drivers/net/mlx5/mlx5_ethdev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2015 Mellanox Technologies, Ltd
   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 <netinet/in.h>
  22 #include <linux/ethtool.h>
  23 #include <linux/sockios.h>
  24 #include <fcntl.h>
  25 #include <stdalign.h>
  26 #include <sys/un.h>
  27 #include <time.h>
  28
  29 #include <rte_atomic.h>
  30 #include <rte_ethdev_driver.h>
  31 #include <rte_bus_pci.h>
  32 #include <rte_mbuf.h>
  33 #include <rte_common.h>
  34 #include <rte_interrupts.h>
  35 #include <rte_malloc.h>
  36 #include <rte_string_fns.h>
  37 #include <rte_rwlock.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                         strlcpy(match, name, sizeof(match));
 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  * Get the interface index from device name.
 183  *
 184  * @param[in] dev
 185  *   Pointer to Ethernet device.
 186  *
 187  * @return
 188  *   Interface index on success, a negative errno value otherwise and
 189  *   rte_errno is set.
 190  */
 191 int
 192 mlx5_ifindex(const struct rte_eth_dev *dev)
 193 {
 194         char ifname[IF_NAMESIZE];
 195         int ret;
 196
 197         ret = mlx5_get_ifname(dev, &ifname);
 198         if (ret)
 199                 return ret;
 200         ret = if_nametoindex(ifname);
 201         if (ret == -1) {
 202                 rte_errno = errno;
 203                 return -rte_errno;
 204         }
 205         return ret;
 206 }
 207
 208 /**
 209  * Perform ifreq ioctl() on associated Ethernet device.
 210  *
 211  * @param[in] dev
 212  *   Pointer to Ethernet device.
 213  * @param req
 214  *   Request number to pass to ioctl().
 215  * @param[out] ifr
 216  *   Interface request structure output buffer.
 217  *
 218  * @return
 219  *   0 on success, a negative errno value otherwise and rte_errno is set.
 220  */
 221 int
 222 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
 223 {
 224         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 225         int ret = 0;
 226
 227         if (sock == -1) {
 228                 rte_errno = errno;
 229                 return -rte_errno;
 230         }
 231         ret = mlx5_get_ifname(dev, &ifr->ifr_name);
 232         if (ret)
 233                 goto error;
 234         ret = ioctl(sock, req, ifr);
 235         if (ret == -1) {
 236                 rte_errno = errno;
 237                 goto error;
 238         }
 239         close(sock);
 240         return 0;
 241 error:
 242         close(sock);
 243         return -rte_errno;
 244 }
 245
 246 /**
 247  * Get device MTU.
 248  *
 249  * @param dev
 250  *   Pointer to Ethernet device.
 251  * @param[out] mtu
 252  *   MTU value output buffer.
 253  *
 254  * @return
 255  *   0 on success, a negative errno value otherwise and rte_errno is set.
 256  */
 257 int
 258 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
 259 {
 260         struct ifreq request;
 261         int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
 262
 263         if (ret)
 264                 return ret;
 265         *mtu = request.ifr_mtu;
 266         return 0;
 267 }
 268
 269 /**
 270  * Set device MTU.
 271  *
 272  * @param dev
 273  *   Pointer to Ethernet device.
 274  * @param mtu
 275  *   MTU value to set.
 276  *
 277  * @return
 278  *   0 on success, a negative errno value otherwise and rte_errno is set.
 279  */
 280 static int
 281 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 282 {
 283         struct ifreq request = { .ifr_mtu = mtu, };
 284
 285         return mlx5_ifreq(dev, SIOCSIFMTU, &request);
 286 }
 287
 288 /**
 289  * Set device flags.
 290  *
 291  * @param dev
 292  *   Pointer to Ethernet device.
 293  * @param keep
 294  *   Bitmask for flags that must remain untouched.
 295  * @param flags
 296  *   Bitmask for flags to modify.
 297  *
 298  * @return
 299  *   0 on success, a negative errno value otherwise and rte_errno is set.
 300  */
 301 int
 302 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
 303 {
 304         struct ifreq request;
 305         int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
 306
 307         if (ret)
 308                 return ret;
 309         request.ifr_flags &= keep;
 310         request.ifr_flags |= flags & ~keep;
 311         return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
 312 }
 313
 314 /**
 315  * DPDK callback for Ethernet device configuration.
 316  *
 317  * @param dev
 318  *   Pointer to Ethernet device structure.
 319  *
 320  * @return
 321  *   0 on success, a negative errno value otherwise and rte_errno is set.
 322  */
 323 int
 324 mlx5_dev_configure(struct rte_eth_dev *dev)
 325 {
 326         struct priv *priv = dev->data->dev_private;
 327         unsigned int rxqs_n = dev->data->nb_rx_queues;
 328         unsigned int txqs_n = dev->data->nb_tx_queues;
 329         unsigned int i;
 330         unsigned int j;
 331         unsigned int reta_idx_n;
 332         const uint8_t use_app_rss_key =
 333                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 334         int ret = 0;
 335
 336         if (use_app_rss_key &&
 337             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 338              rss_hash_default_key_len)) {
 339                 DRV_LOG(ERR, "port %u RSS key len must be %zu Bytes long",
 340                         dev->data->port_id, rss_hash_default_key_len);
 341                 rte_errno = EINVAL;
 342                 return -rte_errno;
 343         }
 344         priv->rss_conf.rss_key =
 345                 rte_realloc(priv->rss_conf.rss_key,
 346                             rss_hash_default_key_len, 0);
 347         if (!priv->rss_conf.rss_key) {
 348                 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
 349                         dev->data->port_id, rxqs_n);
 350                 rte_errno = ENOMEM;
 351                 return -rte_errno;
 352         }
 353         memcpy(priv->rss_conf.rss_key,
 354                use_app_rss_key ?
 355                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 356                rss_hash_default_key,
 357                rss_hash_default_key_len);
 358         priv->rss_conf.rss_key_len = rss_hash_default_key_len;
 359         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 360         priv->rxqs = (void *)dev->data->rx_queues;
 361         priv->txqs = (void *)dev->data->tx_queues;
 362         if (txqs_n != priv->txqs_n) {
 363                 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
 364                         dev->data->port_id, priv->txqs_n, txqs_n);
 365                 priv->txqs_n = txqs_n;
 366         }
 367         if (rxqs_n > priv->config.ind_table_max_size) {
 368                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 369                         dev->data->port_id, rxqs_n);
 370                 rte_errno = EINVAL;
 371                 return -rte_errno;
 372         }
 373         if (rxqs_n == priv->rxqs_n)
 374                 return 0;
 375         DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 376                 dev->data->port_id, priv->rxqs_n, rxqs_n);
 377         priv->rxqs_n = rxqs_n;
 378         /* If the requested number of RX queues is not a power of two, use the
 379          * maximum indirection table size for better balancing.
 380          * The result is always rounded to the next power of two. */
 381         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 382                                      priv->config.ind_table_max_size :
 383                                      rxqs_n));
 384         ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
 385         if (ret)
 386                 return ret;
 387         /* When the number of RX queues is not a power of two, the remaining
 388          * table entries are padded with reused WQs and hashes are not spread
 389          * uniformly. */
 390         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 391                 (*priv->reta_idx)[i] = j;
 392                 if (++j == rxqs_n)
 393                         j = 0;
 394         }
 395         /*
 396          * Once the device is added to the list of memory event callback, its
 397          * global MR cache table cannot be expanded on the fly because of
 398          * deadlock. If it overflows, lookup should be done by searching MR list
 399          * linearly, which is slow.
 400          */
 401         if (mlx5_mr_btree_init(&priv->mr.cache, MLX5_MR_BTREE_CACHE_N * 2,
 402                                dev->device->numa_node)) {
 403                 /* rte_errno is already set. */
 404                 return -rte_errno;
 405         }
 406         return 0;
 407 }
 408
 409 /**
 410  * Sets default tuning parameters.
 411  *
 412  * @param dev
 413  *   Pointer to Ethernet device.
 414  * @param[out] info
 415  *   Info structure output buffer.
 416  */
 417 static void
 418 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 419 {
 420         struct priv *priv = dev->data->dev_private;
 421
 422         /* Minimum CPU utilization. */
 423         info->default_rxportconf.ring_size = 256;
 424         info->default_txportconf.ring_size = 256;
 425         info->default_rxportconf.burst_size = 64;
 426         info->default_txportconf.burst_size = 64;
 427         if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
 428                 info->default_rxportconf.nb_queues = 16;
 429                 info->default_txportconf.nb_queues = 16;
 430                 if (dev->data->nb_rx_queues > 2 ||
 431                     dev->data->nb_tx_queues > 2) {
 432                         /* Max Throughput. */
 433                         info->default_rxportconf.ring_size = 2048;
 434                         info->default_txportconf.ring_size = 2048;
 435                 }
 436         } else {
 437                 info->default_rxportconf.nb_queues = 8;
 438                 info->default_txportconf.nb_queues = 8;
 439                 if (dev->data->nb_rx_queues > 2 ||
 440                     dev->data->nb_tx_queues > 2) {
 441                         /* Max Throughput. */
 442                         info->default_rxportconf.ring_size = 4096;
 443                         info->default_txportconf.ring_size = 4096;
 444                 }
 445         }
 446 }
 447
 448 /**
 449  * DPDK callback to get information about the device.
 450  *
 451  * @param dev
 452  *   Pointer to Ethernet device structure.
 453  * @param[out] info
 454  *   Info structure output buffer.
 455  */
 456 void
 457 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 458 {
 459         struct priv *priv = dev->data->dev_private;
 460         struct mlx5_dev_config *config = &priv->config;
 461         unsigned int max;
 462         char ifname[IF_NAMESIZE];
 463
 464         /* FIXME: we should ask the device for these values. */
 465         info->min_rx_bufsize = 32;
 466         info->max_rx_pktlen = 65536;
 467         /*
 468          * Since we need one CQ per QP, the limit is the minimum number
 469          * between the two values.
 470          */
 471         max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
 472                       priv->device_attr.orig_attr.max_qp);
 473         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 474         if (max >= 65535)
 475                 max = 65535;
 476         info->max_rx_queues = max;
 477         info->max_tx_queues = max;
 478         info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
 479         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
 480         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
 481                                  info->rx_queue_offload_capa);
 482         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
 483         if (mlx5_get_ifname(dev, &ifname) == 0)
 484                 info->if_index = if_nametoindex(ifname);
 485         info->reta_size = priv->reta_idx_n ?
 486                 priv->reta_idx_n : config->ind_table_max_size;
 487         info->hash_key_size = rss_hash_default_key_len;
 488         info->speed_capa = priv->link_speed_capa;
 489         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
 490         mlx5_set_default_params(dev, info);
 491 }
 492
 493 /**
 494  * Get supported packet types.
 495  *
 496  * @param dev
 497  *   Pointer to Ethernet device structure.
 498  *
 499  * @return
 500  *   A pointer to the supported Packet types array.
 501  */
 502 const uint32_t *
 503 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 504 {
 505         static const uint32_t ptypes[] = {
 506                 /* refers to rxq_cq_to_pkt_type() */
 507                 RTE_PTYPE_L2_ETHER,
 508                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 509                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 510                 RTE_PTYPE_L4_NONFRAG,
 511                 RTE_PTYPE_L4_FRAG,
 512                 RTE_PTYPE_L4_TCP,
 513                 RTE_PTYPE_L4_UDP,
 514                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 515                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 516                 RTE_PTYPE_INNER_L4_NONFRAG,
 517                 RTE_PTYPE_INNER_L4_FRAG,
 518                 RTE_PTYPE_INNER_L4_TCP,
 519                 RTE_PTYPE_INNER_L4_UDP,
 520                 RTE_PTYPE_UNKNOWN
 521         };
 522
 523         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 524             dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
 525             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 526                 return ptypes;
 527         return NULL;
 528 }
 529
 530 /**
 531  * DPDK callback to retrieve physical link information.
 532  *
 533  * @param dev
 534  *   Pointer to Ethernet device structure.
 535  * @param[out] link
 536  *   Storage for current link status.
 537  *
 538  * @return
 539  *   0 on success, a negative errno value otherwise and rte_errno is set.
 540  */
 541 static int
 542 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
 543                                struct rte_eth_link *link)
 544 {
 545         struct priv *priv = dev->data->dev_private;
 546         struct ethtool_cmd edata = {
 547                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 548         };
 549         struct ifreq ifr;
 550         struct rte_eth_link dev_link;
 551         int link_speed = 0;
 552         int ret;
 553
 554         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 555         if (ret) {
 556                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 557                         dev->data->port_id, strerror(rte_errno));
 558                 return ret;
 559         }
 560         memset(&dev_link, 0, sizeof(dev_link));
 561         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 562                                 (ifr.ifr_flags & IFF_RUNNING));
 563         ifr.ifr_data = (void *)&edata;
 564         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 565         if (ret) {
 566                 DRV_LOG(WARNING,
 567                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 568                         dev->data->port_id, strerror(rte_errno));
 569                 return ret;
 570         }
 571         link_speed = ethtool_cmd_speed(&edata);
 572         if (link_speed == -1)
 573                 dev_link.link_speed = ETH_SPEED_NUM_NONE;
 574         else
 575                 dev_link.link_speed = link_speed;
 576         priv->link_speed_capa = 0;
 577         if (edata.supported & SUPPORTED_Autoneg)
 578                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 579         if (edata.supported & (SUPPORTED_1000baseT_Full |
 580                                SUPPORTED_1000baseKX_Full))
 581                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 582         if (edata.supported & SUPPORTED_10000baseKR_Full)
 583                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 584         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 585                                SUPPORTED_40000baseCR4_Full |
 586                                SUPPORTED_40000baseSR4_Full |
 587                                SUPPORTED_40000baseLR4_Full))
 588                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 589         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 590                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 591         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 592                         ETH_LINK_SPEED_FIXED);
 593         if ((dev_link.link_speed && !dev_link.link_status) ||
 594             (!dev_link.link_speed && dev_link.link_status)) {
 595                 rte_errno = EAGAIN;
 596                 return -rte_errno;
 597         }
 598         *link = dev_link;
 599         return 0;
 600 }
 601
 602 /**
 603  * Retrieve physical link information (unlocked version using new ioctl).
 604  *
 605  * @param dev
 606  *   Pointer to Ethernet device structure.
 607  * @param[out] link
 608  *   Storage for current link status.
 609  *
 610  * @return
 611  *   0 on success, a negative errno value otherwise and rte_errno is set.
 612  */
 613 static int
 614 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
 615                              struct rte_eth_link *link)
 616
 617 {
 618         struct priv *priv = dev->data->dev_private;
 619         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 620         struct ifreq ifr;
 621         struct rte_eth_link dev_link;
 622         uint64_t sc;
 623         int ret;
 624
 625         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 626         if (ret) {
 627                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 628                         dev->data->port_id, strerror(rte_errno));
 629                 return ret;
 630         }
 631         memset(&dev_link, 0, sizeof(dev_link));
 632         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 633                                 (ifr.ifr_flags & IFF_RUNNING));
 634         ifr.ifr_data = (void *)&gcmd;
 635         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 636         if (ret) {
 637                 DRV_LOG(DEBUG,
 638                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
 639                         " failed: %s",
 640                         dev->data->port_id, strerror(rte_errno));
 641                 return ret;
 642         }
 643         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 644
 645         alignas(struct ethtool_link_settings)
 646         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 647                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 648         struct ethtool_link_settings *ecmd = (void *)data;
 649
 650         *ecmd = gcmd;
 651         ifr.ifr_data = (void *)ecmd;
 652         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 653         if (ret) {
 654                 DRV_LOG(DEBUG,
 655                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
 656                         " failed: %s",
 657                         dev->data->port_id, strerror(rte_errno));
 658                 return ret;
 659         }
 660         dev_link.link_speed = ecmd->speed;
 661         sc = ecmd->link_mode_masks[0] |
 662                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 663         priv->link_speed_capa = 0;
 664         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 665                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 666         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 667                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
 668                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 669         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
 670                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
 671                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
 672                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 673         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
 674                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
 675                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 676         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
 677                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
 678                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
 679                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
 680                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 681         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
 682                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
 683                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
 684                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
 685                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 686         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
 687                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
 688                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
 689                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 690         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
 691                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
 692                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 693         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
 694                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
 695                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
 696                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
 697                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 698         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
 699                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 700         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 701                                   ETH_LINK_SPEED_FIXED);
 702         if ((dev_link.link_speed && !dev_link.link_status) ||
 703             (!dev_link.link_speed && dev_link.link_status)) {
 704                 rte_errno = EAGAIN;
 705                 return -rte_errno;
 706         }
 707         *link = dev_link;
 708         return 0;
 709 }
 710
 711 /**
 712  * DPDK callback to retrieve physical link information.
 713  *
 714  * @param dev
 715  *   Pointer to Ethernet device structure.
 716  * @param wait_to_complete
 717  *   Wait for request completion.
 718  *
 719  * @return
 720  *   0 if link status was not updated, positive if it was, a negative errno
 721  *   value otherwise and rte_errno is set.
 722  */
 723 int
 724 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 725 {
 726         int ret;
 727         struct rte_eth_link dev_link;
 728         time_t start_time = time(NULL);
 729
 730         do {
 731                 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
 732                 if (ret)
 733                         ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
 734                 if (ret == 0)
 735                         break;
 736                 /* Handle wait to complete situation. */
 737                 if (wait_to_complete && ret == -EAGAIN) {
 738                         if (abs((int)difftime(time(NULL), start_time)) <
 739                             MLX5_LINK_STATUS_TIMEOUT) {
 740                                 usleep(0);
 741                                 continue;
 742                         } else {
 743                                 rte_errno = EBUSY;
 744                                 return -rte_errno;
 745                         }
 746                 } else if (ret < 0) {
 747                         return ret;
 748                 }
 749         } while (wait_to_complete);
 750         ret = !!memcmp(&dev->data->dev_link, &dev_link,
 751                        sizeof(struct rte_eth_link));
 752         dev->data->dev_link = dev_link;
 753         return ret;
 754 }
 755
 756 /**
 757  * DPDK callback to change the MTU.
 758  *
 759  * @param dev
 760  *   Pointer to Ethernet device structure.
 761  * @param in_mtu
 762  *   New MTU.
 763  *
 764  * @return
 765  *   0 on success, a negative errno value otherwise and rte_errno is set.
 766  */
 767 int
 768 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 769 {
 770         struct priv *priv = dev->data->dev_private;
 771         uint16_t kern_mtu = 0;
 772         int ret;
 773
 774         ret = mlx5_get_mtu(dev, &kern_mtu);
 775         if (ret)
 776                 return ret;
 777         /* Set kernel interface MTU first. */
 778         ret = mlx5_set_mtu(dev, mtu);
 779         if (ret)
 780                 return ret;
 781         ret = mlx5_get_mtu(dev, &kern_mtu);
 782         if (ret)
 783                 return ret;
 784         if (kern_mtu == mtu) {
 785                 priv->mtu = mtu;
 786                 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
 787                         dev->data->port_id, mtu);
 788                 return 0;
 789         }
 790         rte_errno = EAGAIN;
 791         return -rte_errno;
 792 }
 793
 794 /**
 795  * DPDK callback to get flow control status.
 796  *
 797  * @param dev
 798  *   Pointer to Ethernet device structure.
 799  * @param[out] fc_conf
 800  *   Flow control output buffer.
 801  *
 802  * @return
 803  *   0 on success, a negative errno value otherwise and rte_errno is set.
 804  */
 805 int
 806 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 807 {
 808         struct ifreq ifr;
 809         struct ethtool_pauseparam ethpause = {
 810                 .cmd = ETHTOOL_GPAUSEPARAM
 811         };
 812         int ret;
 813
 814         ifr.ifr_data = (void *)&ethpause;
 815         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 816         if (ret) {
 817                 DRV_LOG(WARNING,
 818                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
 819                         " %s",
 820                         dev->data->port_id, strerror(rte_errno));
 821                 return ret;
 822         }
 823         fc_conf->autoneg = ethpause.autoneg;
 824         if (ethpause.rx_pause && ethpause.tx_pause)
 825                 fc_conf->mode = RTE_FC_FULL;
 826         else if (ethpause.rx_pause)
 827                 fc_conf->mode = RTE_FC_RX_PAUSE;
 828         else if (ethpause.tx_pause)
 829                 fc_conf->mode = RTE_FC_TX_PAUSE;
 830         else
 831                 fc_conf->mode = RTE_FC_NONE;
 832         return 0;
 833 }
 834
 835 /**
 836  * DPDK callback to modify flow control parameters.
 837  *
 838  * @param dev
 839  *   Pointer to Ethernet device structure.
 840  * @param[in] fc_conf
 841  *   Flow control parameters.
 842  *
 843  * @return
 844  *   0 on success, a negative errno value otherwise and rte_errno is set.
 845  */
 846 int
 847 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 848 {
 849         struct ifreq ifr;
 850         struct ethtool_pauseparam ethpause = {
 851                 .cmd = ETHTOOL_SPAUSEPARAM
 852         };
 853         int ret;
 854
 855         ifr.ifr_data = (void *)&ethpause;
 856         ethpause.autoneg = fc_conf->autoneg;
 857         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 858             (fc_conf->mode & RTE_FC_RX_PAUSE))
 859                 ethpause.rx_pause = 1;
 860         else
 861                 ethpause.rx_pause = 0;
 862
 863         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 864             (fc_conf->mode & RTE_FC_TX_PAUSE))
 865                 ethpause.tx_pause = 1;
 866         else
 867                 ethpause.tx_pause = 0;
 868         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 869         if (ret) {
 870                 DRV_LOG(WARNING,
 871                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
 872                         " failed: %s",
 873                         dev->data->port_id, strerror(rte_errno));
 874                 return ret;
 875         }
 876         return 0;
 877 }
 878
 879 /**
 880  * Get PCI information from struct ibv_device.
 881  *
 882  * @param device
 883  *   Pointer to Ethernet device structure.
 884  * @param[out] pci_addr
 885  *   PCI bus address output buffer.
 886  *
 887  * @return
 888  *   0 on success, a negative errno value otherwise and rte_errno is set.
 889  */
 890 int
 891 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
 892                             struct rte_pci_addr *pci_addr)
 893 {
 894         FILE *file;
 895         char line[32];
 896         MKSTR(path, "%s/device/uevent", device->ibdev_path);
 897
 898         file = fopen(path, "rb");
 899         if (file == NULL) {
 900                 rte_errno = errno;
 901                 return -rte_errno;
 902         }
 903         while (fgets(line, sizeof(line), file) == line) {
 904                 size_t len = strlen(line);
 905                 int ret;
 906
 907                 /* Truncate long lines. */
 908                 if (len == (sizeof(line) - 1))
 909                         while (line[(len - 1)] != '\n') {
 910                                 ret = fgetc(file);
 911                                 if (ret == EOF)
 912                                         break;
 913                                 line[(len - 1)] = ret;
 914                         }
 915                 /* Extract information. */
 916                 if (sscanf(line,
 917                            "PCI_SLOT_NAME="
 918                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
 919                            &pci_addr->domain,
 920                            &pci_addr->bus,
 921                            &pci_addr->devid,
 922                            &pci_addr->function) == 4) {
 923                         ret = 0;
 924                         break;
 925                 }
 926         }
 927         fclose(file);
 928         return 0;
 929 }
 930
 931 /**
 932  * Device status handler.
 933  *
 934  * @param dev
 935  *   Pointer to Ethernet device.
 936  * @param events
 937  *   Pointer to event flags holder.
 938  *
 939  * @return
 940  *   Events bitmap of callback process which can be called immediately.
 941  */
 942 static uint32_t
 943 mlx5_dev_status_handler(struct rte_eth_dev *dev)
 944 {
 945         struct priv *priv = dev->data->dev_private;
 946         struct ibv_async_event event;
 947         uint32_t ret = 0;
 948
 949         if (mlx5_link_update(dev, 0) == -EAGAIN) {
 950                 usleep(0);
 951                 return 0;
 952         }
 953         /* Read all message and acknowledge them. */
 954         for (;;) {
 955                 if (mlx5_glue->get_async_event(priv->ctx, &event))
 956                         break;
 957                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
 958                         event.event_type == IBV_EVENT_PORT_ERR) &&
 959                         (dev->data->dev_conf.intr_conf.lsc == 1))
 960                         ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
 961                 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
 962                         dev->data->dev_conf.intr_conf.rmv == 1)
 963                         ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
 964                 else
 965                         DRV_LOG(DEBUG,
 966                                 "port %u event type %d on not handled",
 967                                 dev->data->port_id, event.event_type);
 968                 mlx5_glue->ack_async_event(&event);
 969         }
 970         return ret;
 971 }
 972
 973 /**
 974  * Handle interrupts from the NIC.
 975  *
 976  * @param[in] intr_handle
 977  *   Interrupt handler.
 978  * @param cb_arg
 979  *   Callback argument.
 980  */
 981 void
 982 mlx5_dev_interrupt_handler(void *cb_arg)
 983 {
 984         struct rte_eth_dev *dev = cb_arg;
 985         uint32_t events;
 986
 987         events = mlx5_dev_status_handler(dev);
 988         if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
 989                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
 990         if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
 991                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
 992 }
 993
 994 /**
 995  * Handle interrupts from the socket.
 996  *
 997  * @param cb_arg
 998  *   Callback argument.
 999  */
1000 static void
1001 mlx5_dev_handler_socket(void *cb_arg)
1002 {
1003         struct rte_eth_dev *dev = cb_arg;
1004
1005         mlx5_socket_handle(dev);
1006 }
1007
1008 /**
1009  * Uninstall interrupt handler.
1010  *
1011  * @param dev
1012  *   Pointer to Ethernet device.
1013  */
1014 void
1015 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1016 {
1017         struct priv *priv = dev->data->dev_private;
1018
1019         if (dev->data->dev_conf.intr_conf.lsc ||
1020             dev->data->dev_conf.intr_conf.rmv)
1021                 rte_intr_callback_unregister(&priv->intr_handle,
1022                                              mlx5_dev_interrupt_handler, dev);
1023         if (priv->primary_socket)
1024                 rte_intr_callback_unregister(&priv->intr_handle_socket,
1025                                              mlx5_dev_handler_socket, dev);
1026         priv->intr_handle.fd = 0;
1027         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1028         priv->intr_handle_socket.fd = 0;
1029         priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1030 }
1031
1032 /**
1033  * Install interrupt handler.
1034  *
1035  * @param dev
1036  *   Pointer to Ethernet device.
1037  */
1038 void
1039 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1040 {
1041         struct priv *priv = dev->data->dev_private;
1042         int ret;
1043         int flags;
1044
1045         assert(priv->ctx->async_fd > 0);
1046         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1047         ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1048         if (ret) {
1049                 DRV_LOG(INFO,
1050                         "port %u failed to change file descriptor async event"
1051                         " queue",
1052                         dev->data->port_id);
1053                 dev->data->dev_conf.intr_conf.lsc = 0;
1054                 dev->data->dev_conf.intr_conf.rmv = 0;
1055         }
1056         if (dev->data->dev_conf.intr_conf.lsc ||
1057             dev->data->dev_conf.intr_conf.rmv) {
1058                 priv->intr_handle.fd = priv->ctx->async_fd;
1059                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1060                 rte_intr_callback_register(&priv->intr_handle,
1061                                            mlx5_dev_interrupt_handler, dev);
1062         }
1063         ret = mlx5_socket_init(dev);
1064         if (ret)
1065                 DRV_LOG(ERR, "port %u cannot initialise socket: %s",
1066                         dev->data->port_id, strerror(rte_errno));
1067         else if (priv->primary_socket) {
1068                 priv->intr_handle_socket.fd = priv->primary_socket;
1069                 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1070                 rte_intr_callback_register(&priv->intr_handle_socket,
1071                                            mlx5_dev_handler_socket, dev);
1072         }
1073 }
1074
1075 /**
1076  * DPDK callback to bring the link DOWN.
1077  *
1078  * @param dev
1079  *   Pointer to Ethernet device structure.
1080  *
1081  * @return
1082  *   0 on success, a negative errno value otherwise and rte_errno is set.
1083  */
1084 int
1085 mlx5_set_link_down(struct rte_eth_dev *dev)
1086 {
1087         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1088 }
1089
1090 /**
1091  * DPDK callback to bring the link UP.
1092  *
1093  * @param dev
1094  *   Pointer to Ethernet device structure.
1095  *
1096  * @return
1097  *   0 on success, a negative errno value otherwise and rte_errno is set.
1098  */
1099 int
1100 mlx5_set_link_up(struct rte_eth_dev *dev)
1101 {
1102         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1103 }
1104
1105 /**
1106  * Configure the TX function to use.
1107  *
1108  * @param dev
1109  *   Pointer to private data structure.
1110  *
1111  * @return
1112  *   Pointer to selected Tx burst function.
1113  */
1114 eth_tx_burst_t
1115 mlx5_select_tx_function(struct rte_eth_dev *dev)
1116 {
1117         struct priv *priv = dev->data->dev_private;
1118         eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1119         struct mlx5_dev_config *config = &priv->config;
1120         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1121         int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1122                                     DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1123                                     DEV_TX_OFFLOAD_GRE_TNL_TSO |
1124                                     DEV_TX_OFFLOAD_IP_TNL_TSO |
1125                                     DEV_TX_OFFLOAD_UDP_TNL_TSO));
1126         int swp = !!(tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
1127                                     DEV_TX_OFFLOAD_UDP_TNL_TSO |
1128                                     DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM));
1129         int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1130
1131         assert(priv != NULL);
1132         /* Select appropriate TX function. */
1133         if (vlan_insert || tso || swp)
1134                 return tx_pkt_burst;
1135         if (config->mps == MLX5_MPW_ENHANCED) {
1136                 if (mlx5_check_vec_tx_support(dev) > 0) {
1137                         if (mlx5_check_raw_vec_tx_support(dev) > 0)
1138                                 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1139                         else
1140                                 tx_pkt_burst = mlx5_tx_burst_vec;
1141                         DRV_LOG(DEBUG,
1142                                 "port %u selected enhanced MPW Tx vectorized"
1143                                 " function",
1144                                 dev->data->port_id);
1145                 } else {
1146                         tx_pkt_burst = mlx5_tx_burst_empw;
1147                         DRV_LOG(DEBUG,
1148                                 "port %u selected enhanced MPW Tx function",
1149                                 dev->data->port_id);
1150                 }
1151         } else if (config->mps && (config->txq_inline > 0)) {
1152                 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1153                 DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1154                         dev->data->port_id);
1155         } else if (config->mps) {
1156                 tx_pkt_burst = mlx5_tx_burst_mpw;
1157                 DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1158                         dev->data->port_id);
1159         }
1160         return tx_pkt_burst;
1161 }
1162
1163 /**
1164  * Configure the RX function to use.
1165  *
1166  * @param dev
1167  *   Pointer to private data structure.
1168  *
1169  * @return
1170  *   Pointer to selected Rx burst function.
1171  */
1172 eth_rx_burst_t
1173 mlx5_select_rx_function(struct rte_eth_dev *dev)
1174 {
1175         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1176
1177         assert(dev != NULL);
1178         if (mlx5_check_vec_rx_support(dev) > 0) {
1179                 rx_pkt_burst = mlx5_rx_burst_vec;
1180                 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1181                         dev->data->port_id);
1182         } else if (mlx5_mprq_enabled(dev)) {
1183                 rx_pkt_burst = mlx5_rx_burst_mprq;
1184         }
1185         return rx_pkt_burst;
1186 }
1187
1188 /**
1189  * Check if mlx5 device was removed.
1190  *
1191  * @param dev
1192  *   Pointer to Ethernet device structure.
1193  *
1194  * @return
1195  *   1 when device is removed, otherwise 0.
1196  */
1197 int
1198 mlx5_is_removed(struct rte_eth_dev *dev)
1199 {
1200         struct ibv_device_attr device_attr;
1201         struct priv *priv = dev->data->dev_private;
1202
1203         if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1204                 return 1;
1205         return 0;
1206 }