);
1102         events = priv_dev_status_handler(priv);
1103         priv_unlock(priv);
1104         if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1105                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1106         if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1107                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1108 }
1109
1110 /**
1111  * Handle interrupts from the socket.
1112  *
1113  * @param cb_arg
1114  *   Callback argument.
1115  */
1116 static void
1117 mlx5_dev_handler_socket(void *cb_arg)
1118 {
1119         struct rte_eth_dev *dev = cb_arg;
1120         struct priv *priv = dev->data->dev_private;
1121
1122         priv_lock(priv);
1123         priv_socket_handle(priv);
1124         priv_unlock(priv);
1125 }
1126
1127 /**
1128  * Uninstall interrupt handler.
1129  *
1130  * @param priv
1131  *   Pointer to private structure.
1132  * @param dev
1133  *   Pointer to the rte_eth_dev structure.
1134  */
1135 void
1136 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1137 {
1138         if (dev->data->dev_conf.intr_conf.lsc ||
1139             dev->data->dev_conf.intr_conf.rmv)
1140                 rte_intr_callback_unregister(&priv->intr_handle,
1141                                              mlx5_dev_interrupt_handler, dev);
1142         if (priv->primary_socket)
1143                 rte_intr_callback_unregister(&priv->intr_handle_socket,
1144                                              mlx5_dev_handler_socket, dev);
1145         if (priv->pending_alarm) {
1146                 priv->pending_alarm = 0;
1147                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1148         }
1149         priv->intr_handle.fd = 0;
1150         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1151         priv->intr_handle_socket.fd = 0;
1152         priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1153 }
1154
1155 /**
1156  * Install interrupt handler.
1157  *
1158  * @param priv
1159  *   Pointer to private structure.
1160  * @param dev
1161  *   Pointer to the rte_eth_dev structure.
1162  */
1163 void
1164 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1165 {
1166         int rc, flags;
1167
1168         assert(priv->ctx->async_fd > 0);
1169         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1170         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1171         if (rc < 0) {
1172                 INFO("failed to change file descriptor async event queue");
1173                 dev->data->dev_conf.intr_conf.lsc = 0;
1174                 dev->data->dev_conf.intr_conf.rmv = 0;
1175         }
1176         if (dev->data->dev_conf.intr_conf.lsc ||
1177             dev->data->dev_conf.intr_conf.rmv) {
1178                 priv->intr_handle.fd = priv->ctx->async_fd;
1179                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1180                 rte_intr_callback_register(&priv->intr_handle,
1181                                            mlx5_dev_interrupt_handler, dev);
1182         }
1183
1184         rc = priv_socket_init(priv);
1185         if (!rc && priv->primary_socket) {
1186                 priv->intr_handle_socket.fd = priv->primary_socket;
1187                 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1188                 rte_intr_callback_register(&priv->intr_handle_socket,
1189                                            mlx5_dev_handler_socket, dev);
1190         }
1191 }
1192
1193 /**
1194  * Change the link state (UP / DOWN).
1195  *
1196  * @param priv
1197  *   Pointer to private data structure.
1198  * @param up
1199  *   Nonzero for link up, otherwise link down.
1200  *
1201  * @return
1202  *   0 on success, errno value on failure.
1203  */
1204 static int
1205 priv_dev_set_link(struct priv *priv, int up)
1206 {
1207         return priv_set_flags(priv, ~IFF_UP, up ? IFF_UP : ~IFF_UP);
1208 }
1209
1210 /**
1211  * DPDK callback to bring the link DOWN.
1212  *
1213  * @param dev
1214  *   Pointer to Ethernet device structure.
1215  *
1216  * @return
1217  *   0 on success, errno value on failure.
1218  */
1219 int
1220 mlx5_set_link_down(struct rte_eth_dev *dev)
1221 {
1222         struct priv *priv = dev->data->dev_private;
1223         int err;
1224
1225         priv_lock(priv);
1226         err = priv_dev_set_link(priv, 0);
1227         priv_unlock(priv);
1228         return err;
1229 }
1230
1231 /**
1232  * DPDK callback to bring the link UP.
1233  *
1234  * @param dev
1235  *   Pointer to Ethernet device structure.
1236  *
1237  * @return
1238  *   0 on success, errno value on failure.
1239  */
1240 int
1241 mlx5_set_link_up(struct rte_eth_dev *dev)
1242 {
1243         struct priv *priv = dev->data->dev_private;
1244         int err;
1245
1246         priv_lock(priv);
1247         err = priv_dev_set_link(priv, 1);
1248         priv_unlock(priv);
1249         return err;
1250 }
1251
1252 /**
1253  * Configure the TX function to use.
1254  *
1255  * @param priv
1256  *   Pointer to private data structure.
1257  * @param dev
1258  *   Pointer to rte_eth_dev structure.
1259  *
1260  * @return
1261  *   Pointer to selected Tx burst function.
1262  */
1263 eth_tx_burst_t
1264 priv_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1265 {
1266         eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1267         struct mlx5_dev_config *config = &priv->config;
1268         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1269         int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1270                                     DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1271                                     DEV_TX_OFFLOAD_GRE_TNL_TSO));
1272         int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1273
1274         assert(priv != NULL);
1275         /* Select appropriate TX function. */
1276         if (vlan_insert || tso)
1277                 return tx_pkt_burst;
1278         if (config->mps == MLX5_MPW_ENHANCED) {
1279                 if (priv_check_vec_tx_support(priv, dev) > 0) {
1280                         if (priv_check_raw_vec_tx_support(priv, dev) > 0)
1281                                 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1282                         else
1283                                 tx_pkt_burst = mlx5_tx_burst_vec;
1284                         DEBUG("selected Enhanced MPW TX vectorized function");
1285                 } else {
1286                         tx_pkt_burst = mlx5_tx_burst_empw;
1287                         DEBUG("selected Enhanced MPW TX function");
1288                 }
1289         } else if (config->mps && (config->txq_inline > 0)) {
1290                 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1291                 DEBUG("selected MPW inline TX function");
1292         } else if (config->mps) {
1293                 tx_pkt_burst = mlx5_tx_burst_mpw;
1294                 DEBUG("selected MPW TX function");
1295         }
1296         return tx_pkt_burst;
1297 }
1298
1299 /**
1300  * Configure the RX function to use.
1301  *
1302  * @param priv
1303  *   Pointer to private data structure.
1304  * @param dev
1305  *   Pointer to rte_eth_dev structure.
1306  *
1307  * @return
1308  *   Pointer to selected Rx burst function.
1309  */
1310 eth_rx_burst_t
1311 priv_select_rx_function(struct priv *priv, __rte_unused struct rte_eth_dev *dev)
1312 {
1313         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1314
1315         assert(priv != NULL);
1316         if (priv_check_vec_rx_support(priv) > 0) {
1317                 rx_pkt_burst = mlx5_rx_burst_vec;
1318                 DEBUG("selected RX vectorized function");
1319         }
1320         return rx_pkt_burst;
1321 }
1322
1323 /**
1324  * Check if mlx5 device was removed.
1325  *
1326  * @param dev
1327  *   Pointer to Ethernet device structure.
1328  *
1329  * @return
1330  *   1 when device is removed, otherwise 0.
1331  */
1332 int
1333 mlx5_is_removed(struct rte_eth_dev *dev)
1334 {
1335         struct ibv_device_attr device_attr;
1336         struct priv *priv = dev->data->dev_private;
1337
1338         if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1339                 return 1;
1340         return 0;
1341 }